The cardiac sympathetic innervation originates from sympathetic preganglionic neurons (S1N) in the upper thoracic segments of the spinal cord, which synapse with neurons (S2N) in the cervical and upper thoracic ganglia. Sympathetic postganglionic fibres innervate the cardiac conduction system more prominently than the working myocardium [1], releasing noradrenaline (norepinephrine) and neuropeptide Y. Noradrenaline binds to cardiac beta-1 adrenergic receptors, increasing intracellular levels of cyclic adenosine monophosphate (cAMP). In myocytes of the sinoatrial node, which is the physiological cardiac pacemaker, increased cAMP levels hasten diastolic depolarization by increasing the inward ‘funny’ (If) cation current gated by hyperpolarization-activated cyclic nucleotide-gated channels [2]. As a result, the heart period (HP) shortens almost linearly with the frequency of sympathetic postganglionic discharge [3]. This response is delayed by 1.7 s and the frequency response essentially filters out the fluctuations of sympathetic activity faster than 0.15 Hz [4], which are typically associated with breathing [5]. By itself, decreases in HP raise ventricular contractility [6] and relaxation rate [7]. These effects are further enhanced by the increased cAMP levels, which increase sarcoplasmic reticulum calcium reuptake because of protein kinase A-dependent phosphorylation of phospholamban [8]. The decrease in HP also increases atrioventricular conduction time, but this effect is counteracted by sympathetic activity [9]. Sympathetic innervation of cardiac automatic, conduction and contractile tissue follows parallel yet distinct intrapericardial pathways [10], which are the basis for selective effects [11]. Sympathetic activity is also associated with ventricular repolarization heterogeneity, as indexed by T-wave alternans in subjects with coronary artery disease [12] and by QT interval variability in dogs with experimental heart failure [13]. The sympathetic nervous system also controls the heart by promoting adrenal medullary release of adrenaline (epinephrine), whose plasma threshold for decreasing HP is in the range of plasma levels attained during active standing [14]. Show The parasympathetic preganglionic neurons (P1N) involved in the control of cardiac function are located in the medulla oblongata within and ventrolateral to the nucleus ambiguus (NAmb), and to a much lesser extent in the dorsal motor nucleus of the vagus nerve (DMNX) and in the reticular formation between these two nuclei [15]. All of these neurons project through the vagus nerves to a complex set of epicardial ganglionated plexi [16], which display functional selectivity for the effects on HP, conduction time and contractility [17], and receive synaptic contacts by local sensory neurons [18]. Parasympathetic postganglionic fibres widely innervate the cardiac conduction system as well as the atrial and ventricular working myocardium [1,16], releasing acetylcholine and vasoactive intestinal peptide. Acetylcholine decreases cAMP levels in myocardial cells by binding to M2 muscarinic receptors [8]. These receptors also open specific potassium channels that gate a hyperpolarizing current, particularly in conditions of marked parasympathetic activity [19]. As a result, HP lengthens almost linearly with parasympathetic preganglionic discharge rate [3]. Resting HP values are higher than the intrinsic HP in the absence of autonomic control [20], highlighting the prominence of the cardiac parasympathetic tone over the cardiac sympathetic tone. Sinoatrial responses to acetylcholine occur with minimal delay and keep up with modulations at least up to 0.4 Hz [4]. Parasympathetic activity also lengthens atrioventricular conduction time [9] and decreases atrial and ventricular contractility [16] and ventricular relaxation rate [21], but does not affect ventricular repolarization heterogeneity as indexed by T-wave alternans [12] or QT interval variability [13]. The sympathetic and parasympathetic cardiac controls are antagonistic in that they exert opposite effects on myocardial cAMP levels, and thereby on HP, atrioventricular conduction time, ventricular contractility and relaxation rate. Evidence that sympathetic noradrenaline release is decreased by ongoing parasympathetic activity also suggests an inhibitory presynaptic interaction [22]. The autonomic control of the mean values of heart rate (i.e. 1/HP) displays accentuated nonlinear antagonism in that even high levels of sympathetic activity exert negligible effects when parasympathetic activity is also high [23]. This nonlinear interaction is much less prominent with respect to values of atrioventricular conduction time [23] and even of the mean values of HP [3], which are well explained by a linear summation of sympathetic and parasympathetic effects [24]. As far as autonomic modulations are concerned, activity of each autonomic branch enhances the fluctuations of heart rate resulting from the modulation of the other branch [25]. This synergistic interaction suggests that the high-frequency fluctuations of heart rate that constitute respiratory sinus arrhythmia are enhanced by sympathetic activity and, therefore, do not result solely from parasympathetic modulation, as would be expected from the frequency response properties detailed previously [25]. In contrast to this suggestion, however, analysis of HP fluctuations has shown that beta-adrenergic blockade increases respiratory sinus arrhythmia in human subjects [26]. Taken together, these discrepancies highlight that the interactions between sympathetic and parasympathetic activities on cardiac function are complex and still incompletely understood, and that differences between analyses based on HP and analyses based on heart rate may contribute to the variability in the reported results. The autonomic outflow to the heart is regulated by a central autonomic network (CAN) of interconnected brain structures, which includes the medial prefrontal cortex (MPFC) and insular cortex, the amygdala and the bed nucleus of the stria terminalis (BNST), the lateral region of the hypothalamus and the paraventricular nucleus (PVN) and dorsomedial hypothalamic (DMH) nucleus, the periaqueductal grey (PAG) matter of the midbrain, the parabrachial Kölliker–Fuse region of the lateral pons, as well as several regions of the medulla, which partly overlap with those involved in respiratory control [27]. The MPFC comprises the anterior cingulate cortex and the prelimbic and infralimbic areas and is involved in both cognitive and visceromotor functions, thus being of potential great relevance for psychosomatic medicine [28]. The insula is a viscerosensory and visceromotor region [29] and plays a key role in physiological and pathological cardiovascular control [30]. In humans, the right (non-dominant) anterior insular cortex is involved in the generation of the mental image of one's physical state, which underlies basic emotional states [31]. The amygdala is primarily involved in the information processing related to negative emotions, whereas positive emotions tend to reduce amygdala activation [32]. The insular cortex, the central nucleus of the amygdala and the BNST constitute a cortico-striatal–pallidal circuit that processes emotional information with autonomic responses [27] and projects to the hypothalamic behaviour control column [33]. The PVN is a master controller of the autonomic nervous system, providing specialized innervation to all autonomic relay centres [27], and together with the DMH it integrates neuroendocrine, homeostatic and stress responses [34]. The PAG takes part in regulating autonomic responses to physical and psychological stressors [35]. The medullary nucleus of the tractus solitarius (NTS) and the pontine parabrachial and Kölliker–Fuse nuclei are reciprocally connected and relay visceral afferent information to other CAN structures [36]. In the medulla, the CAN includes the NAmb and DMNX, the rostral (RVLM) and caudal portions of the ventrolateral medulla, and the rostral ventromedial medulla (RVMM), which comprises the midline medullary raphe and the parapyramidal area [36]. Each cardiac P2N receives inputs from only a single active P1N [37]. Accordingly, the functional specificity of cardiac P2N is matched by that of the P1N in the NAmb, separate groups of which selectively and independently control HP, atrioventricular conduction and ventricular contractility [38]. The P1N in the NAmb have myelinated B-fibre axons, while those in the DMNX have unmyelinated slow-conducting C-fibre axons. The P1N in the DMNX may regulate to some extent coronary blood flow or cardiac contractility, but, in contrast to P1N in the NAmb, they have relatively little effect on HP [39]. The P1N in the DMNX have a low level of ongoing activity that is not related to cardiac or respiratory activity, and are unaffected by baroreceptor inputs [40], although they are activated by stimulation of pulmonary C-fibre afferents [41]. Conversely, the P1N in the NAmb have a relatively high level of activity that shows a distinct cardiac and respiratory modulation [42]. Inputs from baroreceptors, chemoreceptors and nasopharyngeal receptors converge on individual P1N in the NAmb [37] and are gated by inspiratory-related inputs, with the result that the reflex bradycardia normally evoked by stimulation of these receptors is greatly attenuated during inspiration [43] (figure 1). The inspiratory gating of reflex bradycardic responses is one of the mechanisms responsible for respiratory sinus arrhythmia and is believed to have the effect of improving the ventilation–perfusion relationship in the lungs [44]. Apart from inputs from peripheral receptors and central inspiratory neurons, the P1N in the NAmb or DMNX also receive inputs from other sources at all levels of the brain, including direct projections from the hypothalamic PVN [45] and the lateral (l) and ventrolateral (vl) parts of the midbrain PAG (figure 1) [46]. The main cortical areas that regulate HP are the MPFC, which is thought to promote cardiac parasympathetic activity [28,47], and the insular cortex, which can increase or decrease HP depending on the stimulated site [30]. The insular cortex sites where stimulation decreases HP and increases arterial blood pressure densely innervate the MPFC and amygdala [48]. In most of these studies, it is not completely clear if the changes in HP are due to sympathetic or parasympathetic effects. Studies using the retrograde trans-synaptic transport of viral vectors revealed indirect descending pathways by which the MPFC and insular cortex may regulate the activity of P1N in the NAmb or DMNX, involving the amygdala, PAG and NTS [49] (figure 1).  Figure 1. Schematic diagram showing the major central pathways regulating the cardiac parasympathetic outflow. No distinction is made between excitatory and inhibitory connections. DMNX, dorsal motor nucleus of the vagus nerve; l, lateral; MDH, medullary dorsal horn of the trigeminal nucleus; MPFC, medial prefrontal cortex; NAmb, nucleus ambiguus, NTS, nucleus of the tractus solitarius; PAG, periaqueductal grey; PVN, hypothalamic paraventricular nucleus; vl, ventrolateral.
The major input to the cardiac S1N arises from sympathetic premotor neurons (S0N) in the RVMM, RVLM, A5 area in the pons and hypothalamic PVN [50] (figure 2). These cell groups differ with respect to their anatomical connections and functional properties. As will be discussed in more detail below, S0N in the RVMM are activated during arousal and stress but do not generate cardiac responses to baroreceptor or chemoreceptor inputs. By contrast, S0N in the RVLM are critical for the expression of cardiac reflex responses to baroreceptor, chemoreceptor and other inputs [36], but do not appear to have a major role in the generation of cardiac responses in stress and arousal [51]. Much less information is available about the precise functional role of S0N in the A5 area and PVN, although in regard to the latter it is likely that they are essential for generating cardiac sympathetic responses to inputs associated with changes in blood volume or plasma osmolality [52]. The chemical properties of the S0N in the different regions are also highly varied. For example, many of the S0N in the RVMM contain serotonin, while those in the RVLM are mainly adrenergic neurons of the C1 group. The S0N in the A5 group in the pons synthesize noradrenaline, and in the PVN many of the S0N contain oxytocin, corticotropin-releasing factor, vasopressin or angiotensin [50]. Neurons at all levels of the brain regulate the activity of the S0N, but many of the descending pathways to the S0N in the medulla include synapses in other nuclei (figure 2). As described above, the MPFC and insular cortex may regulate HP through changes in cardiac sympathetic or parasympathetic activity. Activation of neurons in the MPFC may also decrease vasomotor sympathetic activity [53]. The descending pathways from the MPFC and insular cortex to the S0N in the medulla are not clearly defined. Anatomical studies in animals have revealed direct descending projections to the RVLM as well as potential indirect projections to both the RVMM and RVLM that include synaptic connections in the amygdala, PAG and NTS [35,53] (figure 2). Figure 2. Schematic diagram showing the major central pathways regulating the cardiac sympathetic outflow. No distinction is made between excitatory and inhibitory connections. CVLM, caudal ventrolateral medulla; DMH, dorsomedial hypothalamus; RVLM, rostral ventrolateral medulla; RVMM, rostral ventromedial medulla (for more abbreviations, refer to figure 1).
Cardiac function can be profoundly altered by the reflex activation of cardiac autonomic nerves in response to inputs from receptors including baroreceptors, chemoreceptors, receptors from skeletal muscles and nasopharyngeal receptors. Cardiac sympathetic and parasympathetic nerve activities are reciprocally altered in response to baroreceptor stimulation, whereas in response to stimulation of chemoreceptor and nasopharyngeal inputs, both cardiac sympathetic and parasympathetic activities are increased [54]. The physiological advantage of reciprocal baroreflex control of the heart by sympathetic and parasympathetic nerves is that it allows for rapid and large compensatory responses to perturbations in blood pressure [54]. The essential central circuits that mediate the baroreceptor reflex control of the heart have been well defined [36]. Primary baroreceptor and chemoreceptor afferent fibres terminate in the NTS on separate and spatially distinct subgroups of the second-order neurons. From the NTS, baroreceptor signals are transmitted via excitatory glutamatergic projection neurons directly to cardiac P1N in the NAmb (figure 1) or to the caudal ventrolateral medulla, where they synapse with GABAergic neurons. In turn, these neurons project to and inhibit S0N in the RVLM, which control the sympathetic outflow to the heart and different vascular beds (figure 2). There is good evidence that the S0N in the RVLM that control the sympathetic outflow to different vascular beds form separate subgroups [55]. It also seems likely that cardiac S0N in the RVLM are distinct from sympathetic vasomotor S0N, given that the cardiac and vasomotor sympathetic activities are reflexly regulated in a differentiated manner, according to the afferent input [36]. The physiological cardiac response to the chemoreceptor reflex has the advantage of maximizing oxygen conservation, while at the same time maintaining an adequate perfusion pressure for the brain and the heart. This is achieved largely by a vagally evoked bradycardia (which reduces cardiac oxygen consumption), while co-activation of cardiac sympathetic outflow maintains an optimum stroke volume and cardiac output to meet the essential metabolic needs of the brain and heart [54]. Chemoreceptor signals are transmitted from the NTS via direct projections to P1N in the NAmb (figure 1) and to S0N in the RVLM [56,57] (figure 2). Given that chemoreceptor stimulation evokes parasympathetic and sympathetic co-activation, it seems likely that both of these projections are excitatory, presumably glutamatergic. In addition, chemoreceptor signals are also transmitted indirectly to the RVLM via the Kölliker–Fuse nucleus in the pons [56]. The mechanical distortion and interstitial accumulation of metabolites produced by skeletal muscle contraction stimulate group III/IV thin fibre afferents from skeletal muscles, which elicit the exercise pressor reflex. This reflex supports muscle perfusion by raising arterial blood pressure as a result of increased sympathetic outflow to the heart and systemic vasculature and of decreased parasympathetic outflow to the heart [58]. The neural pathway of the exercise pressor reflex is not known with certainty, except for the site of the first synapse, which is in the spinal cord dorsal horn [58]. However, neurons responsive to group III muscle afferents have been demonstrated in the NTS [59], suggesting that the NTS is a key relay also for this reflex. The diving reflex is perhaps the most powerful autonomic reflex known, as it is associated with intense sympathetically mediated vasoconstriction and profound vagally mediated bradycardia [60]. Like the chemoreceptor reflex, it has the effect of conserving oxygen during submersion. In non-diving animals such as rabbits, a very similar reflex is evoked by irritant vapours [61]. In both cases, nasopharyngeal receptors innervated by trigeminal afferents trigger the reflex responses. There is evidence for both a direct projection from trigeminal afferents to the NAmb region that contains cardiac P1N, as well as indirect projections relayed via the medullary dorsal horn of the trigeminal nerve [60] (figure 1). Physical exercise, either dynamic or static (isometric), entails a decrease in HP, which is abrupt in onset, maintained or even enhanced while exercise occurs, and progressively reversed after exercise is over [62]. The changes in HP at the onset and offset of exercise result primarily from parasympathetic withdrawal and reactivation, respectively [62,63]. During steady-state dynamic exercise, the decrease in HP results from a continuum of balanced sympatho-vagal control, with parasympathetic withdrawal playing the greater role the lower the workload [64]. These autonomic changes result from the integration of the baroreceptor and exercise pressor reflexes with central autonomic commands [65]. These commands are classically defined as feed-forward modulation of P1N and S1N exerted by the descending somatic motor pathways, but also incorporate feedback components associated with the perception of effort [66]. The neural structures that issue the central autonomic commands associated with physical exercise are still uncertain, and have been suggested to overlap with those of the CAN at the level of the NTS, PVN, PAG, insular cortex and MPFC [62,65]. In this last respect, a remarkable study that coupled magnetic resonance neuroimaging with measurements of HP, muscle sympathetic nerve activity and arterial blood pressure in human subjects provided evidence that decreased activity of the ventral MPFC is involved in the central commands that cause parasympathetic cardiac withdrawal during static handgrip exercise [67]. During exercise, both central commands and the exercise pressor reflex reset the operating point of the arterial baroreceptor reflex towards higher values of blood pressure and heart rate [65]. There is evidence that this resetting is mediated by neuronal circuitry in the NTS [59]. Chronotropic incompetence, i.e. the failure to increase heart rate to a level commensurate with increased workload, may occur in patients with heart failure because sympathetic overactivation desensitizes myocardial beta-adrenergic receptors, and portends a poor prognosis [68]. Aerobic exercise training may decrease sympathetic activation and ameliorate chronotropic incompetence in patients with heart failure [68,69]. The mechanisms of this beneficial effect are thought to involve inhibition of the NFκB-dependent inflammatory cascade, angiotensin II receptor expression and oxidative stress in RVLM and PVN neurons [69]. Stress may be defined as a condition in which expectations, whether genetically programmed, learned or deduced, do not match perceptions of the internal or the external environment, this discrepancy eliciting compensatory responses [70]. The concept of arousal is state-dependent, and, during wakefulness, it refers to the cerebral, autonomic and behavioural activation in response to internal and environmental stimuli [71]. In the immediate response to a novel psychological stress (e.g. air-puff stress), the changes in HP are often biphasic, due to co-activation of cardiac sympathetic and parasympathetic nerves, consistent with an orienting response to the arousing stimulus [54,72]. With repeated exposures, however, there is a decrease in HP which is due predominantly to cardiac sympathetic activation [54]. There is good evidence that S0N located in the RVMM (especially the midline raphe) generate stress-evoked increases in cardiac sympathetic activity [73]. The midline raphe receives a direct projection from the DMH, a region that is critical for the expression of the cardiovascular response to stress [72,73], and from the lPAG [74], which as mentioned above is a region that generates cardiovascular responses to physical stressors [35,72]. In addition, both the DMH and lPAG receive direct inputs from the amygdala and direct or indirect inputs from the insular cortex (figure 2), both of which are also involved in generating responses to physical stressors [35,72]. Furthermore, air-puff stress induces an increase in c-Fos expression (a marker of neuronal activation) in the RVMM, but not in the RVLM, which, as mentioned above, is activated when cardiac sympathetic activity is reflexly increased by inputs from peripheral receptors [51]. The RVMM contains more S0N than any other brain region [50]. The descending pathway from the DMH to the RVMM is functionally asymmetric, such that disinhibition of the right DMH induces significantly larger decreases in HP and increases in cardiac contractility and number of ectopic beats than those evoked from the left DMH [75]. Thus, the pathway from the right DMH to the RVMM may contribute to stress-related increases in cardiac sympathetic activity [75], which is a main cause of malignant tachyarrhythmias [76], and further explains the fact that intense emotional stress can cause sudden death (see, for example, [77]), as also discussed in §2d(iv). Finally, vagally mediated heart rate variability may also decrease with stress, and according to a remarkable meta-analysis of neuroimaging studies may index the extent to which threat representations encoded in the amygdala are inhibited by the ventral MPFC based on the external and internal perceptions of safe contexts [47]. The transition from wakefulness to non-rapid eye movement (non-REM) sleep brings about an increase in HP [78], which is attributed to an increase in parasympathetic activity [79]. On passing from non-REM sleep to REM sleep, which accounts for approximately 20% of total sleep time, HP generally shows a mild tonic decrease, which is superimposed on large transient changes in HP associated with bursts of rapid eye movements, phasic increases in the activity of the pyramidal tract, ponto-geniculo-occipital waves, and acceleration of the hippocampal theta rhythm (reviewed in [78]). These sleep-related changes in HP are best explained by sleep-related central autonomic commands, which override or reset baroreflex function [80]. The hypothalamic suprachiasmatic nucleus (SCN) is the master pacemaker responsible for the circadian rhythms of wake–sleep states [81] and of blood pressure and HP [82]. Neurons in the SCN may issue circadian autonomic commands involving sympathetic and parasympathetic outflows by projecting to the PVN [83], which, as mentioned previously, is a master autonomic controller in the CAN [27]. However, the main SCN contribution to the circadian cardiovascular rhythms is thought to come indirectly from the circadian rhythms of rest and activity [84] or, more precisely, of wakefulness, non-REM sleep and REM sleep [85], each of which appears associated with specific central autonomic commands [86]. The central autonomic commands during non-REM sleep may involve inhibitory projections from sleep-active neurons of the hypothalamic ventrolateral preoptic nucleus to the S0N of the PVN [87], central thermoregulatory pathways including the RVMM, central baroreflex pathways including the NTS, the parabrachial and Kölliker–Fuse nuclei, and command neurons in the pedunculopontine tegmental (PPT) nucleus [86]. During REM sleep, central autonomic commands may involve the PAG, the pontine sublaterodorsal and PPT nuclei, and the vestibular and raphe obscurus medullary nuclei [86]. Neurodegenerative disorders frequently cause slowly progressive failure of the autonomic control, mainly concerning its sympathetic component, resulting in multiple clinical manifestations (orthostatic hypotension, impaired sweating, neurogenic bladder with erectile dysfunction in men, gastrointestinal dysmotility), which may be subtle due to compensatory mechanisms [88]. Conversely, vascular, inflammatory or traumatic lesions of the autonomic nervous system and drug adverse effects most often manifest acutely with signs of autonomic hyperactivity, including abnormal excessive control of the cardiovascular system. Sustained and chronic autonomic hyperactivity may also appear in association with other chronic neurological disorders, in particular sleep disorders [89]. Both acute and chronic manifestations of an imbalanced brain–heart interaction represent a risk factor for the development of cardiovascular diseases or acute cardiovascular events, which can also lead to sudden cardiac death. It may be quite difficult to separate out the cardiac component in these clinical conditions. By contrast, viewing derangements of the brain–heart interaction from the perspective of the underlying autonomic syndromes may provide key pathophysiological insights and therapeutic success. Given these clinical implications, in the present section we will characterize the clinical features, the supposed pathogenetic mechanisms, and the common causes of autonomic hyperactivity, focusing on its cardiovascular consequences. Autonomic hyperactivity is characterized by excessive sympathetic activation, either in isolation or in association with excessive parasympathetic activation (table 1) [90]. Excessive sympathetic activity may result from activation of descending sympathoexcitatory pathways, disinhibition of sympathoexcitatory reflexes [91], sympathetic activation by hypoxia or ischaemia [92], loss of baroreflex buffering [93] or loss of inhibitory GABAergic control at diencephalic [94], brainstem or spinal levels [95]. Sympathetic overactivity is thought to be the common phenomenon that links the major cardiac pathologies seen in neurological catastrophes [96] (table 2). In general, when physical or psychological stressors challenge the body, a transient sympathetic overactivity, mainly with active coping mechanisms, is adaptive as part of the short-term survival machinery. This has been referred to as ‘allostasis’, which is the re-establishment of homeostasis through change in the level of operation of the physiological system (see [70]). By contrast, chronic stress states produce passive or withdrawal coping mechanisms and elicit a long-term autonomic response referred to as a ‘hyperarousal state’, characterized by chronic sympathetic and hypothalamo-pituitary–adrenocortical system activation. The consequences of overactivity of the allostatic systems are referred to as ‘excessive allostatic load’, which is maladaptive and leads to chronic diseases (table 3).
Episodes of paroxysmal sympathetic hyperactivity frequently occur in the setting of severe diffuse axonal injury and in the post-resuscitation period after severe anoxic–ischaemic brain insults [97]. These paroxysms usually start 5–7 days after injury and follow a regular pattern (1–3 times per day, lasting 1–10 h). Over time, the episodes tend to become less frequent and more prolonged and are associated with poor outcome. Release of hypothalamic CAN structures from cortical inhibitory control may explain acute episodes of autonomic hyperactivity of diencephalic origin, which may occur after closed-head injury associated with a decorticate state and widespread axonal damage [27]. Subarachnoid haemorrhage is associated with acute and massive sympathetic hyperactivity [98]. This may produce electrocardiographic changes indicative of cardiac injury [99,100] and different arrhythmias, particularly supraventricular tachyarrhythmias such as atrial fibrillation. Acute hydrocephalus after subarachnoid haemorrhage may cause episodes of acute autonomic hyperactivity linked to disinhibition of hypothalamic CAN structures such as the PVN [27]. Mechanical and chemical stimulation of the insular cortex has also been postulated as a mechanism for autonomic derangements caused by subarachnoid haemorrhage in the Sylvian fissure [30]. Acute stroke can impair central autonomic control, resulting in myocardial injury, electrocardiographic abnormalities, cardiac arrhythmias and ultimately sudden death. Cardiac and renal diseases, diabetes and arrhythmogenic medications increase the risk of cardiac morbidity and mortality after stroke [100]. Autonomic imbalance is more frequent after infarcts involving the insular cortex, which is part of the CAN [30]. Despite evidence indicating a lateralization of autonomic control by the insular cortex [27], there is no obvious laterality in the impact of insular cortex stroke on cardiovascular response or stroke prognosis [30]. Focal seizures involving CAN structures such as the MPFC, insular cortex and amygdala entail various autonomic manifestations [27] including sympathetic hyperactivity and cardiac arrhythmias [101]. The most common cardiac manifestation is sinus tachycardia but paroxysmal atrial fibrillation, supraventricular and ventricular tachycardia or ventricular fibrillation may also occur. Other autonomic manifestations of temporal lobe seizures include paroxysmal hypertension, ictal piloerection, sweating and facial flushing or pallor [102]. Temporal lobe seizures, particularly originating from the left hemisphere, may also manifest with excessive parasympathetic activity leading to ictal bradycardia and asystole which can cause syncope [103]. Seizure-related cardiac arrhythmias have been implicated as a potential pathogenetic mechanism of sudden unexpected death in epilepsy. However, ictal asystole usually has a self-limiting course, and postictal rather than ictal arrhythmias together with respiratory abnormalities seem of greater importance to the pathophysiology of sudden unexpected death in epilepsy [104]. In seizures occurring during sleep, as in nocturnal frontal lobe epilepsy, an increase in sympathetic–parasympathetic balance, similar to that associated with physiological arousal from sleep, usually precedes the onset of the motor manifestations. This suggests a role of autonomic activation, which is part of the arousal response, in triggering seizures [105]. Sympathetically mediated hypertension, cardiac arrhythmias or myocardial injury may occur as a consequence of brainstem lesions, particularly those involving the lateral medulla [106]. Rarely, paroxysmal sympathetic activity may be the presentation of baroreflex failure due to bilateral involvement of the NTS [107]. The baroreflex failure syndrome refers to the cardiovascular manifestations resulting from interruption of the afferent limb of the baroreflex at the level of the carotid sinus, baroreceptor afferents or medulla. The main clinical manifestations of baroreflex failure are acute hypertension or fluctuating hypertension with or without orthostatic hypotension or orthostatic tachycardia. Episodes of severe bradycardia and hypotension, referred to as malignant vagotonia, may also occur [108]. Autonomic dysreflexia refers to episodes of massive sympathetic hyperactivity triggered by reflex stimuli below the lesion in patients with cervical or thoracic spinal cord injury above T5. These episodes occur upon recovery from the acute spinal shock [109] and are characterized by severe hypertension that may result in hypertensive encephalopathy, intracranial or retinal haemorrhage, seizures or even sudden death [110]. Headache, anxiety, facial flushing and extensive diaphoresis above the level of the lesion often precede the onset of the hypertensive crisis. The most common stimuli originate from the bladder or rectum. Sympathetic or parasympathetic hyperactivity or hypofunction in different combinations may be detected in about two-thirds of patients with Guillain–Barré syndrome [111,112]. Autonomic hyperactivity may manifest with sinus tachycardia, sustained or paroxysmal hypertension, episodes of flushing, orthostatic hypotension and cardiac arrhythmias including cardiac arrest. The main mechanism of cardiovascular instability in Guillain–Barré syndrome is impaired baroreflex modulation of the sympathetic cardiovascular output due to demyelination of baroreceptor afferents [112]. Episodes of severe paroxysmal hypertension may result in subarachnoid haemorrhage or posterior leukoencephalopathy syndrome or takotsubo syndrome [113], whereas some patients are at risk of severe vagally mediated bradycardia and asystole in response to reflex stimuli such as tracheal suction [112]. The systemic inflammatory response to infection includes tachycardia in a context of sympathetic overactivity and myocardial contractile dysfunction, which converts into autonomic failure preceding shock in severe cases [114]. Inflammatory mediators act directly on the heart to cause tachycardia in experimental sepsis [115]. However, the sympathetic outflow to the heart also markedly increases in these conditions [116]. Cardiac sympathetic activation during sepsis may result from disinhibition of the DMH and the RVMM caused by prostaglandin E2 binding to neurons in the medial preoptic hypothalamus [116,117]. This is a dramatic example of the complex and still poorly understood interactions between the brain, autonomic activity and the immune system, which are a burgeoning area of research [118]. Several drugs can trigger syndromes characterized by mental status changes, increase of muscle tone, hyperthermia and autonomic hyperactivity. These include the neuroleptic malignant syndrome, which may be elicited by dopamine D2 receptor blocking antipsychotics [119]; malignant hyperthermia, which is a hypermetabolic response to volatile anaesthetics or to succinylcholine driven by excessive calcium release in the skeletal muscle [120]; the serotonin syndrome, which results from overactivation of central and peripheral serotonin receptors, often by antidepressant drugs [121]; and the anticholinergic syndrome, which results from the inhibition of muscarinic cholinergic neurotransmission [122]. The brain structures responsible for autonomic hyperactivity associated with these iatrogenic causes are still unclear. Fatal familial insomnia (FFI) is a rare familial prion disease whose clinical hallmark is the agrypnia excitata syndrome [123]. Somatomotor abnormalities (e.g. pyramidal signs, myoclonus, dysarthria/dysphagia and gait dysfunctions) also occur with variable latency and degree during the disease course. The term agrypnia (from the Greek expression for ‘to chase sleep’), as proposed by Lugaresi & Provini [124], describes an organic insomnia, which is characterized by severe or complete lack of sleep, especially deep sleep. The sleep disorder is associated with sympathetic and motor hyperactivation (excitata) and with episodes of a peculiar oneiric behaviour (oneiric stupor) [125]. Agrypnia excitata is typical of but not specific for FFI [94]. At rest, FFI patients show from the onset of agrypnia a progressive worsening of hypertension, tachycardia and hyperthermia sustained for 24 h resulting in a progressive decline in the amplitude of circadian oscillations of the autonomic parameters [126]. FFI is characterized by unbalanced autonomic control with preserved parasympathetic activity but higher background and stimulated sympathetic activity [127]. Clinico-pathological relations and functional imaging in FFI implicate the anterior ventral and mediodorsal nuclei of the thalamus in the regulation of the wake–sleep cycle and other autonomic functions, with sparing of hypothalamic and brainstem structures [128]. The mediodorsal thalamic nucleus is an integral part of the circuits that connect the hypothalamus with the amygdala and the MPFC components of the CAN [33]. The preferential thalamic lesions in FFI would act by disconnecting the limbic cortical areas involved in the control of instinctive behaviour and the cortical and subcortical regions that promote sleep and regulate autonomic functions. Such a disconnection syndrome results in a shift to persistent wakefulness behaviour and sympathetic hyperactivation, which makes FFI the paradigm of what can occur in humans if the arousal system cannot be shut off. Dysfunctions of the medial thalamus and related limbic areas are believed to underlie also the agrypnia excitata associated with delirium tremens, Morvan syndrome [123], and Whipple disease [129]. The obstructive sleep apnoea syndrome (OSAS) is characterized by repetitive episodes of complete (apnoea) or partial (hypopnoea) upper airway obstruction occurring during sleep, which usually result in blood oxygen desaturation and often terminate with brief arousal. The apnoeic episodes are associated with sympathetic hyperactivity and blood pressure increase, and with HP increase at onset of apnoeas followed by tachycardia on resumption of breathing [130]. OSAS with this continuous repetition of obstructive respiratory events during the night is a paradigm of how a sleep breathing disorder can lead to a permanent dysregulation of the autonomic cardiovascular control resulting in sustained sympathetic hyperactivity. Recent data on animal models highlight the key role played in this process by the NTS and PVN [131]. A remarkable functional imaging study on patients suggested that the elevated sympathetic activity associated with OSAS may be driven by changes in activity in higher cortical regions, such as the MPFC [132]. An increase in sympathetic tone is thought to underlie the cardiovascular complications responsible for the increased mortality rate in patients with sleep apnoea [133]. In addition, inhibition of the excitatory pathway from the PVN to P1N during chronic hypoxia/hypercapnia may decrease cardiac parasympathetic activity, further increasing the risk of adverse cardiac events associated with OSAS [45]. At least in women, OSAS severity is significantly associated with incident heart failure [134], which brings about cardiopulmonary reflex desensitization and severe cardiac and renal sympathoexcitation [135]. The OSAS is also related to the development of daytime hypertension [136] and to excessive daytime sleepiness [137]. A reduced baroreflex sensitivity is also a well-documented feature of OSAS before the onset of cardiovascular complications [138]. Daytime hypertension and excessive daytime sleepiness may share a common underlying cause: the cardiovascular effects of OSAS may be due to the top-down dysfunction of the baroreflex, while the reduced diurnal vigilance may be explained by a bottom-up dysfunction of the baroreceptor afferent effect [71]. This could be better understood from an allostatic perspective where the chronic hypertensive state associated with OSAS might be viewed as the result of autonomic nervous system adaptation to the episodic recurrence of sympathetic surges during the night. Therefore, the reduced baroreflex sensitivity consistently described in OSAS could be an indirect index of an as yet unknown maladaptive mechanism resulting in a decreased baroreflex function. Sleep has important homeostatic functions. Whether sleep deprivation and fragmentation is due to sleep disorders such as restless legs syndrome [89] or anxiety or depression or a hectic lifestyle, it may disturb many essential homeostatic mechanisms with adverse effects on autonomic, hormonal and metabolic systems [139]. There is evidence that short sleep duration is associated with cardiovascular morbidity in epidemiological surveys [140] and that sleep deprivation profoundly affects the autonomic nervous system [141]. Acute [142] and chronic [143] sleep deprivation are associated with increased sympathetic and decreased parasympathetic cardiovascular modulation. Long-term sleep deprivation is a chronic stressor that may decrease resting HP and increase systolic and diastolic blood pressure [144], entailing a modest but independent increase in the risk of cardiac ischaemia, stroke and sudden cardiac death [145]. The central neural mechanisms responsible for the autonomic effects of sleep debt are still unclear. However, it seems reasonable to surmise that they consist of altered function of those same structures of the CAN which underlie the sleep-related central autonomic commands (see §1d(iii)). Another sleep disorder with possible autonomic involvement is narcolepsy with cataplexy, a chronic disorder of the sleep–wake behaviour associated with the loss of neurons releasing hypocretin/orexin peptides [146]. Preliminary evidence indicates that autonomic control of cardiac variability by baroreflex and central autonomic (feed-forward) mechanisms is altered in narcolepsy with cataplexy patients during spontaneous sleep–wake behaviour, and particularly during wakefulness before sleep [147]. The extent to which narcolepsy with cataplexy entails derangements in the control of the cardiovascular system is still a matter of debate [146]. Interestingly, however, hypocretin/orexin neurons project to all of the brain structures that are thought to underlie the sleep-related central autonomic commands [86]. After the pioneering work of Walter B. Cannon in 1942 [148] with an article proposing a scientific basis for ‘voodoo’ death, several studies were performed to clarify the neurobiological basis of the link between emotional and cardiovascular events that are a major cause of morbidity and mortality in the developed world. Today we have strong evidence that cardiovascular events can be triggered by acute mental stress caused by events such as an earthquake, a televised high-drama soccer game, job strain or the death of a loved one [149]. Animal studies suggest that stressors presumed to provoke acute negative emotions affect cardiovascular physiological control and increase the risk of sudden cardiac death through haemodynamic and electrophysiological pathways increasing sympathetic output, impairing endothelial function and creating a hypercoagulable state [150]. Interestingly, post-traumatic stress disorder, which is an anxiety disorder initiated by exposure to a traumatic event, is also independently associated with increased risk of incident coronary heart disease and mortality [151]. From the perspective of the clinician, the analysis of heart rate variability (HRV) based on the electrocardiogram holds promise as an attractively simple tool for detecting autonomic impairments and for predicting the prognosis of some neurological disorders through the assessment of the brain–heart connections. Unfortunately, we are still far from the realization of this very important unmet need. According to the Task Force on HRV evaluation (1996) [152], there are only two clinical situations where HRV analysis should be performed: to assess mortality risk in patients after myocardial infarction and to detect early evidence of cardiac autonomic neuropathy in diabetic patients. In the light of the clinical relevance of the brain–heart connection for so many diseases, this implies that what is really lacking to develop specific clinical applications of the knowledge on heart–brain interactions are simple, widely available and reliable cardiovascular markers of the sympathetic tone and of the sympathetic–parasympathetic balance. Such markers would be invaluable for the early detection of signs of cardiovascular dysautonomia, the treatment of which can avoid the occurrence of the life-threatening paroxysmal or chronic autonomic hyperactivity. This state of affairs highlights the need for a deeper physiological understanding of the links between brain dynamics and the corresponding autonomic cardiac dynamics. The key issue at stake here is to combine different sources such as: information on animal models and human subjects, information on basic physiology and clinical disorders, but also, and critically, information from different biosignals in each given setting. In the physiology laboratory, key advances can be obtained by multi-modal recordings of electroencephalogram, electrocardiogram, electromyogram, blood pressure and respiration in animal models with congenital or acquired (viral vectors, drug-inducible expression systems) genetic modification of brain circuits. Multi-modal recordings including beat-to-beat blood pressure (finger volume clamp), respiration, muscle sympathetic nerve activity (peroneal nerve microneurography) and body temperature are also desirable when investigating brain–heart interactions in the clinical laboratory, in order to interpret cardiac control in its true physiological context. Taking advantage of the recent developments in magnetic resonance neuroimaging, as well as in advanced processing techniques of electroencephalographic, electrocardiographic and cerebrovascular flow signals, interdisciplinary and multi-modal research approaches now have the chance of bringing our understanding of the brain–heart interactions to the next level. A.S., G.C.-B., R.A.L.D. and P.C. carried out the bibliographic research, analysed the data, interpreted the results and wrote sections of the paper. R.A.L.D. constructed the figure schemes. A.S. drafted the full manuscript, which was reviewed by G.C.-B., R.A.L.D. and P.C. P.C. coordinated the project. All authors gave final approval for publication. We have no competing interests. No specific funding was dedicated to this project. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 2Magnetic resonance imaging (MRI) has become the primary technique for investigating the macroscopic, systems-level structure and function of the living human brain. As the sensitivity of MRI to detecting neural activity improves with increasing field strength [1], there has been a continual movement towards higher-field imaging. Developments in ‘ultra-high-field’ MRI (UHF, referring to static magnetic fields of 7 T and above) have provided substantial gains in the spatial resolution and localization of neural activity, facilitating investigation of fine-scale cortical structure at the level of columns and layers, and of substructure in subcortical grey matter regions such as the basal ganglia and the thalamus (e.g. [2–4]). Throughout the past decade, high-field MRI has paved the way for novel discoveries and applications in multiple domains of human neuroscience (reviewed in [5–9]). Interactions between the heart and the brain are vital for maintaining homeostasis and survival in an ever-changing environment. Moreover, abnormalities in cardiovascular function and other aspects of the autonomic nervous system are reciprocally linked with numerous brain disorders (e.g. [10–14]; see also [15]). While current knowledge about the pathways underpinning brain–heart interactions is based primarily on studies in the anaesthetized animal [16,17], much is unknown about its circuitry in the human. The interaction between cardiac activity and attentional, cognitive and emotional processes (e.g. [18–20]) is also rich ground for discovery, and there is growing recognition of the concept that the brain may best be studied in the context of its unification with the rest of the body (e.g. [20,21]; also [22]). Methodologically, the scope of research on brain–heart interactions in the human has to date been limited by the difficulty of non-invasively localizing and measuring the neural activity of key subcortical structures, and in modelling their temporal interactions with the rest of the brain. A number of key nuclei of the brainstem and forebrain cannot be adequately resolved with positron emission tomography (PET) or conventional functional magnetic resonance imaging (fMRI) protocols because of their limited spatial resolution [23]. However, recent advances in neuroimaging technology and data analysis are showing promise in surmounting these barriers. In particular, UHF MRI appears well suited to applications in autonomic neuroscience, including the investigation of brain–heart interactions. The improvement in spatial localization at UHF may better enable the study of small structures involved in the control and mediation of cardiac activity, along with their interactions within distributed brain networks that facilitate adaptive modulation to environmental and internal demands. While limitations of fMRI at presently ‘standard’ field strengths of 3 T and 1.5 T effectively restrict spatial resolution to above 2×2×2 mm3 (with voxel dimensions of 2.5–4 mm being typical of whole-brain acquisitions), field strengths of 7 T and above—in conjunction with progress in hardware and accelerated imaging methods—are permitting voxel sizes of 1 mm3 and below. Accordingly, as discussed below, data acquired at high field allows for differentiating between sub-regions of brainstem, amygdala, thalamus and other substrates of autonomic function. UHF MRI also has substantial advantages for the study of brain anatomy: large increases in contrast with susceptibility-weighted imaging at UHF have allowed structural resolution to approach 200 μm [4] which, in combination with fMRI, can help to better characterize the regions involved in brain–heart interactions. Yet, the increasing severity of spatial and temporal artefacts with increased field strengths (see below) necessitates careful treatment as well as continued methodological development in both data acquisition and post-processing aspects. This article offers a review and perspective regarding the potential opportunities of high-field MRI (focusing on fMRI) in the study of brain–heart interactions, and outlines several major accompanying technological and physiological limitations. We also discuss recent and future directions for which high-field imaging data can be used in combination with novel data analysis techniques for probing reciprocal relationships between neural dynamics and cardiac activity. As our primary aim is to motivate the role of high-field imaging in studying brain–heart interactions, we refer the reader to review articles such as [5,6,24] for further information about high-field MRI physics and image acquisition. fMRI [25–27] is a method for non-invasively measuring changes in brain activity over time (see [28] for a recent review). A time series of images is acquired wherein the fluctuation in intensity at a voxel over time indirectly reflects the dynamics of the local neural activity. The vast majority of fMRI studies are based on the blood-oxygen-level-dependent (BOLD) effect [29], by which changes in neural activity are coupled with local changes in blood flow and blood oxygenation. As the magnetic properties of oxygenated versus deoxygenated blood differ, this haemodynamic response to neural activity is detectable by MRI. The spatial and temporal scales (millimetres and seconds, respectively) of BOLD fMRI are too coarse to fully capture the breadth of information contained in neuronal activity; rather, it is limited to detecting slow changes in activity of large neuronal assemblies, and hence sophisticated experimental design and data analysis are required to optimally study brain function with fMRI. The functional involvement of a brain region can be examined by contrasting its time course of relative signal intensity in response to two or more well-controlled experimental conditions. More recently, spontaneous fluctuations in the fMRI signal, i.e. those that occur in the absence of (or seemingly independently of) explicitly presented stimuli and behaviour, have also been widely investigated and found to reveal topographies of distributed neurocognitive networks [30]. In both cases, the sensitivity of fMRI to detecting changes in neural activity depends on the ratio of contrast to noise (contrast-to-noise ratio, CNR). ‘Contrast’ reflects the magnitude of the fMRI signal change caused by neural activity, and ‘noise’ refers to temporal signal changes caused by other processes, often called ‘temporal noise’. Temporal noise contains contributions from the intrinsic MRI noise in each image (also called ‘thermal noise’) as well as structured fluctuations that are not related to neural activity, including head motion, instrumental instabilities and physiological (cardiac and respiratory) processes. In general, both BOLD contrast and physiological noise increase with the square of the field strength, while thermal noise increases linearly with field strength. Hence, CNR increases with field strength, though the gains are limited under conditions where there is a strong contribution from physiological noise [31]. The spatial specificity of BOLD contrast to neural activity also improves with increasing field strength, as the contribution from neural tissue increases at a faster rate than that of large vessels [1]. Increases in CNR can be harnessed to image at higher spatial resolution, which is one of the main advantages of high-field MRI. The exact dependence of CNR on field strength (and on resolution) is complicated. For example, when neural activity produces BOLD contrast over a spatial extent greater than the voxel size, reducing the volume of the voxel will lead to a proportional decrease in contrast; however, when the spatial extent of contrast is smaller than the voxel size, reducing the voxel volume may not lead to appreciable reduction in contrast. In addition, while thermal noise is largely independent of voxel size, noise from physiological processes, subject motion and instrumental instabilities scale up with the signal level, and therefore will decrease when the voxel size is reduced (again, provided that the spatial extent of the noise source exceeds the voxel size) [32]. Thus, it is recommended that one decrease the voxel size until thermal noise is the dominant noise source. In practice, for voxel sizes of 1 mm3 and below, thermal noise will be the dominant noise source and the spatial extent of the functional contrast will often exceed the voxel size. Further reductions in voxel volume will lead to proportional (rapid) reductions in CNR. In summary, the increased contrast at high field allows one to reduce the voxel size. At a given voxel size, the benefits of high field are limited by non-thermal noise that is due to sources including physiological processes and head motion. In addition to reducing the voxel size such that thermal noise dominates, isolating and modelling structured physiological noise may further improve the CNR and is an active area of research (see §4). Importantly, imaging at higher spatial resolution does not necessarily improve the spatial accuracy of measuring neural activity. In addition to head movement and within-brain pulsatility, which are of greater concern as one moves to higher resolution, BOLD contrast arises from vascular sources that may be somewhat distant from the active neural tissue. This is particularly the case for the larger (pial) veins, which may contain blood draining from an active site several millimetres away [33]. In this regard, a notable advantage of UHF fMRI is the inherent increase in the signal contribution of small vessels/capillaries close to the site of neural activity. Briefly, increasing the field strength changes the degree to which different vascular compartments contribute to BOLD contrast: the intravascular (blood) contribution decreases (due to the substantial reduction in the T2* of blood), but the extravascular contribution increases linearly near large vessels and quadratically near capillaries [34]. While the extravascular contribution near large vessels can still compromise spatial specificity, it is possible to suppress its contribution by performing a spin-echo acquisition (though at the expense of sensitivity and increased power deposition) [35] or by identifying them based on their appearance and location [36]. We note that, while BOLD fMRI contrast benefits substantially from UHF, this is not the case for all MRI contrasts. For example, the gains are much decreased with diffusion-weighted contrast due to the accelerated T2 relaxation at high field. A potential drawback of UHF stems from magnetic field inhomogeneities caused by tissue interfaces with air and bone. Such inhomogeneities increase linearly with field strength and may cause image distortion as well as signal loss due to intra-voxel dephasing, also referred to as ‘susceptibility artefacts’. Improvements in accelerated imaging techniques and distortion-correction methods have been (and will continue to be) critical for minimizing these artefacts [5,7,37]. In typical fMRI studies, shimming cannot fully compensate for localized field inhomogeneities from susceptibility differences; this might only be possible for very small fields of view. Spin-echo methods, which suffer less from susceptibility-related signal loss compared to gradient-echo methods, benefit markedly at 7 T over 3 T due to the aforementioned changes in the BOLD contrast mechanism as well as from the increased CNR [35]. However, despite its gains at UHF, spin-echo techniques have reduced sensitivity and also result in greater RF power deposition and resulting tissue heating, limiting the extent of volume coverage ([38] and see [39] for a review). Capitalizing on recent advances in UHF imaging, a number of studies have demonstrated the ability to map features of neural organization with exquisite spatial detail. fMRI studies with in-plane resolutions of 1 mm and below have revealed columnar structure [2,40] and layer-dependent specialization in cortex [3,41–43], tonotopic organization of the inferior colliculus [44] and benefits for characterizing networks at the whole-brain level [45]. Knowledge about the neurocircuitry underlying brain–heart interactions has primarily been established through invasive studies in anaesthetized animals [17,46,47]. More recently, non-invasive imaging techniques such as PET and fMRI have enabled further insight into central autonomic processing in the human, and in the context of higher-order emotional and cognitive responses [18,48,49]. The improved functional specificity of fMRI at higher field strengths is opening further opportunities for understanding the substrates and pathways of brain–heart interactions; many of the structures implicated in cardiovascular control and other autonomic functions consist of small-volume nuclei in the brainstem, as well as sub-regions of subcortical structures including hypothalamus, thalamus and amygdala, that cannot be adequately resolved at conventional field strengths and could benefit greatly from the increased resolution available at higher field (figure 1). Yet, critically, these regions are prone to susceptibility artefacts and physiological noise due to their locations near air–tissue boundaries and large vessels, limiting potential gains in functional contrast and requiring careful attention to acquisition and post-processing techniques. Below, we discuss findings from several domains of neuroscience that exemplify the promise of high-field fMRI in studying brain regions reported to be common to, or of similar size to and imaging difficulty as, structures likely implicated in human brain–heart interactions. Figure 1. Several key areas involved in brain–heart interactions. Selected areas of the forebrain, brainstem and spinal cord involved in human autonomic function are depicted. Arrows illustrate major pathways of the baroreceptor reflex, mediating the homeostatic control of blood pressure (left and right branches stemming from the baroreceptor afferents indicate parasympathetic and sympathetic outflow, respectively). Scale bars provide approximate dimensions in the human brain. For reviews, see [17,18,46,47]. Figure design inspired by Benarroch et al. [50]. (Online version in colour.)
The brainstem houses small and interconnected nuclei which, via ascending and descending projections to cortex and spinal cord, play a vital role in maintaining homeostatic control and adaptively modulating physiological responses to environmental demands. The brainstem presents substantial challenges for neuroimaging and, as such, has remained one of the least-studied structures in the human brain in vivo. For instance, the brainstem is prone to high levels of cardiac and cerebrospinal fluid-induced pulsatile motion due to its proximity to large vessels and ventricles [51,52], and may also undergo bulk motion with the cardiac cycle [53,54]. In addition, prominent susceptibility artefacts may result from the bone and air-filled cavities such as the pontine cistern. Consequently, the BOLD signal quality in the brainstem is considerably lower than that of most other brain regions [55,56]. See [57] for a recent, detailed discussion focused on fMRI of the midbrain. Nonetheless, with attention to acquisition and post-processing strategies (reviewed in [56] and see §4), several groups have reported success in resolving fMRI signals from functionally distinct brainstem areas. The central circuitry underlying the baroreflex, i.e. the homeostatic regulation of blood pressure, was examined at 3 T with a concurrently recorded index of sympathetic output [58,59]. Using an interleaved acquisition scheme and optimized coverage of brainstem areas, increases in muscle sympathetic activity were associated with fMRI signal increases in the rostral ventrolateral medulla along with decreases in its caudal aspect and in the nucleus of the solitary tract, consistent with known baroreflex pathways (figure 2). At 7 T, the internal organization of the periaqueductal gray (PAG) rostrocaudally and into columns was shown to be functionally differentiated in emotional and volitional respiration tasks, using spatial resolutions of 0.75 mm and 1 mm, respectively [60,61]. Figure 2. fMRI signal intensity changes correlated with sympathetic outflow. fMRI signal intensity changes correlated with spontaneous fluctuations in muscle sympathetic nerve activity (MSNA) in eight subjects. Increases (hot colour scale) and decreases(cool colour scale) in fMRI signal intensity with increases in MSNA are overlaid onto a series of axial fMRI slices from an individual subject. RVLM, rostral ventrolateral medulla; NTS, nucleus tractus solitarius; CVLM, caudal ventrolateral medulla; SI, signal intensity. Adapted from [58]. (Online version in colour.)
In addition to studying the activation of individual brainstem nuclei, exploring their functional connectivity may also inform us about the mechanisms involved in brain–heart interactions. However, brainstem functional connectivity in the human is largely uncharted territory, and this endeavour may also be advanced as the possible gains of UHF fMRI are realized. Recently, Beissner et al. [23] demonstrated patterns of cortical connectivity with distinct brainstem sub-regions in human resting-state fMRI at 3 T, which was achieved by masking areas containing high physiological noise prior to assessing functional connectivity (a technique also suggested in [62]). A subsequent analysis of 7 T data yielded robust detection of brainstem nuclei in single subjects, whereas this detection required multi-subject averaging at 3 T [63]. Functional connectivity of the ventral tegmental area and substantia nigra, structures in the midbrain, displayed resting-state functional connectivity patterns within the brainstem (figure 3) and to various cortical regions at 7 T [55]. Figure 3. Functional connectivity in the midbrain at 7 T. Functional connectivity maps at 7 T in a representative participant, obtained with a three-dimensional fast field echo (FFE) sequence. An uncorrected voxel level threshold of p<10−6 (cluster size = 30 voxels) was used. Seed regions ofinterest in the midbrain include left substantia nigra (L_SN), right substantia nigra (R_SN), left ventral tegmental area (L_VTA) and right ventral tegmental area (R_VTA). Colour bar represents T-statistic. FFE voxel size = 1.33 mm3. Axial midbrain slice sections are displayed at the level of the superior colliculus for N=6 participants. L denotes the left side of the brain and A denotes the anterior portion of the brain. Data underwent motion correction, RETROICOR, band-pass filtering between 0.01 and 0.1 Hz. Robust bilateral spatial distribution of functional connectivity was observed in most ROIs. Adapted from [55]. (Online version in colour.)
Autonomic pathways of the brainstem and spinal cord are integrated with, and mediated by, higher centres in cortex as well as subcortical and limbic regions including hypothalamus, thalamus and amygdala. These structures adaptively modulate cardiovascular activity in response to changing emotional, cognitive and physical demands [50]. Like the brainstem, these regions have characteristics that present formidable challenges for neuroimaging. For one, they are small and themselves contain heterogeneous subdivisions, though the majority of fMRI studies effectively treat them as uniform areas, as it is difficult to resolve their finer structure. This may give rise to apparent variability across the literature in their reported functions and may vastly oversimplify any resulting inferences. Further, ventral brain areas (e.g. amygdala, hypothalamus, nucleus accumbens) suffer from susceptibility artefacts as well as elevated amounts of physiological fluctuation. Advances in UHF fMRI are showing promise in obtaining sensitive measures of neural activity from forebrain areas and in revealing their fine internal organization. Considerable improvement in the BOLD CNR of the amygdala was observed at 7 T compared to 3 T in an emotion discrimination task [64], and the use of smaller voxels has been shown to mitigate susceptibility-induced signal dropout in the amygdala and orbitofrontal cortex [65]. Interestingly, high-resolution imaging at 3 T has also indicated the possible involvement of the amygdala in the default-mode network, an observation not typically made at lower resolutions [66], and has revealed valence-dependent responses to emotional stimuli in the hypothalamus, suggesting a broader role for the latter in emotional processing than is typically recognized [67]. Data-driven clustering methods at conventional image resolutions (3 mm isotropic) at 3 T have suggested that three primary amygdalar subdivisions can be identified on the basis of distinct resting-state connectivity patterns [68], and there is evidence that the performance of connectivity-based parcellation techniques will improve with even higher field [63]. Very recently, resting-state functional connectivity of the bed nucleus of the stria terminalis, a portion of the ‘extended amygdala’ that is implicated in autonomic arousal and anxiety, was mapped in humans at 7 T at 1.3 mm isotropic resolution [69]. In contrast with a similar analysis conducted at 3 T [70], greater specificity was reported in identifying functional connectivity to small structures (e.g. sublenticular extended amygdala, PAG, hypothalamus and habenula) and sub-regions of hippocampus and thalamus. Collectively, these studies suggest that improvements in resolving and measuring neural activity in ventral brain areas may enable novel studies of their involvement in emotional and stress-induced cardiac responses, as well as their integration with other cortical and subcortical regions. Cortical areas, prominently the insular, cingulate and prefrontal cortices, also have a well-established involvement in cardiovascular modulation (reviewed in [18,71,72]). Non-invasive imaging is making exciting headway in extending findings from the animal to the human and in shedding light on the integration between the autonomic nervous system, cognition and emotion. At present, an array of findings point to the complexity of cortical autonomic processing, suggesting that different pathways may be recruited depending on the particular task employed (e.g. affective, cognitive, motor [73]) and also perhaps as a function of individual experience [48]. Meta-analyses are beginning to elucidate the segregation and integration between parasympathetic and sympathetic divisions [73,74] and to interrogate their relationship with large-scale networks such as default-mode and salience networks (discussed in [18,73]; see also [19] with regard to attentional and affective networks). Of particular relevance to UHF fMRI, regions attributed to parasympathetic and sympathetic involvement described in [73] were situated in closely overlapping clusters of anterior insula, angular gyrus and amygdala that may be more precisely delineated with high spatial resolution. fMRI at UHF may convey unique benefits in studying interactions between cortical and subcortical areas. For instance, the thalamus—containing nuclei through which information related to brain–heart interactions is relayed between the cortex and subcortical/spinal cord regions [14,17]—has a well-described functional parcellation in animals that is also supported by non-invasive human studies (e.g. [75,76]), though its nuclei are not commonly differentiated in human fMRI reports. Functional integration of different thalamic sub-regions with distinct cortical and basal ganglia circuits may be studied with greater precision with UHF fMRI [77]. In one study at 7 T, distinct activation of mediodorsal versus centromedian/parafascicular thalamic nuclei was observed in emotional versus anticipatory attention phases of a task; further, these nuclei exhibited distinct co-activation with affective and cognitive divisions, respectively, of the anterior cingulate cortex (ACC) and insula [78]. Although much of the neocortex can be imaged with adequate sensitivity at 3 T and below, examination of cortical subdivisions having differential autonomic involvement, e.g. within the insula [79], may also be better accomplished with UHF fMRI. UHF MRI is a burgeoning component of large multi-site neuroimaging projects, including the Human Connectome Project ([80], which includes a 7 T protocol). Progress towards optimizing imaging hardware, acquisition and data analysis in association with these projects [81] is likely to continue advancing the use of UHF and its role in understanding brain–heart interactions. The growing number of open data repositories are also broadening the accessibility of UHF data ([82,83] and see [5] for further discussion of practical issues concerning UHF scanners). While UHF is not presently established in clinical practice, its advantages over lower field strengths for visualizing certain pathological features offers clear promise for the diagnosis and understanding of disease processes [5,84,85]. As one possible future direction for fMRI, the activity of deep-brain and brainstem nuclei—along with their functional interactions with the rest of the brain—may be contrasted between normal and pathological conditions for insight into how human brainstem abnormalities may underlie autonomic impairments. Preliminary research at 3 T provides evidence that in major depression, a condition marked by reduced heart rate (HR) variability [86,87], autonomic dysregulation may be linked with deficient functional connectivity between autonomic brainstem nuclei and the rostral anterior cingulate [88]. While the enhanced temporal contrast and spatial fidelity of UHF MRI make it an appealing tool for the study of brain–heart interactions, there are a number of counteracting technical and analytic hurdles. Several issues pertaining to the study of brain–heart interactions are discussed below. One challenge in studying large-scale functional interactions with small nuclei involves balancing optimized acquisition of specific brain areas with the desire for broad spatial coverage. Owing to limitations in MRI acquisition speed, high-resolution fMRI of specific brain regions has typically been achieved by restricting the field of view such that it covers only a fraction of the brain. Fortunately, continued progress in hardware and accelerated imaging techniques is enabling the acquisition of high-resolution images with extensive spatial coverage [89,90]. Leveraging advances in gradient and RF hardware along with parallel imaging, De Martino et al. [45] examined resting-state networks as a function of resolution (1–2 mm isotropic) in 7 T data with whole-brain coverage [45]. Beyond demonstrating the detection of canonical resting-state networks at high resolution (with 1.5 mm reported to be the best compromise in their comparisons), higher spatial resolutions were found to yield significantly reduced partial volume effects and display tighter correspondence with the subject-specific anatomic images. However, default-mode network connectivity in frontal brain areas (ACC) was reduced compared to that seen in lower-field data acquired at lower resolution, an effect suggested to result from reduced SNR in the high-resolution data. Indeed, optimal spatial resolutions may depend on the brain region and particular neuroscience question of interest [6]. Also, consistent with observations in the visual cortex by Bianciardi et al. [91], BOLD signals of opposite temporal phase near large draining vessels were identified in the precuneus, suggesting that large-vessel effects are still present at high field and may interfere with interpretation of resting-state fMRI. Localizing activation observed with BOLD contrast fMRI, particularly within zones of small, densely situated nuclei, can be exceedingly difficult and requires precise alignment with an anatomical reference image. Fortunately, structural MRI also benefits from increased resolution and contrast available at UHF (figure 4). Susceptibility-weighted imaging at field strengths of 7 T and above has improved the visualization of structures, including white matter fibres and cortical layers, of the order of hundreds of micrometres [5]. Anatomic parcellation of amygdalar and hypothalamic subdivisons, and delineation of midbrain dopaminergic structures, has also seen improvement at 7 T compared to lower field strengths [93–96]. Efforts to evaluate different anatomic contrasts (and combinations thereof) for in vivo visualization of brainstem nuclei are under way [97,98], as are assessments of optimal processing and alignment strategies for both functional and structural data [99–101]. Recent studies have taken initial steps towards mapping human brainstem nuclei into a standard-space atlas (at 7 T [97]) and towards visualizing neuroanatomic connections between autonomic nuclei in the brainstem and limbic forebrain sites using high-resolution diffusion spectrum imaging tractography (at 3 T [102]). Figure 4. High-resolution anatomical contrast at 7 T. This sagittal image was acquired at 7 T using a two-dimensional gradient-echo sequence and elongated head coil for full brainstem coverage.Spatial resolution is 0.25 mm in-plane (slice thickness = 2 mm), allowing for detailed discrimination of brainstem structures and nuclei. Th, thalamus; AC, anterior commissure; Mi, midbrain; RN, red nucleus; P, pons; Me, medulla. Adapted from [92].
Importantly, the increased severity of distortions at UHF poses a major challenge for aligning functional and anatomic data. Tools for distortion correction are available and under development (e.g. [103,104]). Acquiring T1-contrast echo-planar images with matched distortion can also facilitate registration to anatomical landmarks or serve as an intermediate step between aligning distorted functional images with a non-distorted T1-contrast anatomic reference [105]. The quality of T1-weighted images may also be compromised at high field, due to the increased propensity for B1 field inhomogeneities, though recent improvements in RF coil technology and pulse sequences (e.g. [106]) can greatly ameliorate these issues. Finally, it is possible that standard tools for whole-brain spatial registration may not perform well at high (less than 1 mm) spatial resolution; beyond the lower tolerance of high-resolution data to head motion, true differences in fine-scale functional localization across individuals introduce challenges in assessing results at the group level [107]. Non-neural sources of temporal fluctuation have been estimated to constitute more than half of the total fMRI signal variability at 7 T [108]; also, with BOLD contrast comprising only a few per cent of the raw signal, effect sizes in cognitive neuroscience studies tend to be small, and accounting for non-neural influences can critically influence the outcome of analysis. Below, we discuss three major classes of temporal noise. Physiological noise refers to temporal noise that arises from systemic physiology, primarily from cardiac and respiratory processes (see [56,109] for reviews). Physiological noise is amplified at high magnetic fields and influences fMRI time series through multiple distinct mechanisms. One class of physiological noise results in non-BOLD fMRI signal deflections that are synchronized with the respiratory and cardiac cycles [51,110,111]. Static magnetic field modulations caused by the motion of the chest during breathing will result in voxel shifts or blurring, causing the greatest deflections at boundaries between tissue compartments and around the edges of the brain [112]; the cardiac cycle can induce cerebral pulsation and inflow (T1) effects [51]. As noted above, these artefacts unfortunately affect the brainstem and other regions that reside near vessels and fluid-filled cavities. Retrospective filtering methods [113,114], including RETROICOR [115] and independent component analysis (ICA [116,117]) have shown success in reducing these cyclic, non-BOLD artefacts, even in the brainstem [118,119]. On the acquisition side, artefacts from respiratory motion can be additionally reduced by the use of navigator methods to demodulate dynamic phase shifts [120,121]. Cardiac gating (synchronizing the fMRI image acquisition with the cardiac cycle) is often used to minimize pulsation artefacts, and undesired fluctuations introduced by the (heart-dependent) variability in TR can be normalized by collecting images at multiple echo times per TR (‘multi-echo imaging’) [100,122]. A second form of physiological noise acts directly on BOLD contrast by dynamically altering cerebral blood flow and volume. In other words, BOLD signal changes that are not coupled to the local neural activity (and thereby regarded as ‘noise’) can be induced through systemic respiratory and cardiac modulation of variables such as arterial CO2 and blood pressure. An extreme example is breath holding, which is known to trigger a spatially extensive, large-amplitude BOLD signal change [123] through physiological mechanisms postulated in, for example, [124,125]. The more subtle, inevitable changes in respiration volume and rate (RV [126,127]) (as well as HR [128,129]) across a typical fMRI are also found to induce considerable fluctuation in widespread, grey matter BOLD signal, compromising the ability to detect resting-state networks. These haemodynamic RV and HR effects are, compared to the non-BOLD artefacts described above, more difficult to disentangle from neurally driven BOLD due to their shared spatial and temporal characteristics; linear transfer function models have been found to provide effective nuisance regressors [129–131], though further investigation is needed. The mechanism by which HR correlates with BOLD is not as well understood as that of respiratory modulation. One possibility is that correlations between HR and the BOLD signal arise indirectly from correlations between respiration and HR (e.g. respiratory sinus arrhythmia [132]); it is also plausible that HR acts independently from respiration to some degree, separately modulating blood pressure and haemodynamics. While correlations between HR and BOLD are observed to be quite global in nature [128,129], suggestive of systemic modulation rather than particular neural substrates of cardiac regulation, it is difficult to rule out (or disentangle) the contribution of neural activity. Another possibility is that changes in vigilance and arousal state, which are associated with changes in both HR and global neural activity [133,134], mediate the observed correlation between HR and BOLD. Indeed, physiological noise (particularly that of the BOLD variety) poses an obvious confound for studies investigating neural correlates of autonomic processing. Such studies have employed tasks that actively modulate physiology (e.g. Valsalva manoeuvre [135], volitional breath holding [136] and hand-grip effort [137,138]), and/or they assess relationships between fMRI signals and simultaneously acquired peripheral autonomic measures (e.g. HR variability (HRV, [74,139]), skin conductance [140,141] and microneurographic recordings of nerve activity [59]). Respiratory manoeuvres and physical tasks can elicit large changes in systemic physiology and subject motion that can obscure specific neural responses. Passive monitoring of autonomic variables may reduce such artefacts, but also elicits less autonomic variation and still does not circumvent the question of disambiguating autonomic neural responses from physiological noise (and from activity underlying cognitive/mental effort or skeletal motor activity in performing a task). How can one disentangle non-neural physiological influences on fMRI signals from those implicated in cardio-respiratory control [142]? One possibility for dissociating BOLD ‘noise’ from neural signals arises when these components possess spatial and/or temporal separability. Birn et al. [143] suggested that the temporal dynamics of the artefactual, respiratory-driven BOLD signal modulation may differ enough from the haemodynamic response to neural activity such that they can be disentangled to some extent using linear modelling [143]. The transfer function between RV and BOLD signal time series indeed exhibits longer latency (and primarily of opposite polarity) compared to the haemodynamic response of BOLD to neural activity [129,144]. Dissociation between vascular and neural signals in respiratory paradigms may also be augmented by experimentally manipulating the time course of the former. As two examples: (i) a cued-breathing task was performed in addition to a respiratory challenge at 7 T in an effort to discern brainstem areas related to volitional control of breathing [60]; and (ii) BOLD signal changes from external CO2 administration were contrasted with those correlated with natural fluctuations in end-tidal CO2, with the hypothesis that the coupling between PaCO2 and BOLD would become stronger in respiratory control centres [145]. Spatial ICA appears to capture well the cyclic cardiac pulsation and respiratory artefact into a collection of ‘noise’ components, though is less effective for BOLD respiratory volume/rate artefacts [146], perhaps due to the prevalence of the latter in grey matter and its strong spatial overlap with task- and resting-state networks. ICA in conjunction with a multi-echo pulse sequence can offer a principled way of sorting BOLD and non-BOLD components based on their echo-time dependence (see below), though again demonstrating better performance in dissociating cyclic, non-BOLD physiological noise [118,147] compared to BOLD physiological noise. Examples of non-physiological noise sources include head motion and instrumental drift. Some of these noise sources are very slow and may be suppressed by filtering out very low frequencies or regressing out low-order polynomials [148]; however, the challenge is to remove such nuisance fluctuations without removing BOLD signal of interest. One alternative is to use regressors based on motion parameters or reference regions within the brain that are known to be neurally inactive [149,150]. Another approach is to examine the dependence on echo time: non-BOLD noise should scale with MRI signal strength and decay approximately exponentially with echo time, whereas BOLD signals show a characteristic gamma-variate dependence on echo time [151]. Thus, by performing a multi-echo acquisition, it may be possible to isolate non-BOLD noise sources [152,153]. Recent work has demonstrated that ICA of multi-echo data can offer a practical approach for mitigating non-BOLD sources of fluctuation [118,147,154,155]. It will be interesting to see whether this, and other filtering approaches, will enable the study of autonomic processes that vary too slowly to be resolved with conventional BOLD fMRI, such as tonically active sites of autonomic control. One drawback of the multi-echo approach is the high data acquisition rate, which complicates the acquisition of fMRI scans with high spatial and temporal resolution; the use of simultaneous multi-slice excitation may facilitate this (discussed in [156]). Recent developments in high-dimensional data analysis may, in conjunction with UHF fMRI, uncover novel information about brain–heart interactions and their deviation in pathological states. We briefly discuss two relevant directions. Methods for studying the interactions between brain regions typically assume that temporal covariation is fixed over time, particularly in resting-state fMRI. While static analysis frameworks have yielded tremendous insight into large-scale functional architecture, considerable variation over time can be observed in the properties of resting-state fMRI time series from individual brain regions [157,158], and in metrics of inter-regional coupling [159,160]. Variability in intrinsic activity has been suggested to arise in part from changes in arousal and autonomic processes, underscoring the state dependence of spontaneous brain activity and connectivity [140,161–163]. A better understanding of autonomic influences on intrinsic fluctuations may shed light on previously unexplained patterns of structured brain activity, as well as provide a new window into the neurobiology of brain–body interactions. In addition to mapping regions whose fMRI activity tracks fluctuations in peripheral autonomic indices such as HRV, one may also ask whether time-varying autonomic states are linked with time-varying associations between brain regions. In an exploratory study of this nature [162], spontaneous fluctuations in HRV—whose power in high frequencies (0.15–0.4 Hz) is regarded as an indirect measure of parasympathetic activity—were accompanied by changes in functional connectivity with regard to the dorsal anterior cingulate cortex (dACC) and amygdala. A cluster in the brainstem whose correlation with dACC and amygdala specifically tracked high-frequency HRV was identified; while the low spatial resolution of the data precluded localization to specific nuclei, the analysis framework could certainly be extended to resolve autonomic nuclei in the brainstem and elsewhere with higher spatial resolution. The extent to which time-averaged functional connectivity differences between clinical populations are mediated by varying autonomic states (and corresponding functional connectivity differences) in neurological and psychiatric disorders is also an open question. A number of approaches for analysing time-varying functional interactions have been proposed. Recently, a multivariate technique was described by Liu et al. wherein each BOLD fMRI image (volume) is regarded as a snapshot of co-activation, and representative co-activation patterns (CAPs) are then derived by clustering a set of volumes based on their spatial similarity [164]. CAPs derived from resting-state data revealed transient associations between brain regions that are not apparent from typical, longer time averages; the default-mode network, for instance, appears to co-activate with different areas and sub-networks within salience and executive control networks at different points in time. We may speculate that greater spatial detail revealed by analysing CAPs (and extensions thereof [165]) at higher resolution, and in conjunction with cardiac measurements, may provide a new technique for investigating the network behaviour underlying brain–heart interactions. Machine learning methods are having a rapidly growing impact on the neuroimaging field (reviewed in [166,167]). Distinct from traditional univariate approaches, which query the activity of individual voxels, multivariate methods capitalize on the joint activity expressed over collections of voxels to understand differences in brain states, differences in structure or function between clinical populations, and neural representations of stimuli, mental imagery and dream content. A possible future direction may use multivariate pattern analysis in the study of brain–heart interactions. For instance, within small-volume regions of highly heterogeneous structure, the pattern of activity across a neighbourhood of voxels (i.e. ‘searchlight’ [168]) spanning functional subdivisions may reveal complementary information about how they jointly encode or underlie changes in autonomic state. We may speculate that UHF imaging and increased spatial resolution enhance the information that can be mapped at fine scales with multivariate methods (an idea also suggested and further described in [6,24]). The benefit of higher resolution may depend on the particular brain regions studied as well as dimensionality reduction trade-offs (e.g. using more, lower-resolution brain areas in a classifier compared to fewer high-resolution areas) and the particular classification methods applied. In addition, there are several examples in which discriminative information is shown to persist at coarse spatial scales, the underlying mechanisms of which are under discussion (e.g. [169–172]). Presently, whether high-resolution fMRI can offer advantages for multivariate pattern analysis in assessing brain–heart interactions has yet to be determined; the potential benefits may be countered by significant trade-offs (see [173] and those discussed above), and is a topic currently under investigation (e.g. [174]). High-resolution fMRI at magnetic fields at and above 7 T has yielded remarkably fine-scale measurements of neural activity and structure, and is currently well poised for novel discoveries about interactions between the heart and the brain. The ability to achieve sub-millimetre functional and anatomic resolution in vivo permits the study of small nuclei in areas such as brainstem and forebrain that serve as key integration and relay nodes for bidirectional brain–heart modulation. Advances in accelerated imaging also permit studying their large-scale functional connectivity, which may be further facilitated by drawing upon recent concepts in spatio-temporal data analysis. However, successfully isolating functional signals from small regions, especially in areas strongly contaminated by physiological and susceptibility artefacts, requires considerable effort in acquisition and post-processing. Subject motion can obscure signals from smaller, densely interwoven nuclei (even if the imaging resolution can be further increased), and disentangling temporal noise from autonomic neural activity remains an outstanding challenge. Nonetheless, evidence suggests that UHF MRI will continue to open unique opportunities for investigating brain–heart interactions in the human. We declare we have no competing interests. This research was supported by the Intramural Research Program of the National Institutes of Health, National Institute of Neurological Disorders and Stroke. This material is based in part upon work supported by the National Science Foundation Graduate Research Fellowship under grant no. DGE-1444316. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 3The last three decades have seen the impressive growth of the brain neuromodulation field, which is a form a targeted, reversible, electrical stimulation of the brain able to induce long-lasting changes of firing neural properties, thereby modifying behaviour. Brain neuromodulation, which usually assists—but does not replace—traditional pharmacological treatments, can be achieved either by invasive (deep brain stimulation, DBS), by minimally invasive (vagal nerve stimulation, VNS) or by non-invasive techniques such as repetitive transcranial magnetic stimulation (rTMS) or transcranial direct current stimulation (tDCS). As discussed in detail in another article in this theme issue [1], an extensive portion of the autonomic nervous system is located intracranially: briefly, afferent pathways from the thoracolumbar sympathetic system, to which cardiac afferents belong, converge into the nucleus of the solitary tract in the medulla oblongata. This nucleus is an important relay station, because it receives complex reciprocal connections with the other components of the intracranial autonomic network, and is the place where a sympathetic or parasympathetic outflow is coded in response to a variety of afferent stimuli (i.e. emotional, chemical, attentional, motivational, etc.). This is particularly effective in modulating blood pressure and cardiac rate during physical exercise [2], under the top-down, feed-forward, ‘supervision’ of central commands originating at different levels (i.e. cortical and subcortical) of the central nervous system: functional neuroimaging demonstrated the involvement of anterior cingulate cortex, insula and thalamus in this complex function [3–5]. Therefore, although neuromodulatory interventions are generally targeted on discrete cortical (rTMS and tDCS) or subcortical regions (DBS), or afferent fibres (VNS), the resulting behavioural effect may impact on the function of brainstem centres controlling vital functions, such as the cardiovascular ones (table 1). Each target, indeed, is a relay node of a network, whose activity is somewhat altered by the stimulation, and this perturbation may have trans-synaptic or system-level effects including vital nervous centres. This aspect has been relatively underinvestigated, but deserves attention for the potential relevance of side effects of neuromodulatory interventions, for enhancing our understanding of the cortical/subcortical mechanisms of autonomic regulation of cardiovascular function, as well as for the possibility to design new therapeutic strategies in patients with otherwise intractable autonomic dysfunctions.
Invasive neuromodulation through electrodes placed in different subcortical nuclei or structures, or DBS, is an established therapeutic option for selected patients with advanced Parkinson's disease (PD) [6,7], tremor of different aetiologies [8–10], as well as for other otherwise intractable movement disorders such as dystonia [11,12] and Tourette syndrome [13]. DBS is also an emerging approach to treat pharmacoresistant epilepsy [14], migraine and other chronic pain syndromes, or severe psychiatric disorders such as depression and obsessive-compulsive disorder (see [15] for an exhaustive review). In DBS, extracellular direct currents of variable pulse frequency, intensity and width (20–200 Hz, 1–6 V, 60–120 μs) are chronically applied through reversible leads indwelled into deep grey nuclei and connected via subcutaneous cables to a controllable internal pulse generator [16]. The mechanisms of action of DBS are multiple and complex, and they are not fully known yet: the basic concept is that electrical stimulation, especially at higher frequencies above 100 Hz, resembles the same inhibitory effect of a destructive neural lesion [17–19]. Neurophysiologically, there is clear evidence of either local or system effects: the former are mainly a consequence of excitation/inhibition of both afferent and efferent axonal fibres, rather than the body cell [20]. System effects [21] imply that DBS modifies the dynamics of the whole network connected with the discrete region being stimulated ([19]; see [15,22] for reviews): possible mechanisms accounting for this effect are synaptic inhibition [18] and the so-called jamming that is a masking of pathological oscillatory signals [23]. Although more than 40 different neural targets have so far been described for DBS treatments of several neurological and psychiatric diseases [21], most of the cardiovascular effects have emerged after implants of the subthalamic nucleus (STN) and internal globus pallidum (GPi) in PD patients, of periventricular–periaqueductal grey matter (PV/PAG) in patients with chronic pain syndromes, and of the hypothalamus (H) for cluster headache. The obvious reason is because (i) the great majority of implanted DBS patients worldwide have advanced PD (STN and GPi) and (ii) these targets are spatially close and/or functionally linked to brainstem vital structures. Moreover, the STN and the PV/PAG are relay stations connecting the limbic system with the motor cortex, and change their firing activity in parallel with exercise-related cardiovascular adaptations [24]. Priori et al. [25] showed some of the first evidence that STN DBS in PD patients influences sympathetic and cardiovascular reactivity: while the sympathetic skin response improved, plasma renin increased during DBS-OFF, but arterial blood pressure remained unchanged throughout the upright tilt test duration. Some acute autonomic effects have been observed during the first reglage phase of stimulus parameters after the DBS implant (i.e. the attempt to search the most clinically effective contacts and the threshold for side effects due to stimulus spread to neighbouring neural structures): in 88% of implanted STN PD patients, tachycardia arose within seconds after switching ON the STN stimulation, whereas hypertension arose within about 1 min [26]. Heart rate (HR) and mean arterial blood pressure (MAP) significantly increased after high-frequency (i.e. greater than 90 Hz) thalamic and STN—but not after GPi—stimulation in PD patients: changes were modest (HR increased by about 5±3 beat per min; MAP increased by 5±3 mmHg), but consistent across patients and, most importantly, not related to improvement of locomotor signs, suggesting that DBS of STN, thalamus and sustantia nigra may activate autonomic central commands [27]. These data fit with the notion that postural hypotension may be improved by STN DBS in the long term, through the increase of peripheral vasoconstriction and baroreflex sensitivity, two factors helping to stabilize blood pressure [28]. However, it should be taken into account that patients with STN DBS (but not those with GPi DBS) usually can decrease the daily dosage of dopaminergic therapy, so that hypotension—a frequent side effect of this therapy—might improve independently by stimulation. In another study, the possibility to enhance sympathetic cardiac regulation was shown by means of HR variability spectral analysis, suggesting even a link between electrode positioning within the STN and HR cumulative effects [29]. However, cardiovascular dysautonomia linked with PD seems to be poorly affected by STN DBS [30], at variance with other autonomic disturbances that may instead improve in the long term after STN DBS (but all are small class IV studies; see [31]). When the target of DBS is the PV/PAG, a midbrain nucleus with an important role in pain signalling and autonomic control, effects on blood pressure and cardiac function seem more consistent, at least in patients with intractable chronic pain. This target has been proposed as a new treatment option for postural hypotension, following experimental observations that sympathetic effects on blood pressure (i.e. increase or decrease) depend, respectively, on electrode location in the dorsal or ventral regions of the PV/PAG [32,33], and are possibly mediated by suppression or enhancement of the baroreceptor reflex activity [34,35]. Some human studies seem to confirm this experimental evidence: Green et al. [36] showed that ventral PV/PAG DBS induced a mean reduction in systolic blood pressure of 14.2±3.6 mmHg in 7/11 patients, whereas dorsal PV/PAG DBS caused a mean increase of 16.7±5.9 mmHg in 6/11 patients; interestingly, these findings were accompanied by analogous changes in diastolic blood pressure, but not in R–R variability [36]. Hypotensive effects seem to be long-lasting (up to 1 year) and unrelated to pain relief [37,38]. These results were confirmed in successive studies carried out on patients who were hypertensive before the DBS implant [37–39]. Interestingly, the degree of analgesia induced by DBS of the rostral PAG was linearly related to the magnitude of reduction in arterial blood pressure [39]. The hypothalamus is the target of choice for DBS treatment for refractory cluster headache [40]. The involuntary stimulation of its posterior portion led to increase of blood pressure and respiratory rate in a PD patient [41]. Chronic DBS of the posterior H in chronic cluster headache patients is associated with an enhanced sympathoexcitatory drive on the cardiovascular system during the ‘head uptilt testing’, thereby suggesting that DBS at this level can improve cardiovascular autonomic function especially during orthostatic challenge [42]. In conclusion, all available data on cardiovascular effects of DBS are necessarily limited to patients in whom the cardiovascular autonomic dysfunction was not the primary aim for which DBS was carried out. Depending on the target, effects on blood pressure seem more consistent than those on HR, which are subtle and variable [27,39,43]. Although generalization of effects may be difficult because they are not representative of the general population, they provide an important translational step from the bench to the bedside [44]. The long-term effect of DBS on the ample constellation of non-motor symptoms in PD, including autonomic dysfunctions, is an emerging clinical issue [31]. Although autonomic effects of DBS have been recently regarded as a possible and tantalizing therapeutic opportunity [45], large ad hoc longitudinal studies in which autonomic functions are included among primary clinical endpoints are still required to fully translate this opportunity into a solid therapeutic option. Indeed, the expected benefits should overcome both the intrinsic risks and the limited—albeit present—adverse events of the DBS procedure [46]. VNS refers to any technique that stimulates the vagus nerve, representing another neuromodulatory opportunity of minimally invasive (or non-invasive) brain stimulation. Moreover, at variance with the other brain stimulation techniques (i.e. DBS, rTMS and tDCS), VNS offers a unique example of how stimulation of autonomic fibres may induce bidirectional effects at central (i.e. modulation of brain activity) and peripheral levels (i.e. cardiovascular effects). Through the anatomical and functional links between the vagus nerve and nucleus tractus solitarius, VNS may target diverse and widespread brain regions. Specifically, in animal experiments, VNS elicited synchronized activity in the orbital cortex [47] and slow waves in the lateral frontal cortex, anterior rhinal sulcus and amygdala [48]. Following this former evidence, VNS became a viable treatment option both in neurology and in psychiatry [49–51]. Nevertheless, disadvantages are related to intraoperative risks, such as lesions of the vagus nerve, or to infection, hoarseness, shortness of breath and the requirement for surgical intervention when the battery runs out [51–53]. Other respiratory complications may include vocal cord movement abnormalities, as well as sleep-related breathing pattern changes, with an associated increase in the number of obstructive apnoeas and hypopnoeas [54]. Therefore, an alternative, non-invasive method, called transcutaneous electrical stimulation of the sensory auricular branch of the vagus nerve, was investigated, and central effects verified by means of functional magnetic resonance imaging (MRI) in healthy volunteers [51]. These newer non-invasive VNS delivery systems do not require surgery, and permit patient-administered stimulation on demand; therefore, they improve the safety and tolerability of VNS, making it more accessible and facilitating further investigations across a wider range of uses [55]. VNS has now obtained approval by the US Food and Drug Administration and EU regulatory agencies for treatment-resistant partial onset seizure disorder [56], and treatment-resistant depression [57]. In particular, non-invasive VNS has gained popularity for its ease of use, broad spectrum of efficacy and tolerability in extremes of age and mental capacity [58]. Investigational uses of this technique include antinociceptive effects in chronic pain disorders [59], sympatholytic/vagotonic effects in patients with coronary artery disease [60] and in pentobarbital anaesthetized dogs [61]. However, VNS mechanisms of action are still unknown. Concerning epilepsy, early evidence [62] suggested an anticonvulsant action of VNS in dogs, probably due to increased periods of spike-free intervals. This may reflect the mechanism of action of VNS in achieving seizure control: alternating synchronization and desynchronization of electroencephalogram (EEG), with the latter being progressively the dominant feature [63]. In patients with epilepsy, the long-term efficacy of VNS seems to be maintained or improved [64], whereas the frequency of adverse events generally decreases as patients accommodate to stimulation [65]. The therapy is highly cost-effective, resulting in considerable long-term savings even though most patents do not become totally seizure-free [66]. Of note, VNS therapy was demonstrated to be particularly effective in children with drug-resistant epilepsy, reducing seizure frequency without major safety issues [67]. In these cases, VNS increased respiratory frequency and decreased respiratory amplitude with a variable effect on cardiac activity [68]. During clinical trials of VNS in patients with epilepsy, several investigators reported mood improvements in their patients that occurred independently of the reduction in seizure frequency [69]. Of note, reversible changes in HR variability were observed in patients with major depression after VNS [70]. Although some concerns on its effectiveness were raised in the past [71–73], neurochemical data have demonstrated VNS effects on neurotransmitters thought to be important in mood disorders, including serotonin, norepinephrine, GABA and glutamate [69,74]. Long-term benefits of VNS in treatment-resistant depression were also reported [75,76]. VNS seems to be most effective in patients with low to moderate, but not extreme, antidepressant resistance. Two major working hypotheses of VNS in depression were proposed: the ‘monoaminergic’ and the ‘neural plasticity’ hypotheses of depression [77]. Of note, noradrenergic neurons from the locus coeruleus play an important role in the antidepressant-like effect of VNS [78]. Furthermore, transcutaneous VNS modulates the default mode network in major depressive disorder [79]. Other potential therapeutic indications [80] of VNS are chronic pain, migraine, cluster headache, obesity, chronic tinnitus and Alzheimer's disease [51,80–82]. Of note, VNS may modulate memory formation in humans [83]. Importantly, VNS increased the expression of brain-derived neurotrophic factor and fibroblast growth factor in the hippocampus and cerebral cortex, decreased the abundance of nerve growth factor mRNA in the hippocampus and, similar to the antidepressant drug venlafaxine, increased the norepinephrine concentration in the prefrontal cortex in the rat brain [84]. Transcutaneous VNS also elicited far field potentials from the brainstem [85], as well as inducing production of IL-1β in the brain and activating the hypothalamic–pituitary–adrenal axis [86]. In healthy subjects, VNS shifts the high-frequency power density of heartbeat dynamics, also inducing a partial cardiorespiratory decoupling [87]. VNS, together with rehabilitative therapy, also enhances functional motor recovery after traumatic brain injury in animals [88]. VNS activates neuronal and astrocytes a7nAchR, and inhibits the apoptosis and oxidant stress responses possibly associated with increased Akt phosphorylation and miR210 expression [89]. VNS has also been linked to cardiac diseases [90–92]. These diseases, in fact, still have a high mortality rate owing to neurohormonal activation and autonomic imbalance with increase in sympathetic activity and withdrawal of vagal activity. VNS could be a viable solution to counteract the sympathetic tone and enhance the vagal tone, with applications for heart failure, atrial fibrillation and coronary heart disease induced by increased sympathetic nerve activity [90,91,93]. The improvement in heart failure and the anti-inflammatory and vasodilatory properties of VNS provide additional antiarrhythmic benefit [94]. VNS, in fact, acts on proinflammatory cytokines, nitric oxide elaboration and myocardial expression of gap junction proteins [95,96]. Of note, conventionally, the left-sided cervical vagus nerve is mostly selected as the site for stimulation because of safety concerns [91,93]. Regarding changes in the ECG morphology, elevated levels of T-wave alternans were found in patients with drug-refractory partial-onset seizures following VNS [97]. Moreover, pure ectopic cycles were observed during sinus arrest caused by VNS [98]. Remarkably, this effect has been suggested to profitably discern between parasystole and extrasystole events in case pure ectopic cycles are not spontaneously observed [98]. In summary, the acute and long-term efficacy of VNS, especially in treatment of epilepsy and depression, is encouraging, but still under debate. The effectiveness of non-invasive transcutaneous VNS for epilepsy, depression and other conditions has not been investigated beyond small pilot studies [99]. Transcutaneous VNS deserves further study as an antidepressant therapy and for its potential effect on physiological biomarkers associated with depression morbidity and mortality, especially because the exact mode of action of VNS is still not well understood [100–102]. Furthermore, given the complexity of the cardiac autonomic nervous system, including the presence of both afferent and efferent as well as parasympathetic and sympathetic fibres in the vagosympathetic trunk, it is critical to carefully elucidate the mechanisms of VNS and the parameters of stimulation to ensure that VNS achieves the desired therapeutic effect [94,103]. Nevertheless, nowadays, left cervical VNS is an approved therapy for refractory epilepsy and treatment-resistant depression, whereas right cervical VNS has proven promising for treating heart failure, along with reducing malignant ventricular arrhythmias, particularly in the setting of ischaemia, as well as atrial arrhythmias. Repetitive TMS is a 25-year-old non-invasive neuromodulatory technique [104] by which electric currents generated by a rapidly varying magnetic field (up to 2 T) applied on the scalp through a coil can reach the superficial layers of the cortex. At this level, neural elements (mainly interneurons) are activated trans-synaptically [105]. Trains of single pulses may be applied regularly spaced in time or in a patterned way: both types of rTMS produce changes in cortical excitability outlasting the stimulation time. Typically, high-frequency (greater than or equal to 5 Hz) rTMS increases cortical excitability, whereas low-frequency (less than or equal to 1 Hz) rTMS reduces cortical excitability [106]. In the patterned theta burst stimulation (TBS), three pulses at 50 Hz, repeated at 5 Hz (i.e. every 200 ms, or theta frequency) are applied: this produces a longer lasting after-effect [107], with a resulting inhibitory net effect if applied continuously for 40 s (cTBS), or gives rise to enhanced excitability if applied intermittently (iTBS, 2 s of stimulation every 10 s). Sustained changes of cortical excitability can be obtained thanks to a modification of the synaptic efficacy, implying the occurrence of long-term depression or potentiation mechanisms, probably via an NMDA-receptor-dependent mechanism (see [108] for a review). The effects of rTMS take place not only locally, but involve a network-mediated modulation of regional brain activity, either at cortical or even subcortical levels [108]: for example, there is evidence that stimulation at the level of the primary motor cortex (M1) or the dorsolateral prefrontal cortex (DLPFC) induces, respectively, suppression of beta oscillatory activity in the STN [109] and modifies striatal binding at dopaminergic terminals [110] in PD patients. The impact of TMS-induced electric field on local structures—as well as distributed anatomical and functional networks—has been investigated using various approaches, ranging from initial spherical models of the head roughly calculating the amount and directionality of injected current [111], to those including brain structural information derived from MRI [112,113]. Recent studies have even attempted to model the impact of cortical gyration at the individual level [114], with results suggesting the field strength as being significantly enhanced when the currents run approximately perpendicular to the local gyral orientation [115,116]. This evidence, implying an even more careful selection of coil orientation in order to realistically target given cortical areas, has been recently expanded by studies focusing on the role of tissue anisotropy (using diffusion tensor (DTI) and diffusion-weighted (DWI) imaging) to determine the probability to target specific white matter tracts [115,117], thus directly perturbing the structural ‘connectome’. Finally, resting-state functional MRI brain networks [118] have also recently been considered, showing how changes in coil orientation and stimulation site location—even within the same anatomical region (e.g. DLPFC)—might lead to the engagement of completely different networks (i.e. default mode versus frontoparietal networks) [119]. Overall, current evidence strongly suggests the need to rely on multimodal imaging efforts in order to reliably assess the impact of TMS, which should be adopted for the modulation of regions related to autonomic functions. With these neurophysiological and modelling premises, rTMS has gained consensus among neurologists and psychiatrists as a non-invasive neuromodulatory intervention to treat several diseases associated with regional or diffuse dysfunctions of cortical excitability: a recent evidence-based survey [108] indicates that rTMS is definitely effective (level of evidence A) in the treatment of chronic pain syndromes when applied at high frequency on the contralateral M1, and on pharmacoresistant depression when applied at high frequency on the left DLPFC. A level B of evidence (i.e. probable efficacy) has been proposed for the antidepressant effect of low-frequency rTMS of the right DLPFC, high-frequency rTMS of the left DLPFC for the negative symptoms of schizophrenia, and low-frequency rTMS of contralesional motor cortex in the recovery phase from a chronic motor stroke. As there is experimental evidence in animals that stimulation of the M1 and of the DLPFC (i.e. the cortical targets which reached a level A of evidence for rTMS interventions; figure 1a) may influence cardiovascular autonomic functions [120,121], and that both cortical regions showed associations with sympathetic activity in humans [122,123], similar effects could consequently be expected in subjects who underwent several sessions of rTMS on these cortical targets. Some early studies confirmed this hypothesis: for example, Udupa et al. [124] treated daily for two weeks a group of 27 depressed patients by high-frequency rTMS of the left DLPFC, for a total of 18 000 pulses. HR variability measures indicated that rTMS produced significantly greater reduction than serotonergic agents (taken by a second group of 25 patients) in the sympathetic/parasympathetic ratio, suggesting improvement in sympathovagal balance [124]. However, a recent review on the topic failed to demonstrate conclusive evidences that rTMS might impact on autonomic function at a clinically useful degree [125]. This might be due to several, non-mutually exclusive factors: the heterogeneity of stimulation parameters (cortical target, frequency intensity and duration of stimulation), and patients' characteristics and design of the studies not including a systematic evaluation of cardiovascular variables among their primary or secondary aims. Figure 1. Sites of stimulation of rTMS and tDCS montages, with effects on the brain. (a) The most commonly reported rTMS stimulation sites for targeting cortical regions related to autonomic functions. (b) A bipolar tDCS montage, with anode and cathode electrodes both placed on the scalp, roughly in correspondence of right M1 and left frontopolar cortex. As shown, the result of tDCS is a constant current flow between the two stimulation sites, with current both bridging between the electrodes through the skin, as well as penetrating the scalp (c). tDCS provides a less focal stimulation effect (areas in red and cyan) underneath each electrode, with the highest current density expressed in brain regions in between the two fields. (d) Extracephalic solutions involving one electrode on the scalp and one on ipsi/contralateral arm/shoulder/neck can be adopted in order to facilitate the current flow towards brainstem structures. (Online version in colour.)
Although long-lasting autonomic effects of rTMS are substantially lacking, there is evidence that rTMS may alter HR during its application: stimulation of M1 induced a short-lasting increase of HR and blood pressure [126,127]; even in vegetative state patients, high-frequency rTMS can transiently increase HR [128], suggesting that rTMS of M1 may modulate the autonomic outflow in the absence of motor responses (and unspecific arousal). Low-frequency repetitive TMS, particularly after stimulation of the right hemisphere, induced a slight increase in the parasympathetic drive (i.e. significant bradycardia) and no effects on the sympathetic outflow in healthy subjects [129]. However, this was not the case following acute high-frequency rTMS of the left DLPFC [130]. In conclusion, available data do not allow conclusive statements regarding clinically relevant effects of rTMS on autonomic regulation of cardiovascular function. The effects of brain stimulation in healthy subjects on the sympathetic outflow posit the need to design ad hoc studies to confirm the encouraging preliminary results, possibly based on neuroimaging investigations coupled with monitoring of vital parameters. Different from rTMS, tDCS considers a transcranially delivered constant electrical field aimed at inducing instantaneous (‘online’) and long-lasting (‘offline’) changes in cortical excitability levels by means of a cellular membrane polarization process [131,132]. Originally applied through two saline-soaked electrodes placed on the scalp and with intensities in the range 0.5–2 mA, tDCS has demonstrated significant modulation of the physiology of several brain systems. The effects of tDCS spread from the modulation of cortical excitability levels [133] up to high-order cognitive networks [134,135], with initial promising results also for the treatment of neurological and psychiatric conditions [136–138]. The application of a constant field on the scalp generates two electric ‘poles’ with opposite charge, constituted by the two electrodes. During anodal tDCS, the current delivered on the ‘anode’ attracts negative ions in the tissues underneath the electrode: this reduces the resting membrane threshold, thereby facilitating neuronal firing (i.e. it increases cortical excitability). On the contrary, negatively charged electrodes (i.e. cathodal tDCS) affects nearby regions by attracting positive charges, increasing the threshold and making the stimulated area less prone to be activated in response to exogenous or endogenous stimuli (i.e. reduction of cortical excitability) (figure 1b). While such a basic mechanism reflects the simplicity of application for tDCS, it also highlights the potential issues related to the estimation of the brain tissues actually being stimulated. Electrical charges leaving each electrode follow diverse routes in their path towards the oppositely charged electrode, passing through biological tissues with different conductivity properties (e.g. skin, muscle, bone, cerebrospinal fluid, grey and white matter, blood vessels). Modelling studies have repeatedly shown how the effect of tDCS is highly dependent on electrode montage, and how an accurate positioning of tDCS electrodes might allow one even to stimulate subcortical structures [139,140]. This made realistic the possibility of targeting both cortical and deep brainstem structures related to autonomic functioning, with interesting scenarios related to the understanding of brain–heart interactions, as well as potential treatment of pathological conditions [141] (figure 1c,d). There is evidence of autonomic effects of brain stimulation on animals [121,142], with electrical stimulation applied to motor and premotor areas in rat, cat and monkeys eliciting cardioautonomic responses [120], as well as stimulation of different regions of the insular cortex triggering cardiovascular responses consistently [142]. Despite this promising scenario and modelling work suggesting its feasibility in humans, just a handful of studies have investigated the application of tDCS for the modulation of cardiovascular human functions [125], with current literature showing very conflicting results [143,144]. Most of the studies involved the delivery of tDCS using a classic bipolar, cephalic electrode montage (i.e. both anode and cathode are placed on the scalp) targeting M1 and DLPFC, but autonomic function was also tested with tDCS delivered on temporal regions [145]. Whenever documented, positive autonomic effects of tDCS were related to anodal stimulation, with only two studies supporting effects for cathodal tDCS delivered on the DLPFC [146] or on M1, the latter addressing vasomotor reactivity changes [147]. However, only two studies were designed ad hoc to investigate the effect of tDCS on brainstem structures through a unipolar scalp montage linked with an extracephalic (arm/shoulder) electrode positioning: both studies, one targeting M1 [148] and the other one targeting the frontal midline [144], suggest null or mild effects on cardiovascular and respiratory functions, with a progressive shift over time in favour of the sympathetic tone, a finding that however was present during sham (i.e. placebo) tDCS. Thus, the small number of studies and the heterogeneity of tDCS parameters (e.g. montage, stimulation intensity and duration, monitoring length) do not allow for depicting a definitive scenario for the effect of tDCS on autonomic function. Despite this, several points can be drawn to inform future investigations: first of all, just a few studies attempted to monitor for long-lasting effects of tDCS stimulation, with the vast majority of the investigations focusing on online effects recorded during stimulation or changes happening just a few minutes after the stimulation was turned off. Given that the effect of a classical tDCS session (i.e. 20 min of stimulation) seems to last for almost 1 h, and that extracephalic stimulation has shown a different timing of cortical excitability modulation [148], future investigations should include longer post-tDCS recordings. Indeed, as shown in Santarnecchi et al. [148], changes in heart-rate variability and systolic/diastolic pressure seem to correlate with the time course of spontaneous motor cortex excitability during sham tDCS, with a strengthening of such dependence during anodal tDCS of the motor cortex. This suggests the possibility to use tDCS as a tool to uncover the causal link between excitability of specific regions of the cerebral cortex and autonomic functions, if appropriately designed studies are able to monitor their interplay at high temporal resolution. In conclusion, available studies demonstrated some potential of brain stimulation techniques in modulating cardiovascular function acutely, with most consistent effects in the long term shown for DBS of the PAV/PAG on blood pressure control (table 1). However, the full clinical applicability of these neuromodulatory approaches in patients with otherwise intractable autonomic dysfunctions will remain a ‘therapeutic opportunity’ [44,45] until such time as autonomic variables are considered as primary or secondary aims in large prospective clinical trials of neuromodulatory therapies, even if performed on patients in which the autonomic dysfunction is not the primary reason for the neuromodulatory intervention. There is no additional or supplementary data. S.R. wrote the DBS section, drafted and supervised the whole paper; E.S. drafted the tDCS section; G.V. drafted the VNS section; M.U. drafted the TMS section. All the authors made a substantial contribution to the conception of the paper and gave final approval of the version to be published. The authors declare no financial or non-financial competing interests. We received no funding for this study. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 4In the emerging field of network physiology [1], the human body is seen as an integrated network composed of different organ systems, which have their own internal regulatory mechanisms, but also interact with each other to preserve the physiological function. In addition to other important physiological systems subjected to neural regulation, such as the circulatory and respiratory systems, two crucial nodes of the human physiological network are the brain and cardiac systems. The dynamical states of these systems are continuously modulated by the rhythmic activity of brain structures devoted to visceral control (i.e. the autonomic nervous system). This modulation is clearly visible in the time course of the physiological variables measured as output of the cardiac and central brain systems. For instance, the analysis of low-frequency (LF) and high-frequency (HF) oscillatory components of heart rate variability (HRV) is ubiquitously used to assess autonomic control in a variety of physiological conditions and pathological states [2]. The continuous modulation of brain activity is reflected in the amplitude of electroencephalographic (EEG) rhythms oscillating in the δ, θ, α, σ, β, and γ frequency bands that cover the whole spectral EEG variability. In particular, the modulating activity of the autonomic function is known to have an impact on the five frequency bands that characterize in detail the EEG during sleep (i.e. δ, θ, α, σ and β) [3]. Sleep is a physiological state that has a significantly more complex and important impact on the neural regulation of cardiac and cerebral physiological variables [3–6]. The stage organization of sleep reflects modulations in the autonomic activity such that an increase in the balance between the variance of the autonomic components of HRV in the LF and HF bands (LF/HF ratio) is known to occur with the transition from non-random eye movement (NREM) sleep to REM sleep [4,7]. Sleep state-related variations are also observed in the spectral power of the sleep EEG, with wave amplitude decreasing significantly in the slower δ, θ and α bands, and increasing in the faster σ and β bands, during REM sleep compared with NREM stages [3,8]. In more recent years, the research on brain and cardiovascular dynamics during sleep has seen a shift of paradigm from the study of individual EEG or HRV activity to the investigation of the relationships between specific EEG wave amplitudes and HRV dynamical indexes. Initially, brain and cardiovascular interactions during sleep were studied only indirectly, correlating indexes of cardiac and cerebral variability to each other through simple non-dynamical analyses [9–14]. Dynamic approaches to the joint characterization of cardiac and neural time series were introduced only recently, and were based on bivariate and non-causal analyses [15–22]. However, complex physiological networks are composed of multiple nodes, where each node may represent the activity of a specific physiological system that exhibits autonomous dynamics but is also connected to other diverse systems. A proper analysis of these networks necessitates the introduction of fully multivariate and causal methods for time-series analysis, to guarantee a faithful reconstruction of the network structure based on the detection of direct and directional effects. To meet this need, we have recently proposed an integrated framework for multivariate time-series analysis, essentially based on information-theoretic and predictability measures, that led us to evidence for the first time the existence during sleep of a structured network of brain–heart interactions, sustained both by the internal dynamics of the various cardiac and brain processes and by the causal interactions between them [23,24]. In this study, we pursue the twofold aim of completing the methodological formulation of this framework, and of advancing its utilization in the clinical description of sleep disorders. Methodologically, the analysis of how information is processed inside a network of multiple interacting processes is often performed under the perspective of distributed computation, whereby the general concept of ‘information processing’ is dissected into the basic components of information storage, transfer and modification [25]. Information storage and transfer refer respectively to how much the uncertainty about the present state of a dynamical system can be resolved by the knowledge of its own past states [26], and by the additional knowledge of the past states of the systems potentially connected to it [27]. Following our recent developments in the detection of information storage and information transfer in brain–heart networks [23,24], in this study, we embed in the framework the concept of information modification, relevant to how two (groups of) source systems interact redundantly or synergistically with each other when they contribute to resolve the uncertainty about the states of the assigned target system [28,29]. The complete framework is formulated, in the context of multivariate linear prediction models, devising a predictability decomposition strategy that leads to decompose the full predictability of the target process into measures of self-predictability, causal predictability and interaction predictability which reflect the notions of information storage, transfer and modification. This opens the way to the thorough investigation of how different cardiac and brain processes retain the information that they produce, transfer information to each other, and mutually interact while they transfer information. From a clinical point of view, a pathological condition that leads to a re-organization of the network of physiological interactions between the cardiovascular and brain rhythms is the sleep apnoea–hypopnoea syndrome (SAHS). This syndrome is a common medical problem that dramatically affects the quality of life, and is associated with hypertension, heart failure, myocardial infarction, stroke and vascular complications [30,31]. SAHS has been associated with significant alterations of the rhythmic autonomic activity during sleep, with blunted shifts of the alternating predominance of cardiac vagal and sympathetic activities during NREM and REM sleep [32], increased δ EEG activity during NREM apnoea [33], and impaired link between cardiac parasympathetic and delta EEG activities [34]. Moreover, the treatment of severely apnoeic patients with nasal continuous positive airway pressure (CPAP), a respiratory therapy which consists in applying mild air pressure to keep the airways continuously open during night-time, has been proven effective in improving sleep architecture and cardiovascular parameters, as well as in reducing some comorbidities of SAHS [35–37]. It has been recently shown that long-term CPAP therapy produces a subtle but clinically important restoration of the neuroautonomic joint regulation of different organ systems: in the presence of similar sleep characteristics and spectral profiles of cardiac variability, patients with severe SAHS exhibiting a weakened link between the cardiac vagal component of HRV and the sleep delta EEG amplitude partially recovered such a link after more than 1 year of nasal CPAP treatment [38]. In this study, we delve into these modified brain–heart interactions, according to the broader perspective of multivariate analysis of network processes, through the evaluation of the patterns of self-predictability, causal predictability and interaction predictability. Let us consider a multivariate stochastic process Ω={Y,X} composed by a predefined ‘target’ process, Y , and by M other possibly interacting processes, X={X1,…,XM}, which are considered as ‘sources’. In this context, we deal with the description of the dynamics of the target process performed in the framework of linear prediction. Setting a temporal frame where n represents the present time, the random variable associated with the present of the target, Yn, is described as resulting from a linear combination of the p past target variables, Yn−1,…,Yn−p, and of the past variables of the source processes, Xm,n−τm,…,Xm,n−p (m=1,…,M), according to the autoregressive (AR) model of order p and with M inputs 2.1 where Ak and Bmk are linear regression coefficients, τm is the minimum delay of the interaction from Xm to Y , and WΩ is a scalar zero-mean innovation process uncorrelated and X1,…,XM. The variance of WΩ, denoted as ε(Y |Ω)=ε(Y |Y,X), is a measure of the unpredictability of Y given Ω={Y,X}, and is bounded between 0 and the variance of Y , ε(Y).Starting from the most complete model structure of equation (2.1), simplified structures that disregard some of the source processes can be obtained as follows. Disregarding all possible sources, the description of Y can be performed by a simple linear AR model 2.2 where the regression coefficients2.3 In equation (2.3), the variances of the innovation processes WYV and WYZ, denoted as ε(Y |Y,V) and ε(Y |Y,Z), can be taken as measures of the unpredictability of the present of Y given its past and the past of V, or given its past and the past of Z, respectively.Based on the representations above, we provide the following definitions of full predictability of Y , self-predictability of Y and causal predictability from X to Y 2.4 2.5 2.6 The full predictability PY |Ω quantifies the portion of the variance of the target process Y that can be predicted from the knowledge of the whole set of available processes Ω={Y,X}. Similarly, the self-predictability PY |Y quantifies the portion of the variance of Y that can be predicted from the exclusive knowledge of its own dynamics. These predictability measures are inversely related to the complexity of the target time series, which is defined in this study as the degree of unpredictability of the series measured through the prediction error variance [39]. The causal predictability PX→Y measures the portion of the variance of Y that can be predicted from the knowledge of the source process X above and beyond the portion that can be predicted from Y considered alone. Quantifying predictability improvement, this last measure is in agreement with the concept of Wiener–Granger causality [40,41]. Combining equations (2.4)–(2.6), one can easily show that the measures of full, self- and causal predictability are related to each other by the equation: PY |Y,X=PY |Y+PX→Y (figure 1). Moreover, separating the source process X into the subsets V and Z, we define the partial causal predictability from V to Y given Z as2.7 Quantifying the portion of the variance of Y that can be predicted from the knowledge of V above and beyond the portion that can be predicted from Y and Z, the measure defined in equation (2.7) reflects the concept of (multivariate) partial Granger causality [41–44].Figure 1. Variance decomposition for the multivariate process Ω={Y,X}= {Y,V,Z}. The diagrams show how the variance of the target process Y (ε(Y), total area) is decomposed in an unpredictable part (ε(Y |Ω), grey area) plus a predictable part PY |Ω, denoting full predictability. The latter further splits in two parts evidencing the self-predictability of the target (PY |Y, red area) and the causal predictability from all sources to the target (PX→Y, yellow+white areas). Then, the overall causal predictability can be decomposed: (a) as the sum of the partial causal predictabilities (PV→Y |Z and PZ→Y |V, yellow) plus the interaction predictability (IZ,V→Y, white) in the case of redundancy or (b) as the sum of the causal predictabilities (PV→Y and PZ→Y, yellow) minus the interaction predictability (−IZ,V→Y, white) in the case of synergy. (Online version in colour.)
The concepts of causal predictability and partial causal predictability can be combined together to investigate how the sources interact with each other in the prediction of the target dynamics. Specifically, defining a so-called interaction predictability as 2.8 it is easy to show that it quantifies the difference between the causal predictability and the partial causal predictability from one source process to the target process as2.9 Then, one can show that the ‘collective’ causal predictability from X={V,Z} to Y can be decomposed as2.10 The predictability decomposition of equation (2.10) is illustrated graphically in figure 1, showing how it allows to quantify the character (redundant or synergistic) of the interaction between V and Z when they are used to predict Y [45]. Specifically, if the causal predictability from V to Y is larger than the partial causal predictability from V to Y given Z, the interaction predictability is positive. In this case, we have that considering the two sources together yields a worse prediction of the target than summing up the individual predictabilities from each source (PX→Y<PV→Y+PZ→Y); this situation denotes redundancy between the sources (figure 1a). If, on the contrary, the causal predictability from V to Y is smaller than the partial causal predictability from V to Y given Z, the interaction predictability is negative. In this case, we have that considering the two sources together yields a better prediction of the target than summing up the individual predictabilities from each source (PX→Y>PV→Y+PZ→Y); this situation denotes synergy between the sources (figure 1b).It is worth noting that the predictability measures introduced above are strictly related to a set of corresponding information-theoretic measures defined for vector AR processes. Specifically, exploiting the close correspondence between conditional entropy and prediction error variance that holds for linear processes, one can show that the measures of full predictability, self-predictability, (partial) causal predictability and interaction predictability defined in this study are directly related to the linear AR definition of the measures of prediction entropy [46], self-entropy [46], (partial) transfer entropy [43,46] and interaction transfer entropy [47], respectively. The theoretical concepts underlying these measures and their practical implementation in simulated time series have been thoroughly tested in a number of recent studies [43,46–48]. This study considers a database previously collected [38], including eight patients (45.6±6.7 years) who were admitted in the Sleep Laboratory for two consecutive nights and were diagnosed as suffering from severe SAHS, detected measuring an apnoea–hypopnoea index (AHI) more than 30 events per hour. The same patients were admitted again (49.9±7.0 years) after at least 1 year of treatment with the CPAP therapy (range: 1.8–7.8 years; mean±s.d.: 4±3 years), which was maintained for at least 5 h every night and led to adjustments documented by the significant reduction of apnoea episodes (AHI<10 events per hour). Moreover, 14 healthy age-matched subjects (44.0±6.2 years), admitted to the Sleep Laboratory for three consecutive nights, were considered as the control group. Participants were not allowed to take any medication, and were instructed to limit alcohol and caffeine consumption and to respect a regular sleep–wake cycle. We analysed the ECG and EEG recordings of the SAHS patients, both before and during CPAP treatment, and of the healthy controls, acquired during the second or third night of their hospitalization. The measurement procedure resembled that of our previous studies [23,24] and consisted in the following processing of ECG and EEG signals, digitized simultaneously with 200 Hz sampling frequency and 12-bit amplitude resolution. The analysis of the EEG recordings (Cz-Ax derivation, with Ax mastoid reference) was performed first applying a fast Fourier transform (FFT) to each consecutive window of 5 s, and then computing the spectral power inside each of the five conventional frequency bands (δ: 0.5–3 Hz; θ: 3–8 Hz; α: 8–12 Hz; σ: 12–16 Hz; β: 16–25 Hz) traditionally explored in sleep studies [1,3,15–17,34,49]. In this study, we stick to these five bands, excluding the dynamics of the EEG amplitude in the γ band (greater than 25 Hz) which are commonly associated with sensory and cognitive functions and during sleep are partly synchronous with β dynamics (γ1 band, 25–35 Hz) and partly correlated only to muscle tone artefacts ((γ2 band, 35–45 Hz) [50]. The time-series representative of the δ, θ, α, σ and β wave amplitudes were finally obtained averaging the power values for the relevant frequency band over non-overlapping windows of 60 s. The ECG (lead V4 or V6) was first up-sampled to 400 Hz to increase precision in the location of the R-peaks, which were used to measure HRV through the sequence of the durations of the time intervals between consecutive R-peaks (RR intervals). Premature ventricular contractions, ectopic beats and other artefacts were automatically detected when RR<0.35 s or RR>1.5 s, and were then removed and linearly interpolated with the surrounding values. The resulting RR time series was interpolated and resampled uniformly to 8 Hz, and then subdivided in consecutive windows of 120 s overlapped by half. For each window, the RR interval series was in turn detrended, Hanning windowed and FFT transformed. Finally, the time-series representative of the cardiac parasympathetic activity was obtained as the sequence of the spectral power values taken in the HF band (0.15–0.4 Hz) divided to the power contained in the LF+HF band (0.04–0.4 Hz). With this overall procedure, six synchronous time series, describing the variability of the five brain wave amplitudes and of the cardiac parasympathetic component, were obtained with sampling frequency of 1 min. In this way, one multivariate time series was collected from the full night polysomnographic recordings of each individual subject. Before predictability analysis, each time series was normalized to zero-mean and unit variance. The six normalized time series measured for each subject were then considered as realizations of the overall stochastic process Ω={Φ,η}, composed by the five-dimensional vector process Φ={δ,θ,α,σ,β} describing the dynamics of the different brain rhythms and by the scalar process η describing the cardiac dynamics. The analysis was then performed, following the derivations presented in §2, as described in the following. In the study of cardiac dynamics, the role of the target Y was assumed by the cardiac process η, whereas the sources X were represented by the brain process Φ. We computed the full, self- and causal predictability measures Pη|Ω, Pη|η, and PΦ→η; further, causal predictability decomposition was performed computing the interaction predictability between each single brain process and the remaining brain processes as Ix,Φ∖x→η, with x taking the role of any process δ, θ, α, σ or β. In the study of brain dynamics, each single brain process was considered as the target and the full, self- and causal predictability were computed as Px|Ω, Px|x, and Pη,Φ∖x→x; predictability decomposition was then performed computing the interaction predictability between the cardiac process and the brain processes other than the target as Iη,Φ∖x→x, (here x was equal to δ, θ, α, σ or β). All indexes were computed from the prediction error variances obtained from the linear regressions described in equations ((2.1)–(2.3)). Regressions were performed through the standard least-squares approach, after optimizing the model order p in the range from 2 to 12 according to the Bayesian information criterion (BIC) applied to the full multivariate AR model fitting the six series. The reduced model structures were separately identified from the data using the same model order obtained through the BIC for the full structure. In the study of cardiac dynamics, the minimum interaction delays τm were set equal to 1 for all sources. In contrast, in the study of brain dynamics, the minimum interaction delay from the cardiac source process η to the considered brain target process x, with x=δ, θ, α, σ or β, was set to 0 to account for the partial overlap of the time series owing to the measurement convention, which introduces ‘causal’ information with no delay from heart to brain (i.e. because the HF power of HRV was measured on windows of 120 s overlapped in the second half with the measurement windows of the EEG bandpowers, the nth sample of the cardiac series is built including HRV points that occur in time before the EEG points forming the nth sample of any brain series) [24]. To test the hypothesis that an individual observed process (i.e. the cardiac rhythm η or one of the brain rhythms δ, θ, α, σ, β) is significantly predictable given either the past of the full process Ω, its own past, or the past of all other processes, we assessed the statistical significance of the measures of full, self- and causal predictability (equations (2.4)–(2.7)); this was performed for each individual subject applying the Fisher F-test with significance p<0.01: the test compares the variances of the residuals obtained from the unrestricted and restricted model structures relevant to the model-based computation of the considered predictability measure. To test the hypothesis that the brain rhythms interact significantly with each other while they contribute to improve the predictability of the cardiac rhythm, and the hypothesis that the cardiac rhythm interacts significantly with the brain rhythms while contributing to improve their predictability, we assessed the statistical significance of the interaction predictability measure (equations (2.8) and (2.9)) on a group basis, by testing whether the causal predictability and partial causal predictability measures computed across subjects come from a distribution with the same median; this was done using the Wilcoxon signed-rank test for paired data, and correcting for multiple comparisons (i.e. setting statistical significance p<0.01). To test the hypothesis that each assigned predictability measure varied significantly when computed before and during the CPAP treatment, we used the Wilcoxon signed-rank test for paired data with significance p<0.05. To test the hypothesis that each assigned predictability measure varied significantly between untreated SAHS patients and controls, or between patients treated with CPAP therapy and controls, we used the Mann–Whitney U-test with significance p<0.05. The length of the time series analysed in this study (from the onset of sleep until the end of the last REM sleep state) was 418±60 min for the control group, 419±55 min for the SAHS patients and 410±81 min for the same patients studied after prolonged CPAP treatment. Complete sleep and demographic parameters for these groups are reported in [38]. The model order estimated for the full multivariate model using the BIC was 3.52 (average for all groups and time series), and did not vary significantly across the three groups (CTRL: 3.4±1.3 for η, 3.8±1.9 for Φ; SAHS: 3.9±2.6 for η, 3.5±1.8 for Φ; CPAP: 2.5±0.5 for η, 3.2±1.9 for Φ). The results of the predictability decomposition of the cardiac dynamics are reported in figure 2, showing the distribution in the three groups of the full predictability of the process η and of its constituent terms, i.e. the self-predictability and the causal predictability derived from the brain process Φ. In the group of healthy subjects, about 50% of the variance of the cardiac dynamics during sleep could be predicted from their past and from the dynamics of the brain processes. This full predictability was significantly lower (36.9%, p=0.007) in patients suffering from SAHS, and remained at lower levels (38.4%, p=0.012) also after the CPAP treatment. This higher complexity exhibited by apnoeic patients compared with healthy controls was the result of a decrease of both the self-predictability of the cardiac dynamics and of the causal predictability brought by the brain dynamics. The indexes of full and self-predictability were statistically significant in all groups, whereas the causal predictability was not significant in three out of eight patients before CPAP treatment, and one patient after the treatment. Figure 2. Distribution of the measures of full predictability (Pη|Ω), self-predictability (Pη|η) and causal predictability (PΦ→η) assessed for the cardiac parasympathetic component of HRV in the network Ω={Φ,η} formed by cardiac process η and the brain process Φ, and expressed as median and quartiles computed in healthy controls (CTRL), in patients with severe sleep apnoea–hypopnoea syndrome (SAHS), and in the same patients studied after prolonged continuous positive airway pressure therapy (CPAP). For each measure, the number of subjects for which it was detected as statistically significant is reported inside the bar. *p<0.05 CTRL versus SAHS or CTRL versus CPAP (Mann–Whitney U-test).
Figure 3 reports the distributions of the measures of interaction predictability reflecting how each single brain process δ, θ, α, σ or β interacts with the other brain processes in the prediction of the cardiac process η. In the group of healthy subjects, the θ, α and β brain wave amplitudes interacted redundantly with the other EEG waves (the interaction predictabilities Iθ,Φ∖θ→η, Iα,Φ∖α→η and Iβ,Φ∖β→η were significantly different from zero). This redundant interaction disappeared in SAHS patients, and was restored after CPAP treatment. Notably, in the SAHS group, the interaction predictability was never significantly different from zero, documenting the absence of any significant interaction between the brain processes while they contribute to the prediction of the cardiac dynamics. Figure 3. Box plots of the measures of interaction predictability, Ix,Φ∖x→η, quantifying the degree of interaction between each brain process (x=δ, θ, α, σ, β) and the remaining brain processes in the prediction of the cardiac process η, computed in healthy controls (CTRL), in patients with severe sleep apnoea–hypopnoea syndrome (SAHS), and in the same patients studied after prolonged continuous positive airway pressure therapy (CPAP). For each measure, significantly non-zero interaction predictability is denoted by the hash symbol (p<0.01, Wilcoxon signed-rank test). *p<0.05 CTRL versus SAHS (Mann–Whitney U-test) or CTRL versus CPAP (Wilcoxon signed-rank test).
Figure 4 depicts the results of predictability decomposition applied to the five brain processes, reporting how the predictable dynamics of each process broke down into components related to the self-predictability of the process and to its predictability arising from the other processes of the brain–heart network. In the control group, the full predictability of the brain processes ranged from 81% for the δ wave dynamics to 73% for the θ waves, approximately 65% for the α and σ waves, and 73% for the β waves, resulting higher than that of the cardiac process (approx. 50%). While these full predictability values and their self- and causal constituents did not differ significantly in the untreated and treated patients compared with the controls when considering the faster α, σ and β waves, important modifications were noted for the slower δ and θ waves. Specifically, the predictability was markedly decreased in SAHS patients for the δ dynamics (62%, p=0.005) and the θ dynamics (59%, p=0.003), but was restored at values comparable to those of the healthy controls after long-term CPAP treatment (74% for δ and 71% for θ, p=n.s.). The drop of predictability and its recovery with treatment were entirely owing to the self-predictability of these brain rhythms, whereas the causal predictability was unchanged across the three groups for the θ wave, and was even showing an opposite trend for the δ wave. Figure 4. Distribution of the measures of full predictability (Px|Ω), self-predictability (Px|x) and causal predictability (Pη,Φ∖x→x) assessed for the amplitude of each EEG brain wave (x=δ, θ, α, σ, β) in the network Ω={Φ,η} formed by cardiac process η and the brain process Φ, and expressed as median and quartiles computed in healthy controls (CTRL), in patients with severe sleep apnoea–hypopnoea syndrome (SAHS), and in the same patients studied after prolonged continuous positive airway pressure therapy (CPAP). For each measure, the number of subjects for which it was detected as statistically significant is reported inside the bar. *p<0.05 CTRL versus SAHS (Mann–Whitney U-test) or CTRL versus CPAP (Wilcoxon signed-rank test).
The higher causal predictability of the sleep EEG amplitude in the δ frequency band observed in untreated SAHS patients (8.3% versus 5.7% of healthy controls, p=0.044), as well as its decrease with CPAP treatment (5.6% versus 8.3%, p=0.039) documented in figure 4, can be related to how the brain and heart processes interact in contributing to the predictability of the δ waves. Figure 5 shows indeed that the interaction predictability between the cardiac dynamics and the brain dynamics other than δ is significantly positive in the control group, cannot be distinguished from zero in the SAHS patients, and is again significantly positive during long-term CPAP treatment. Such a redundant interaction between brain and cardiac dynamics for healthy subjects and treated patients was a peculiarity of the δ waves, as it was not observed for any of the other brain wave amplitudes. Figure 5. Box plots of the measures of interaction predictability, Iη,Φ∖x→x, quantifying how the prediction of each brain process (x=δ, θ, α, σ, β) results from the interaction between the remaining brain processes and the cardiac process η, computed in healthy controls (CTRL), in patients with severe sleep apnoea–hypopnoea syndrome (SAHS), and in the same patients studied after prolonged continuous positive airway pressure therapy (CPAP). For each measure, significantly non-zero interaction predictability is denoted by the symbol hash (p<0.01, Wilcoxon signed-rank test).
This study deals with the characterization of the physiological network of brain–heart interactions during sleep, extending the methodology and the results of our previous works in this field [23,24] in several respects. First, complexity and causality of the brain and cardiovascular dynamics during sleep are investigated exploring predictability measures rather than information measures. Complexity is a defining feature of physiological systems, and is commonly quantified as the departure of a signal from a fully predictable time course [39,51]; in this study, we assessed complexity through the full and self-predictability measures, quantifying respectively how much of the unpredictability of an observed signal can be reduced by the knowledge of its own dynamics, or by the knowledge of the dynamics of the whole set of signals composing the observed physiological network [39]. Causality is a concept related to the existence of directed interactions between the observed signals, and is often associated with the physiological mechanisms of joint modulation of interconnected organ systems [52]; here, causality was assessed in agreement with the well-known Wiener–Granger concept of predictability improvement [41]. Furthermore, in this work, the notion of causality is broken down by exploring, for the first time, to the best of our knowledge, in the study of brain–heart networks, the concept of source interaction: defining the novel measure of interaction predictability, we investigate whether two (groups of) source signals act redundantly or in synergy while they contribute to reduce the unpredictability of the target signal. The complete framework, investigating the concepts of complexity and causality from a different perspective than in [23,24] and including the new concept of redundancy/synergy, is applied for the first time on brain–heart networks assessed in pathological states, considering a group of severely apnoeic patients before and after their treatment with the ventilation therapy. When applied to these unique data, our comprehensive approach for the prediction of multivariate time series led to the following original findings: (i) the proposed framework detected consistently the presence of structured dynamics in the nocturnal time course of the cardiac parasympathetic component of HRV and of the EEG wave amplitudes, documented by the statistically significant predictability measures which explained, in all subjects and patients, more than 50% of the variability of heart and brain dynamics during sleep; (ii) severe apnoea–hypopnoea increased the complexity of the nocturnal dynamics of the brain and cardiovascular systems, significantly reducing the predictability of the time courses of the parasympathetic component of HRV and of the δ and θ EEG bandpowers; (iii) this higher complexity of cardiac and slow-wave brain rhythms was due in its largest part to a marked reduction of the self-predictability of these rhythms; (iv) prolonged treatment of SAHS patients with the CPAP therapy restored the full and self-predictability of the δ and θ brain wave amplitudes, but not those of the cardiac parasympathetic component; (v) the magnitude of the causal effects sustaining the network of physiological brain–heart interactions during sleep decreased along the brain-to-heart direction, and increased significantly towards the δ brain process, in apnoeic patients compared with healthy subjects; and (vi) brain–heart communications during normal sleep were characterized by redundant interactions between different brain rhythms contributing to the cardiac dynamics, and between brain and cardiac rhythms contributing to the δ EEG dynamics, which were lost in severe SAHS and were restored by long-term nasal CPAP. The analysis framework introduced in this study was designed to quantify together different aspects of the dynamics of networks composed by multiple interacting dynamical processes, i.e. the reduction in uncertainty about an assigned target process that results from its own dynamics, from the dynamics of the other processes in the network, and from the interaction between these other processes. The relevant quantitative measures are the self-predictability, causal predictability and interaction predictability, and are intimately related to the information-theoretic concepts of information storage, transfer and modification [25]. While traditionally these concepts are quantified through entropy-based functionals such as the storage entropy [43], the transfer entropy [27] and a dynamical formulation of the interaction information [53], in this work, we adopted predictability-based functionals essentially exploiting the prediction error variance. The information-theoretic and predictability frameworks are closely connected as regards the estimation of information storage and information transfer, as the self- and transfer entropies are strictly related to self- and causal predictability, and are formally equivalent in the linear Gaussian approximation [42,46]. An important advantage of the predictability framework regards the evaluation of information modification: the information-theoretic formulation of the interaction information, which for dynamical systems reflects the synergy or redundancy between two sources transferring information to a target process [53], leads to a distorting effect that may produce net synergy in the presence of statistically independent sources [48]. This confusing behaviour, which leads to the prevalence of synergy over redundancy in multivariate Gaussian systems, is avoided if one quantifies information as reduction in variance rather than reduction in entropy. Thus, also according to recent studies [45,48], our formulation based solely on subtracting prediction error variances allows an unbiased evaluation of how information is modified when comparing the combined contribution of two sources with the sum of the individual contributions in the formula for the interaction predictability. In agreement with our previous findings obtained applying the information-theoretic framework in a different group of healthy subjects [23], the results of this work confirm that during undisturbed sleep each node of the brain–heart physiological network displays structured dynamics that can be significantly predicted from the past activity at the node, as well as from the past activity at the other nodes of the network. These structured physiological dynamics are likely the result of changes in the regulation of sleep that determine fluctuations in the autonomic nervous activity. In turn, these fluctuations induce dynamical changes in the activity of both the brain and cardiac systems [3,5,54], which are manifested at multiple time scales [21]. The characteristic scales explored in this and previous works [15,16,23,24,34] are relevant to brain and cardiovascular oscillations occurring with periods longer than 15 min. In a previous study, we documented that this network of brain–heart interactions is sustained mostly by the transitions across the different sleep states, because it is much less connected when assessed for specific individual sleep stages [23]. This study documents that these brain–heart dynamics are also detectable in the sleep recordings of patients with severe SAHS. More importantly, we observe a number of significant alterations in the temporal structure of these dynamics, which were quantified by our measures of full, self-, causal and interaction predictability. First, we found that cardiac dynamics (η rhythm) and slow wave brain dynamics (δ and θ rhythms) are significantly more complex during disturbed apnoeic sleep (figures 2 and 4). SAHS induces sudden surges in sympathetic and vagal cardiac activity [55] as well as in NREM EEG δ activity [33] and, as a result, suppresses the nocturnal rhythmic physiological shifts in the sympathovagal balance. In turn, this suppression likely induces a weakened autonomic control of the heart and brain rhythms and may thus explain their higher complexity, which we measured in this study in terms of markedly reduced values of full and self-predictability. After long-term CPAP treatment, the differences in full and self-predictability observed in SAHS patients compared with healthy controls were maintained for the cardiac dynamics (figure 2) and disappeared for the δ and θ brain dynamics (figure 4). A factor that may contribute to the restoration of the predictability of the brain dynamics is the fact that in this group of SAHS patients the duration of REM sleep, which was significantly lower before the CPAP treatment, increased to values comparable to those of the healthy controls after the treatment [38]. Because the slow wave EEG activity was found to be much more irregularly affected by apnoeas during NREM sleep compared with REM [33], the restoration of REM duration favoured by the treatment may be helpful to restore also the complexity of the dynamical structure of the EEG waves. In contrast, the absence of any difference in all sleep parameters between treated patients and controls, documented in reference [38], was not sufficient to recover the drop in complexity of the dynamics of the HF component of HRV (see figure 2, showing that the drop in complexity documented by lower predictability of η for SAHS was not restored after CPAP). The higher complexity of the cardiac dynamics displayed by SAHS patients both before and after prolonged CPAP treatment documents an important and peculiar alteration of the neural regulation of the heart, because it was observed in the absence of any significant change of the mean and spectral HRV components across the same three groups [38]. A possible factor related to this higher complexity observed for the cardiac time series even after treatment may be the presumed increase of the cardiac vagal tone induced by CPAP, that was documented in previous studies [37,56] and only partially observed in the patients of this work [38]. Nevertheless, future studies are needed to investigate the underlying pathophysiological mechanisms and clinical relevance of this lack of recovery of the predictability of cardiac dynamics. The analysis of the dynamical interactions in the brain–heart physiological network was performed in this study by exploring the intertwined concepts of information transfer and information modification through the measures of causal predictability and interaction predictability. Focusing on the network nodes for which the patterns of predictability changed significantly across groups, i.e. the cardiac η node and the δ brain node, we observed opposite changes of the full causal predictability comparing SAHS patients and healthy subjects: the causal interactions were weaker when directed to the η node (figure 2), and stronger when directed to the δ node (figure 4). This suggests that in apnoeic patients the network of physiological interactions re-organizes its sleep topology in a way such that information flows more towards the δ brain node and less towards the cardiac node. After prolonged CPAP treatment, the information flowing to the δ brain node decreased to values comparable with the healthy controls (figure 4), whereas the flow towards the cardiac node showed only a tendency to increase to the level of the controls (figure 2). These results extend to a broader perspective the findings of a previous study [38], and agree in suggesting that long-term CPAP treatment restores some aspects, but not the full connectivity structure of brain–heart interactions. Contrary to causal predictability, the measure of interaction predictability revealed a common pattern in the comparison among healthy controls and untreated/treated patients, i.e. the loss of redundant interactions in the brain–heart network associated with SAHS and their recovery with CPAP. This was documented by the interactions between each of the θ, α and β waves and the other brain waves when predicting the cardiac activity (figure 3), and between the cardiac dynamics and the brain dynamics when predicting the δ wave activity (figure 5), that were significantly positive for the controls (indicating redundant cooperation), non-significant for the SAHS group (indicating statistical independence), and again significantly positive for the CPAP group. These results suggest that redundancy is a prevailing feature of correctly working brain–heart networks, and its loss during SAs generates a kind of segregation of the network nodes which are still transferring information to each other but perform this task in isolation. Limitations of this work regarding the study design include the small size of the considered group of patients, the fact that patients and controls, though being matched in almost all demographic parameters, differed in the body mass index [38], and the highly variable duration of the CPAP treatment. Combined with the small population size, the different time at which follow-up measurements were taken may have an impact on the interpretation of our results. Nevertheless, because the CPAP therapy was prolonged up to the achievement of beneficial effects in terms of sleep efficiency, duration and quality, its different duration was unavoidable and in some sense purposeful to the formation of a group of treated patients for which therapy was really effective [38]. Given the size and characteristics of our group of patients, our results should be confirmed by larger population studies in which they can be tested also across the duration of long-term treatments. Methodologically, the linear parametric approach adopted in this work guarantees the full detection of the system dynamics only when these dynamics are distributed as a jointly Gaussian process. While we have shown in a recent comparative study that linear and nonlinear estimators detect similar brain–heart networks in normal undisturbed sleep [24], we cannot exclude that nonlinear dynamics which cannot be fully disclosed by our linear approach may significantly contribute to self-, causal and interaction predictability during SAHS and after CPAP treatment. Thus, it may be possible that some of the results presented in this study would change when the same measures are computed relaxing the assumption of linearity. Therefore, we advocate studies comparing the linear approach proposed in this study with model-free nonlinear prediction methods [43,52] to assess the actual contribution of nonlinear dynamics to full night brain–heart interactions during apnoeic sleep and after long-term treatment. As to the theoretical development, the proposed measure of interaction predictability may be compared with novel concepts of information modification based on the simultaneous evaluation of redundant and synergistic contributions to a given target [48]. Finally, the approach proposed here could be extended in future studies to the identification of larger physiological networks comprising different physiological systems. This is particularly important given the known relevance to physiological interactions of variables not considered in the present study such as respiration, arterial pressure or cerebral blood flow [57,58]; in particular, the inclusion of respiratory dynamics would allow to relate the alterations of the brain–heart dynamics observed in this study to the cardiopulmonary instability which breaks down the scaling function of HRV in SA patients [59]. The network size can be enlarged even more within an individual complex system such as the brain, e.g. extending the intrachannel EEG network to the inclusion of the γ frequency band [19,20] and including a new subnetwork which explores the spatial level through the simultaneous interchannel analysis of the EEG at different scalp locations [20]. In conclusion, we have documented specific alterations of the sleep dynamics of the cardiac parasympathetic and brain wave activities in patients with SAHS, which are evident even in the absence of substantial modifications in the sleep characteristics or in the cardiovascular parameters. These alterations, documented for the first time using our predictability framework, are manifested through the falloff of the regularity of cardiac activity and slow wave brain activity, and of the redundant information shared within the network of brain–heart physiological interactions. We documented also that prolonged treatment of apnoeic patients with nasal CPAP is effective in restoring almost completely the sleep dynamics of the slow EEG waves as well as the links of the brain–heart network, but is unable to recover the structure of the cardiac dynamics. These findings help in elucidating how the complex influence of the autonomic nervous system on the sleep physiological dynamics is modified by sleep disorders, and may contribute to assess the efficacy of therapeutic approaches as well as to adjust the delivery of such therapies. All participants gave written informed consent to the procedure that was approved by the Ethical Committee of the Erasme Academic Hospital of Brussels. The datasets supporting this article have been uploaded as part of the electronic supplementary material, in the file Faes_etal-PTRSA2015-data.zip. L.F.: conception and design of the work, analysis of data, interpretation of data, drafting of the manuscript, final approval of the version to be published; F.J.: acquisition of data, critical revision of the work, final approval of the version to be published; D.M., S.S., A.P., G.N.: interpretation of data, critical revision of the work, final approval of the version to be published. We have no competing interests. This work was supported by the Healthcare Research Implementation Programme (IRCS), Provincia Autonoma di Trento and Bruno Kessler Foundation, Italy. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 5As an integrated physiologic system under neural regulation, the cardiac system exhibits complex behaviour characterized by continuous fluctuations and transient, nonlinear and scale-invariant temporal dynamics. Reflecting modulation in neural autonomic control and sympatho-vagal balance [1–4], the linear and nonlinear characteristics of cardiac dynamics [5–7] change with sleep–wake cycle [8,9], across circadian phases [10] and sleep-stage transitions [11–13]. Various measures derived from linear and nonlinear characteristics of cardiac dynamics have been established as robust biomarkers for diagnosis and prognosis under a broad range of conditions [14,15], young versus elderly [16,17] and under pathological perturbations [18–23]. To account for these extensive empirical observations, modelling approaches have been developed (i) to investigate the underlying mechanisms of neuroautonomic control and associated nonlinear feedback loops acting on a wide range of time scales, and (ii) to study how these mechanisms change across different physiologic states and conditions [24]. Despite the importance of understanding the basic mechanisms of neural regulation of organ systems, it is not well understood how the brain and the cardiac system dynamically interact and coordinate their functions to generate a variety of physiologic states. Specifically, the role of different brain rhythms and their temporal dynamics in mediating brain–heart communications remains an open question. Probing dynamics of brain–heart interactions is a major challenge due to several levels of complexity. At the individual systems level: the heart and the brain are very different integrated systems, each with its own structural and functional complexity, leading to output dynamics with distinct characteristics, i.e. the cardiac system exhibits a pronounced oscillatory pattern on the scale of seconds, whereas brain dynamics are characterized by multiple rhythms with different origins and functions that operate on much shorter time scales. At the level of pairwise interactions: physiological systems often interact through multiple forms of coupling, which are of transient nature, can switch on/off, and can simultaneously coexist [25–29]. In the context of brain–heart communications [30], synchronization [31–33], coherence [34], time delay [35,36] and information transfer [37,38] play important roles. At the organism level where dynamical networks of diverse organ systems are essential: integrated physiologic function emerges as a global phenomenon from hierarchical networks representing the dynamical interactions among organ systems, and cannot be simply described by summing up the behaviours of individual systems. Within the framework of ‘Network Physiology’ [36,39,40], using the concept of time-delay stability (TDS), recent work [26,40] has demonstrated that the cardiac system communicates with the brain not only through one but rather through multiple brain rhythms simultaneously. Further, empirical analyses have shown that, during different physiologic states, brain–heart communications are predominantly mediated through different brain rhythms, where patterns of brain–heart networked interactions depend on brain locations, and undergo complex hierarchical reorganization with transitions across physiologic states [40]. To quantify brain–heart interactions, here we extend the TDS approach [36,40–42] and we propose a generalized time-delay analysis based on the novel concept of delay-correlation landscape (DCL) to investigate coordination of bursting activities in the brain and heart output signals. We hypothesize that key properties of the brain–heart DCL reflect changes in neuroautonomic control of cardiac and brain dynamics associated with distinct physiologic states such as sleep or wake and different sleep stages. Specifically, we hypothesize that the characteristic time delays and directionality of brain–heart communications between each physiologically relevant brain rhythm and cardiac output dynamics exhibit unique signature profiles reflecting physiologic function during distinct physiologic states. We analyse continuously recorded multichannel physiological data obtained from 34 healthy young subjects (17 female, 17 male, with ages between 20 and 40, average 29 years) during night-time sleep [43] (average record duration is 7.8 h). This allows us to track the dynamics and evolution of brain–heart interactions during different sleep stages and sleep-stage transitions. We focus on physiological dynamics during sleep as sleep stages are well-defined physiological states, and external influences due to physical activity or sensory inputs are reduced during sleep. Sleep stages are scored in 30 s epochs by sleep laboratory technicians based on standard criteria [44,43]. In particular, we focus on the electroencephalogram (EEG) and the electrocardiogram (ECG). To compare these very different signals with each other and to study interrelations between them, we extract the following time series from the raw signals: the spectral power of five physiologically relevant frequency bands of the EEG, derived from the central C3 channel, in moving windows of 2 s with a 1 s overlap (namely δ wave (0.5–3.5 Hz), θ wave (4–7.5 Hz), α wave (8–11.5 Hz), σ wave (12–15.5 Hz) and β wave (16–19.5 Hz)); heartbeat RR intervals are re-sampled to 1 Hz (1 s bins) after which values are inverted to obtain the instantaneous heart rate (HR). Thus, all time series have the same time resolution of 1 s before our analyses are applied. We calculate the fast Fourier transform (FFT) in 2 s EEG windows and determine the spectral power in the EEG frequency bands mentioned above. As there is a problem of power leakage from one frequency bin to others, we taper the window by a Hann function, and because tapering itself introduces the problem of weighting the edge of the windows much less than the data in the middle, we choose an overlap of half the window length, i.e. 1 s. According to Press et al. [45], tapering and choosing an overlap that is half the window length resolves the problems of power leakage and different weights, respectively. Because we are analysing EEG data that were recorded during sleep, we use the five EEG band definitions that are commonly accepted in sleep medicine [46] as defined above. We originally extended the definition for β to include ‘high β waves’ (20–30 Hz); however, we noted that, past 20 Hz, the EEG is more susceptible to electromyography (EMG) movement artefacts, and therefore we chose the traditional 16–19.5 Hz frequency band. The ECG data are analysed and annotated by a semi-automatic R-peak detector (see below). EEG recordings were filtered by a high-pass filter (0–0.4 Hz) and a low-pass filter (30–70 Hz). We apply the high-pass filter in this range to filter out slow movement artefacts without much affecting δ frequencies. The low-pass filter filters out high-frequency artefacts (e.g. from EMG). In addition, the EEG recording device had a 50 Hz notch filter. R-peaks are extracted from the ECG data using the semi-automatic peak detector Raschlab developed by the cardiology group of Klinikum Rechts der Isar, Munich, Germany (R. Schneider. Open source toolbox for handling cardiologic data, available on the internet: www.librasch.org). A beat classification (normal beat, ventricular beat, artefact) is assigned to each R-peak by the detector. Then we calculate the series of RR time intervals between each pair of consecutive heartbeats and obtain the HR time series by inverting the RR series. Ectopic beats and artefacts are detected by Raschlab. Additionally, we examine more carefully the obtained RR intervals and exclude RR intervals from our calculations, if (i) the beat at the beginning or at the end of the interval is not normal, (ii) the calculated interval is shorter than 330 ms or longer than 2000 ms or (iii) the interval is more than 30% shorter or more than 60% longer than the preceding interval. The purpose of the last filter is to eliminate extrasystoles and ectopic beats unnoticed by the peak detector. This procedure led to ≈1% removal of original ECG RR intervals and corresponding ≈1% reduction in the original EEG data. One potential approach to study brain–heart interaction mediated by different brain rhythms is to use the absolute spectral power in each EEG frequency band. However, our preliminary results (not shown) indicate that the bursting activity in HR is strongly modulated by trends in the total EEG power within the frequency range of 0.5–19.5 Hz (sum of all five frequency bands)—a masking effect leading to very similar results for each pair of HR and brain rhythm interaction. In order to eliminate this masking effect and to isolate spurious synchronization among all brain rhythms caused by modulations in the total EEG spectral power, we use normalized (relative) spectral power in our analyses. First, we calculate the time series of the total power of EEG (SEEG) as the sum of all spectral powers from the five frequency bands listed above. Next, we obtain the relative spectral power in each frequency band with a 1 s resolution. The relative spectral power (Sδ(t),Sθ(t),Sα(t),Sσ(t),Sβ(t)) is obtained as the ratio between the spectral power in the specific frequency band and the total spectral power of all five bands. Thus, the obtained normalized relative spectral power represents the relative contribution of each brain rhythm to the total brain activity, and allows one to investigate the individual role of each brain rhythm in facilitating brain–heart communications. The TDS method proposed in earlier studies [36,41,40] focuses on the time evolution and stability of the time delay defined as the time shift corresponding to maximum degree of correlation/ anti-correlation between two signals. While the percentage of data segments exhibiting time-delay stability (%TDS) was found to have important physiological relevance, as it undergoes a pronounced transition from one physiologic state to another, it does not provide information on the type of correlation (positive or negative) and directionality of interaction based on the sign of the time delay. To address these limitations and to further quantify the dynamical aspects of brain–heart interaction, we extend our TDS methodology to a more generalized time-delay analysis framework, which keeps track of both the time t evolution of cross-correlation C as well as the time shift τ dependence of the cross-correlation, i.e. C(t,τ). As shown in figure 1, for each time window t with size L=30 s, we obtain the Spearman cross-correlation as a function Cxy of the time shift τ∈[−30,30] between two signals x and y, where rx and ry represent the ranks of the values in the signals x and y, respectively. The functional form of Cxy(t,τ) can be written as: 2.1 2.2 whereFigure 1. Generalized time-delay analysis and delay-correlation landscape. (a) Normalized spectral power Sδ of δ brain rhythm (0.5–4 Hz) derived from EEG recording at the central C3 channel (C3–M2 set-up), and (b) heart rate (HR) from a healthy subject during 30 min of night-time sleep. Black dashed lines and background colours in (a) and (b) represent sleep stages (denoted on the right vertical axis) as defined by traditional sleep-stage scoring criteria [43,44]. (c) Spearman cross-correlation function Cδ,HR(τ,t) between Sδ and heart rate (HR) are obtained in 30 s windows moving with a step of 30 s and plotted as delay-correlation landscape (colour map). Horizontal axis t indicates the time corresponding to the centre of the two aligned 30 s windows used for the cross-correlation calculation. At each time step t, we shift the 30 s window of the HR relatively to the 30 s window of EEG spectral power, and calculate the cross-correlation as a function of the relative time shift τ (vertical axis): τ>0 when the 30 s window for brain rhythm signal precedes the window of cardiac signal and vice versa for τ<0. Colour of the delay-correlation landscape (DCL) represents the value of cross-correlation: red corresponds to positive correlation C>0 and blue corresponds to negative correlation C<0. Black triangle symbols in (c) mark the time evolution (in 30 s steps) of the maximum correlation time delay τMC(t) defined as the time shift corresponding to the maximum absolute value of the cross-correlation function. Positive and negative correlation maps are plotted separately in (d) and (e), where only positive or negative correlation values are shown. We define positive correlation time delay τPC(t) (red circles) and negative correlation time delayτNC(t) (blue squares) as the time shift corresponding to the maximum positive or maximum negative correlation. (Online version in colour.)
Theoretically, the DCL contains the information provided by both the traditional cross-correlation analysis and the original TDS method [36,41,40]. The cross-section of DCL along the black dashed line τ=0 in figure 1c represents the cross-correlation of the two signals without any time shift, i.e. Cxy(t)|τ=0. Black triangle symbols in figure 1c mark the evolution of time delay defined in the original TDS method as the time shift between the two signals at which the maximum in the absolute value of cross-correlation is observed, i.e. To quantify the structure of DCL and to better understand the nature of interactions that generate the delay-correlation configuration at each time t, it is important to differentiate the two types of cross-correlation—positive versus negative. Thus, we construct the subset of DCL with only positive correlation, i.e. only ‘hills’ in the DCL, as shown in figure 1d, where red solid circles mark the positive correlation time delay, In our analyses of brain–heart interactions, we fix the second signal to be the instantaneous heart rate y(t)≡HR(t), and we assign the first signal x(t) as the relative spectral power of five physiologically relevant brain rhythms (Sδ(t),Sθ(t),Sα(t),Sσ(t),Sβ(t)) and the total spectral power SEEG(t). Under this definition, a positive time delay τ>0 always corresponds to a situation when the modulation in brain rhythms precedes corresponding changes in the cardiac signal, and vice versa for τ<0. Our previous work has shown that TDS is a reliable measure of interaction and coupling between dynamical systems, and that it is sensitive to differentiate between physiological states and conditions even in cases when the amplitude of cross-correlation cannot provide a statistically significant separation between real and surrogate data [36]. Thus, to probe brain–heart interactions, here we quantify the temporal dynamics and statistical properties of the time delay inherent to different pairs of brain rhythms and heart interactions by calculating the probability distribution P(τ) of three types of time delay, including maximum correlation time delay (τMC), positive correlation time delay (τPC) and negative correlation time delay (τNC). In other words, we are interested in the most probable time delay when significant cross-correlations are observed. The probability distribution is represented by a renormalized histogram of time delays and the summation of the histogram values over all bins (1 s bin) equals 1. As shown in figures 2 and 3, we obtain the probability distributions for all three types of time delay for a typical individual subject as well as for the entire group of subjects. Panels along the horizontal direction are colour-coded to represent interactions between different brain signals and the HR output, whereas different rows of panels correspond to different types of time delay. Figure 2. Probability distribution profiles for maximum, positive and negative correlation time delays τMC, τPC and τNC: (a) for an individual subject and (b) for the group average of 34 healthy subjects representing the entire night-time sleep period. Solid black line in each panel is a moving average of the probability distribution (smoothed profile with 3 s moving window). The bin size for the time delay τ is 1 s. Similarity between the individual subject profiles and the group average profiles for each pair of brain rhythm–heart interaction indicates a universal mechanism underlying time delays in brain–heart communication. (Online version in colour.)
Figure 3. Statistical patterns of characteristic time delays underlying brain–heart communications. Probability distributions of the maximum correlation time delay τMC (first row), positive correlation time delay τPC (second row) and negative correlation time delay τNC (third row) are obtained from the generalized time-delay analysis (figure 1) for each pair of brain–heart interactions. (a) Individual distributions for all 34 subjects (plotted in different colours) and (b) the group average, where black error bars represent the standard deviation across subjects. Distributions are obtained for the entire night-time sleep period. Distribution profiles skewed to the right with peak at τ>0 indicate that activations in brain dynamics precede cardiac dynamics. Considering interactions between the relative spectral power of each brain rhythm and the HR, we find unique profiles for the probability distribution of the time delay τ, indicating a specific role of each brain rhythm in mediating brain–heart interactions. Note the different distribution profiles of τMC, τPC and τNC for the interaction between each brain rhythm and the heart, indicating that positive and negative cross-correlations are characterized by different time delays. Interactions between the total EEG power (all five brain rhythms) and the HR are characterized by a significant peak at τPC=0, indicating synchronized bursting activity in the brain–heart network where modulations in the same direction for the heart and the brain occur simultaneously. By contrast, total EEG power and HR interactions exhibit a sharp peak at τNC>0, indicating that modulations in the HR that are in opposite direction to changes in brain oscillation occur with a positive time delay. The double-peak distribution profile of τMC reflects a combination of the profiles for τPC and τNC. Remarkably, the ensemble of probability distribution profiles for all three types of time delays is consistently observed for all subjects with small standard deviation, indicating a universal mechanism underlying time delays in brain–heart communication. (Online version in colour.)
For a given pair of brain–heart interaction, the probability distributions of τPC and τNC can be further combined into one histogram profile by plotting P(τPC) in the upper half plane with corresponding colours as in figures 2 and 3, and inverting and plotting P(τNC) in the lower half plane with the same light shade (figure 4). Combining histograms in this way enables us to better demonstrate the changes in the probability distribution profile of the time delay for a given pair of brain–heart interaction with transitions from one physiologic state (sleep stage) to another, as well as the corresponding hierarchical reorganization of the entire set of distribution profiles for all pairs of brain rhythms and heart interactions (shown in figure 4). Figure 4. Change in brain–heart time-delay distribution profiles with transitions across physiologic states. Joint profiles of the probability distribution of τPC and τNC for each pair of brain–heart interactions during different physiologic states (sleep stages) obtained by pooling data from all subjects (see §2). Distributions for τPC are plotted in the upper half plane in each panel (different colours for different brain–heart interactions as in figure 3). Distributions for τNC are inverted and plotted in the lower half plane with the same light shade. Each sleep stage (horizontal row) is characterized by a specific set of profiles representing the time-delay characteristics for each brain–heart interaction. Considering each pair of brain rhythm and heart interaction (column), the joint distribution profile of τPC and τNC changes from one sleep stage to another, leading to a complex reorganization for the entire set of profiles across different physiologic states. This reorganization in the time-delay profiles across sleep stages is also observed for the total EEG spectral power (right column), indicating a pronounced change in the coupling between the overall brain activity and cardiac dynamics. Note that, for all pairs of brain–heart interactions during all sleep stages, there are no peaks with significant negative time delay, indicating that brain–heart communications are mainly mediated through directional interaction from the brain to the heart. Each sleep stage is characterized by a specific ensemble of joint time-delay distribution profiles indicating that these profiles are a robust signature of physiologic state. (Online version in colour.)
Peaks in the probability distribution profiles of time delay correspond to the characteristic time delays that underlie brain–heart communications, and the sign of these characteristic time delays is indicative of the directionality of brain–heart communication: τ>0 for directional interaction from the brain to the heart, whereas τ<0 indicates that cardiac dynamics precede modulations in brain activity. Integrated physiological systems, in general, are coupled by feedback and/or feed-forward loops with a broad range of time delays that underlie physiologic interactions. Combination of these feedback loops leads to different types of coordinated modulation in the output dynamics of physiological systems that can simultaneously coexist [25–27]. Characteristics of physiologic coupling and interaction, such as the range of time delays and different modes of coordination, are essential for the entire organism to optimize its function during different physiologic states and to generate proper response to external perturbation. Consequently, we focus on the characteristic time delays involved in different modes of brain–heart interaction. We perform empirical analyses of EEG and HR data recorded in healthy subjects during night-time sleep to probe the interaction between distinct brain rhythm and the heart, and how these interactions change with different sleep stages (well-defined physiologic states). As brain rhythm activation and cardiac dynamics continuously change even within the same physiologic state, we expect a high degree of complexity in the DCL representing brain–heart communications (§2). As shown in figure 1c, the DCL for the Sδ−HR interaction is characterized by pronounced heterogeneity where different types of cross-correlation (positive or negative correlation as represented by regions with different colours) form a complex mixture. To investigate whether there are characteristic time delays associated with brain–heart interactions, we construct the DCL (figure 1) for each pair of brain rhythm and HR signals. We obtain probability distributions for three distinct types of time delays:
We find that, for each pair of brain–heart interaction, τMC, τPC and τNC are characterized by markedly different patterns in their probability distribution P(τ) (vertical columns in figures 2 and 3). For example, Sδ−HR interaction has a pronounced peak for P(τPC) at τPC≈0 s, while the peak for P(τNC) is located at τNC≈6 s, indicating that different modes of brain–heart coordination operate at markedly different time delays. For τMC, we consistently observe a double-peak pattern in its probability distribution for all pairs of brain rhythm and HR interactions, which reflects features of the distribution profiles for both τPC and τNC—this double-peak pattern is most pronounced for the SEEG–HR interaction (figures 2 and 3). Comparing pairs of interaction between the HR and different brain rhythms, we find that for each type of time delay (horizontal rows in figures 2 and 3) different pairs of brain–heart interactions exhibit distinct profiles of the probability distribution P(τ). For example, while P(τPC) for the pair Sδ−HR has a dominant peak at τPC≈0 s, the pair Sθ−HR is characterized by a dominant time delay of τPC≈6 s, and Sσ−HR has characteristic time delays at both τPC≈0 s and τPC≈6 s. By contrast, comparing P(τNC) for different pairs of brain rhythm–HR interaction, we observe a pronounced peak at τNC≈6 s for Sδ−HR and peak at τNC≈0 s for Sθ−HR; this is exactly opposite to the peak locations observed in P(τPC) (figures 2 and 3). These distinct profiles indicate that each brain rhythm plays a specific role in mediating brain–heart interactions. Remarkably, the entire ensemble of distribution profiles for all three types of time delays is robust, as it is consistently observed for all individual subjects (figure 3a) as well as for the group average behaviour (figure 3b). This consistency in time-delay distribution profiles is demonstrated by the small standard deviation across subjects (error bars around group average value, figure 3). Our statistical analysis (electronic supplementary material, figure S5a) reveals that the most pronounced characteristic time delays underlying brain–heart communications are associated with the following pairwise interactions: (i) Sδ–HR, (ii) Sθ–HR and (iii) SEEG–HR. These interactions exhibit statistically significant peaks in the profiles of P(τPC) and P(τNC), coupled with consistent time delays for τPC and τNC across subjects: (i) for Sδ–HR, τPC=0 s and τNC=6 s; (ii) for Sθ–HR, τPC=6 s and τNC=0 s; and (iii) for SEEG–HR, τPC=0 s and τNC=6 s. To further explore the interrelation between the characteristic time delays in the brain–heart communication and distinct physiologic functions, we calculate the probability distributions for τPC and τNC for different sleep stages, including deep sleep (DS), light sleep (LS), rapid eye movement sleep (REM) and wake/brief arousals (W). We find that each sleep stage is characterized by a specific set of joint profiles of the probability distribution for τPC and τNC (horizontal rows in figure 4). Following each pair of brain–heart interaction (vertical columns in figure 4), we observe that the joint profile of τPC and τNC changes significantly from one sleep stage to another, reflecting changes in the neural regulation of cardiac dynamics. Moreover, we find that the joint τPC and τNC profile for each pair of brain–heart interaction follows a specific transition pattern across sleep stages (figure 4). Thus, with transition across physiological states there is a complex reorganization of the entire ensemble of time-delay distribution profiles of the different brain rhythm–HR interactions. Intriguingly, our results for the total power SEEG–HR interaction show that, with increase in sympathetic tone from DS to LS, REM and W, the peak in P(τNC) at τNC=6 s completely vanishes (figure 4), corresponding to the loss of significant time delays for negative brain–heart cross-correlation (i.e. for modulations in the opposite directions between the EEG and HR signals). Indeed, increased bursting activity in brain dynamics associated with dominant parasympathetic tone during DS and LS is associated with dipping in the HR, leading to a pronounced anti-correlation profile, which disappears under dominant sympathetic tone during wake. This pronounced τNC=6 s time delay is consistently observed for all subjects (electronic supplementary material, figure S5). While it may be associated with the baroreflex feedback loop, the underlying physiologic mechanism for this characteristic time delay remains to be explored. These observations identify characteristic time delays of brain–heart communications as a new hallmark of physiologic state and function. Notably, we find that, for both τPC and τNC, there are no peaks at significant negative time delays as shown by the profiles in figure 4, indicating that brain–heart communications are mainly mediated through directional interaction from the brain to the cardiac system. To understand the basic mechanisms of neuroautonomic control of the cardiac systems, we develop a generalized time-delay analysis framework and a novel DCL approach to investigate the role of distinct physiologically relevant brain rhythms in mediating brain–heart interactions. Compared with the traditional cross-correlation analysis with a fixed time delay or the original TDS approach, where the emergence of stable time delay marks the onset of dynamical coupling between two systems, the approach proposed here keeps track of both the time evolution and the delay dependence of the cross-correlation between two signals. We find that brain–heart interactions exhibit characteristic time delays and that different modes of interaction (i.e. positive or negative cross-correlations) are characterized by different time delays. Our results demonstrate that the interactions between different brain rhythms and the HR are characterized by distinct distribution profiles for time delays, indicating that each brain rhythm has a specific role in mediating brain–heart communications. Furthermore, we find that the time-delay profile for each pair of brain rhythm and HR interaction follows a unique transition pattern from one sleep stage to another, leading to a complex reorganization of the entire ensemble of time-delay profiles. As sleep-stage transitions are closely associated with changes in sympatho-vagal balance, the uncovered ensemble of time-delay profiles representing brain–heart interactions reveals previously unknown dynamical aspects of cardiac neural regulation that are a hallmark of physiologic state and function. Remarkably, the uncovered time-delay distribution profiles for all pairs of brain rhythm and HR interactions are consistently observed in all healthy subjects, and exhibit a similar reorganization with transition across sleep stages in each subject. Thus, these new measures can potentially be used not only as robust markers of physiologic states and functions under healthy condition, but also as diagnostic and prognostic indicators of pathological perturbations. The main purpose of this work is to present a first proof-of-concept demonstration of a DCL approach to identify and quantify the characteristic time delay underlying brain–heart interactions. Thus, in this study, we use data from the C3 EEG channel only, which is most commonly used in sleep research, and we use EEGC3–HR interaction as an example to present our computational framework. Naturally, follow-up work will extend to other EEG leads to identify the different roles of brain location in mediating brain–heart interactions. These extended analyses may include not only instantaneous HR time series, as presented here, but also high- and low-frequency HR components as well as other static and dynamic local characteristics of the cardiac output. Further, modelling approaches based on surrogate time series with different autocorrelations and other dynamical characteristics as observed in the brain and heart output signals can help elucidate the origin and structure of the DCL representing brain–heart interactions during different sleep stages. The proposed time-delay approach is general and can be applied to other types of dynamical systems with complex output signals where the existence and the nature of coupling and interactions are not known a priori. Moreover, the novel concept of delay-correlation landscape encompasses all key elements of cross-correlation, and provides a comprehensive picture of the coupling strength, characteristic time delays and time evolution of correlation between dynamical systems. Thus, the approach presented here can serve as a general analytical tool to understand basic mechanisms underlying physiological interactions, which is essential for the development of the new field of ‘Network Physiology’. The research protocol was approved by the Institutional Review Boards of Boston University (Boston, MA, USA) and was conducted according to the principles expressed in the Declaration of Helsinki. The data we used in this work are pre-existing multi-channel physiologic recordings from EU SIESTA databases. The detailed protocol of the SIESTA database can be found in Klösch et al. [43]. All participants provided written informed consent. A.L. and K.K.L.L. contributed equally to this paper. A.L. and K.K.L.L. designed the analysis algorithm. A.L., K.K.L.L., R.P.B. and P.Ch.I. analysed the data. K.K.L.L., R.P.B. and P.Ch.I. prepared the manuscript. P.Ch.I. initiated the investigation and supervised all aspects of the work. All authors discussed the results and commented on the manuscript. The authors declare that they have no competing interests. This research was supported by W. M. Keck Foundation, National Institutes of Health (NIH grant 1R01- HL098437), the Office of Naval Research (ONR grant 000141010078), the US–Israel Binational Science Foundation (BSF grant 2012219), EC-FP7 Marie Curie Fellowship (IIF 628159) (to R.P.B.), the National Natural Science Foundation of China (grant no. 61304145) (to A.L.) and the Research Fund for the Doctoral Program of Higher Education (grant no. 20130009120016) (to A.L.). We acknowledge the generous support given by the above organizations. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 6General anaesthesia plays a crucial role in many surgical procedures, and it therefore has an enormous impact on human health. It is a drug-induced, reversible state characterized by unconsciousness, anti-nociception or analgesia, immobility and amnesia [1,2]. On rare occasions, however, the patient can remain unconscious longer than intended, or may regain awareness during surgery. There are no precise measures for maintaining the correct dose of anaesthetic, and there is currently no fully reliable instrument to monitor depth of anaesthesia. Although a number of devices for monitoring brain function or sympathetic output are commercially available [3], the anaesthetist also relies on clinical assessment and experience to judge anaesthetic depth. The undesirable consequences of overdose or unintended awareness might in principle be ameliorated by improved control if we could understand better the changes in function that occur during general anaesthesia, in particular the dynamical brain states, the dynamics of cardiovascular oscillations and their mutual interactions [4]. General anaesthesia can be induced by different anaesthetics which can affect different physiological regions, receptors and channels [5,6]. In this study, we used two of the most widely used anaesthetics—propofol and sevoflurane, i.e. we used one of the two in each anaesthesia measurement. Propofol is introduced intravenously, while sevoflurane is a sweet-smelling, non-flammable type of ether that is inhaled [7–10]. The central enigma in general anaesthesia is the nature of the unconscious state mediated in the brain. Neuronal states often manifest themselves as changes in brain electrophysiological activity, which emanates from the dynamics of large-scale cell ensembles oscillating synchronously [11,12] within characteristic frequency intervals. Individual ensembles communicate to integrate their local information flows into a common brain network. One way to describe such an integration or communication is through cross-frequency coupling, a method that has led to numerous studies elucidating the respective roles of cognition, attention, memory and anaesthesia [9,10,13–15]. Jirsa & Müller [13] recently identified different types of cross-frequency coupling based on use of the power, phase or frequency domains; in what follows, we focus on phase–phase cross-frequency couplings. Unlike earlier cross-frequency coupling methods, the approach that we will discuss assesses neuronal states through the computation of coupling functions describing the functional forms of individual cross-frequency interactions. Coupling functions prescribe the physical rule specifying how the inter-oscillator interactions occur. They determine the possibility of qualitative transitions between the oscillations, e.g. routes into and out of phase synchronization [16]. Their decomposition can describe the functional contribution from each separate subsystem within a single coupling relationship. In this way, coupling functions offer a unique means of describing mechanisms in a unified and mathematically precise way. It is a fast growing field of research, with much recent progress on the theory [17,18] and especially towards being able to extract and reconstruct the coupling functions between interacting oscillations from data, leading to useful applications in cardiorespiratory interactions [19–21], chemistry [16], mechanics [22] and communications [23]. We will show that, in neuronal analysis, the cross-frequency coupling function describes much more than just a new way of measuring effects: it opens up a whole new perspective on the functional mechanisms underlying the functionality of the brain network. The oscillatory processes of the brain are not only individually important to the function of the central nervous system, but they can also interact, both mutually and with other physiological oscillations. The latter comprise, e.g., the oscillatory processes of the cardiovascular system [24] including the heart and the lungs, which are closely associated because, working together, they provide the blood supply with oxygen and nutrients for the whole body including the brain. The brain's functional state is obviously of crucial importance in general anaesthesia and as such it provides the basis for number of measures [3] (including, e.g., the BIS (bispectral index) monitor by Medtronic (formerly aspect medical), the Entropy monitor by GE Healthcare, the Narcotrend index by MonitorTechnik, and others). However, although traditional anaesthetic monitoring includes only the on–off awake versus unconscious classification, indirect or surrogate measures of brain function, such as movement, blood pressure, heart rate, sweating and other anaesthesia-induced changes to the cardiovascular system [25–29], also provide valuable indicators [1]. Moreover, the two systems are connected in many ways, and some signatures of causal interaction have already been demonstrated [15]. For a comprehensive assessment of general anaesthesia one should therefore add a consideration of the (complex) interactions between the cardiovascular and brain oscillations [4], and the integration of their functions into what are interconnected physiological networks [4,15,30]. One may thus investigate the mechanisms and connections between the brain and the loss of consciousness [1,2] on the one hand, and, on the other hand, the cardiovascular system which is closely related to the function of the autonomous nervous system including anti-nociception, analgesia and the perception of pain [31–34]. In this paper, we seek to establish the functional laws defining the mutual interactions between the brain, heart and the lungs (figure 1) in general anaesthesia. The study is based on three complementary pillars: (i) anaesthesia with two of the most widely used anaesthetics, using the same experimental set-up; (ii) application of the novel methodology of cross-frequency coupling functions to determine phase-causal links and to probe the interaction mechanisms directly, and (iii) assessment of general anaesthesia based on the combined dynamics and interactions of the brain, lungs and heart oscillations. Figure 1. Schematic of the main aims of the study. We seek to investigate the interactions between oscillatory processes in the brain, lungs and heart, and to establish how they are affected by general anaesthesia. The interactions are assessed by reconstruction of the coupling functions. The analyses are performed on non-invasive measurements of the electroencephalogram (EEG), the respiration signal from expansion of the thorax and the electrocardiogram (ECG). Samples of raw measurements are shown adjacent to each of the organs, as are also the relevant cross-frequency intervals. (Online version in colour.)
Cross-frequency couplings are usually inferred by methods based on the statistics of the coupled signals, such as the correlation and (bi-) coherence measures. Such approaches tell one about the functional connectivity [35], but they do not provide information about causality or about the form of the coupling functions. By contrast, however, we now show that inference of cross-frequency couplings based on a model of coupled phase oscillators [36,37] and dynamical Bayesian inference [19,38,39] enables us to infer the effective connectivity [35], i.e. to estimate the coupling functions and the underlying causality. We note that the effective connectivity was initially discussed in relation to the spatial segregation of brain functions [40] and is often used in this sense by the neuroscience community. In this work, we do not study spatial connectivity but, rather, we exploit the mathematical concept of effective connectivity in determining the influence that one oscillator exerts on another, under a particular model of causal dynamics [35,41]. With its ability to infer time-evolving coupled dynamics in the presence of noise, dynamical Bayesian inference is ideal for the calculation of effective connectivity from neuronal oscillations. The signals derived from the chosen cross-frequency intervals are oscillatory, and their interactions can be studied effectively through their phase dynamics. We therefore consider a model of two coupled phase oscillators [36] described by the stochastic differential equation 2.1 with i≠j for i,j={1,2} and where ωi(t) is the parameter for the natural frequency. The deterministic part given by the base functions qi(ϕi,ϕj,t) describes the self and the interacting dynamics. The external stochastic dynamics ξi(t) is considered to be Gaussian white noise 〈ξi(t)ξj(τ)〉=δ(t−τ)Dij. Owing to the periodic nature of the deterministic dynamics, the base functions can be decomposed into infinite Fourier seriesDynamical Bayesian inference [19,39] enables us to evaluate the model parameters 2.2 where summation over n=1,…,N is assumed, and summation over repeated indices k and w is implicit. We used informative priors and a special procedure for the propagation of information between consecutive data windows [19], which permitted the inference parameters that varied with time (for implementation and usage, see [42,43]). Given its ability to infer time-varying and noisy dynamics, our Bayesian method is especially well fitted for applications to EEG, ECG and respiration signals. A block diagram summarising the analysis procedure is provided in the electronic supplementary material.Once we have the inferred parameters The form of a coupling function depends on the differing contributions from individual oscillations. Changes in form may depend predominantly on only one of the phases (along one-axis), or they may depend on both phases, often resulting in a complicated and intuitively unclear dependance. This demonstrates the need for a model able to distinguish the individual functional contributions to a coupling. Accordingly, following the cardiorespiratory model [21], we present a generalized coupling decomposition model (figure 2). Figure 2. Model of the coupling decomposition. The dynamical equation (top) represents how one phase oscillator (index 1) is influenced by another (index 2). The net coupling q1(ϕ1,ϕ2) is decomposed into two functional entities: the direct d1(ϕ2) and the indirect ii(ϕ1,ϕ2), coupling functions. The dynamics is also characterized by a natural frequency parameter ω1 and external noise perturbations ξ. (Online version in colour.)
Previous coupling treatments, including the cross-frequency coupling in neuroscience, have focused on the net coupling in one direction. Instead, we decompose the net coupling into two components depending on their functional roles: the direct and the indirect couplings (figure 2). Direct-coupling describes the influence of the direct (unidirectional) driving that one oscillator exerts on the other. Arguably, it is the most studied interaction in physiology, often linked to modulation mechanisms. We will see that direct-coupling is the dominant mechanism in most of the coupling functions. The second component, indirect-coupling, often called common-coupling, depends on the shared contributions of the two oscillators. The indirect coupling includes also the diffusive coupling given with the phase difference terms. The mechanism behind this coupling component (the small circle on the arrow in figure 2) can lie in some functional dependence from both of the current phase states, or it can be induced by a third system or process. Although we present the model in relation to phase dynamics, a similar functional decomposition of the couplings can also be applied to amplitude state dynamics. In terms of the general theory of phase dynamics [36] (and equation (2.1)), the coupling function q1(ϕ1,ϕ2) can be expressed as the product of two functions 2.3 where Z1(ϕ1) is the phase response curve (PRC) of the first oscillator and shows how it responds to external perturbations, while I1(ϕ2) is the perturbation function through which the second oscillator acts on the first one. (The perturbation function is often given in a more general form like I1(ϕ1,ϕ2) [36]). In terms of equation (2.3), the direct-coupling component results from the existence of the constant part of the PRC Z1(ϕ1), while the common or indirect-coupling component results from the existence of the phase-dependent part of the PRC Z1(ϕ1) and the perturbation function I1(ϕ2).We measured 25 awake and 29 anaesthetized healthy subjects, aged 18–60 years, who were about to undergo elective surgery, all of whom had given their informed consent in writing. The research was approved by the relevant research ethics committees in Norway and the UK. Of 29 anaesthetized subjects, 14 were anaesthetized with propofol and 15 with sevoflurane. There are two sets of recordings for every subject: the first while the subject was awake and resting, and the second while anaesthetized with either propofol or sevoflurane by random choice. Propofol anaesthesia was induced by infusing propofol until a plasma target concentration of 6.0 μgm l−1 was reached [44]. A laryngeal mask airway was inserted 2 min after the start of the infusion. After insertion, the target concentration was reduced to 3.0 μgm l−1 and the infusion was maintained at this rate throughout the measurement period. Some of the propofol patients became restless during induction (while unconscious) and eight of them were given a small dose (50–100 μg) of the very short-lived (Thalf-life≃4 min.) opioid remifentanil during induction. Owing to the small dose and the short half-life, this would not have significantly affected the signals. The other group of subjects were asked to breathe 8% sevoflurane through a close-fitting facemask until an end-tidal concentration of 5% was reached. A laryngeal mask airway was inserted, and then the sevoflurane turned off until the end-tidal concentration fell to 2%. The sevoflurane was then reinstituted to maintain the end-tidal concentration at 2% throughout the measurement period. After a further stabilization period, the anaesthetized set of signal recordings took place. Subjects breathed spontaneously during both sets of recordings. The BIS EEG electrode was placed frontally on the forehead (similar to the FP1 electrode from the 10–20 international system). All data were recorded simultaneously using a Cardio & Brain Signals signal conditioning system (Jožef Stefan Institute, Ljubljana, Slovenia) specially designed for the BRACCIA study. Following 24-bit A/D conversion at 1200 Hz, the signals were stored on a computer for subsequent analysis. They included the three-lead ECG, and the respiration signal measured with a thorax-belt, as well as the frontal EEG signal. All were of 22–32 min duration. The analyses were performed on equal-length segments of 20 min. The signals were first inspected visually, followed by automated artefact removal by interpolation. Data from subjects whose signals had many artefacts were disregarded and not analysed. The cross-frequency intervals were estimated by standard digital filtering procedures, including a FIR filter followed by a zero-phase digital filtering procedure (filtfilt in Matlab) to ensure that no time or phase lags were introduced by the filtering. The boundaries of the intervals extracted from the EEG signal were δ=0.8–4 Hz, θ=4–7.5 Hz, α=7.5–14 Hz, β=14–22 Hz and γ=22–100 Hz; the interval extracted from the respiration signal was r=0.145–0.6 Hz; and the extraction of the heart activity from the ECG signal was h=0.6–2 Hz. Wavelet power and coherence analyses, together with further clinical interpretation, will be presented elsewhere. For the EEG oscillations special care was taken in dealing with frequency spillage between intervals, heart artefacts and powerline artefacts [45]. The cardiac activity has been widely studied through heart rate variability (HRV) analysis [46]. Usually, the HRV signal is constructed by interpolation of the times of the R-peaks marked on an ECG signal, whence the variations in heart rate can be obtained up to a frequency of approximately 0.5 Hz, i.e. up to half of the main (fundamental) cardiac oscillation frequency at approximately 1 Hz. In this study, we focused on the coupled-oscillator approach [36,47], which meant that we required the cardiac main oscillation mode at approximately 1 Hz, which would of course get lost in an HRV estimation. By contrast, by band-filtration of the signal in the interval h=0.6–2 Hz around the main oscillation we were able to obtain well-defined phase estimates with intra-cycle resolution. Hence, we could analyse the phase interactions of the cardiac main oscillation mode, as required; we note, however, that this procedure would have led to the loss of some of the HRV variations, and especially those at the lower frequencies. The phases of the filtered signals were estimated by use of the Hilbert transform [48], and the protophase-to-phase transformation [22] was then applied to the resultant protophases to obtain invariant observable-independent phases. To determine whether the coupling strength and coupling functions were not genuine, i.e. whether they happened by chance, the coupling of each of the relationships investigated was tested against intra- and intersubject surrogates [49]: the former were generated by randomizing the phase signals, and the latter by taking one of the phases from a different subject. In this way, the surrogates should be independent and any apparent coupling from the surrogate phases should be very low. From the large number of investigated relationships, only those exhibiting a statistically significant difference compared with their corresponding surrogates are discussed in the study. Similarly, for simplicity we present only the coupling in the predominant direction because that in the weaker direction was usually insignificant. To assess the statistical difference between groups of awake, propofol- and sevoflurane-anaesthetized subjects (and because of the non-normal distributions), we used the Wilcoxon statistical test, with p<0.05 considered as significant. The couplings were assessed independently; they did not form a statistical family and multiple comparison tests were not used. To present visually the differences between the distributions, we used standard boxplots which refer to the descriptive statistics (median, quartiles, maximum and minimum). The application of dynamical Bayesian inference to bivariate phase signals leads to the parameters of the coupled phase model, from which the coupling functions can then be reconstructed. For clarity, we first present in detail the coupling function for one relationship only—the delta–alpha coupling figure 3. Examples of the delta–alpha coupling function for single representative subjects in their awake and anaesthetized states are shown in figure 3a–c. In the averaged coupling function for all subjects (figure 3d–f) the inter-subject variations are averaged out, and the remaining coupling function signifies a functional form that represents a deterministic law for all of the subjects. Figure 3. Cross-frequency coupling functions between δ and α brain oscillations. Each δ-to-α coupling function qα(ϕδ,ϕα) is evaluated from the α-dynamics and depends on the bivariate (ϕδ,ϕα) phases. (a–c) The coupling functions for one individual subject, while (d–f) the average coupling functions from all subjects within the group. Note that for comparison the vertical scale of coupling amplitude is shown on same interval for (a–c), and then for (d–f). Here, and throughout, we refer to Awake as the state when the subject is awake and resting; and Propofol and Sevoflurane when the subject is anaesthetized with propofol or sevoflurane, respectively. (Online version in colour.)
Comparison of the coupling function shapes for individual subjects (figure 3a–c) with the corresponding averages over all subjects (figure 3d–f) reveals considerable similarity between subjects. The coupling functions for awake resting [50], propofol and sevoflurane (figure 3a–f) are, however, quite different from each other, both in the form and strength of the coupling. The delta–alpha coupling function for the awake state has a relatively complex and varying form, and low amplitude. The coupling functions for propofol and sevoflurane are similar and they look significantly different from those for the awake state. The sevoflurane coupling function has the largest coupling amplitude. The qualitative form of the delta–alpha coupling function (figure 3f) has a sine-like wave form along the ϕδ-axis, while it is nearly constant along the ϕα-axis. This strongly implies that much of the delta–alpha coupling comes from the direct contribution of the delta oscillation. The specific form of the delta–alpha coupling function (e.g. figure 3f) reveals the underlying functional coupling mechanism, i.e. it shows that, when the delta oscillations are between π and 2π, the sine-wave coupling function is higher and the delta activity accelerates the alpha oscillations; similarly, when the delta oscillations are between 0 and π, the coupling function is decreased and delta decelerates the alpha oscillations. In figure 4, we summarize our results for the coupling functions of all significant coupling relationships. They include the cross-frequency coupling functions that emerge within the brain, and between the brain, the lungs and the heart oscillations (figures with enhanced resolution are provided in the electronic supplementary material). The delta–alpha relationship is presented again for completeness and comparison. The theta–gamma coupling functions (figure 4e–h) have different forms, depending on the state of awakeness, with propofol and sevoflurane taking similar forms. The coupling amplitude of the propofol theta–gamma coupling (figure 4f) is lowest. The form of the functions looks like a second-order sine wave which changes predominantly along the ϕθ-axis. The alpha–gamma coupling functions (figure 4i–l) are of similar form, but their coupling amplitudes increase in anaesthesia, with the sevoflurane coupling function again being the highest. Interestingly, the qualitative form of the functions changes along both axes. This implies that the alpha–gamma coupling depends on both of the oscillations (alpha and gamma), or on the same indirect influence that affects them both. Figure 4. Cross-frequency coupling functions between neuronal and cardiorespiratory oscillations: (a–d) δ–α; (e–h) θ–γ; (i–l) α–γ; (m–p) r–θ; (q–t) h–θ and (u–x) r–h. The coupling functions are arranged in columns, and the states and surrogates are aligned horizontally. The coupling functions shown are the average over all subjects within a group and the vertical coupling scales are the same for each state within a relationship. The notation and interpretation of the individual coupling functions are the same as in figure 3. (Online version in colour.)
The influence of respiration on brain theta oscillations is shown in figure 4m–p. There exist similarities in the form of the coupling functions between the awake and sevoflurane states, while the form of the propofol function seems qualitatively different. The direct influence of the phase ϕr of respiration is dominant in the awake and sevoflurane coupling functions. The coupling of the heart to theta oscillations is weak with a less-stable and time-varying form (figure 4q–t). The two anaesthetized heart–theta coupling functions are of similar form and are stronger than in the awake state. The strong coupling function between respiration and the heart oscillations (figure 4u–x) is the only one to have been studied previously, and our results confirm the earlier work [19–21]. More importantly, the propofol and sevoflurane anaesthesia made the form of the cardiorespiratory coupling function more time-varying and unstable—which is opposite to the effect of anaesthesia on the delta–alpha coupling (cf. figure 4a–d). In order to assess the influence of anaesthesia, we first quantify the coupling (amplitude) strength. The latter has been extensively studied in earlier work [13,15,45,51]: wherever reference was made to coupling causality and directionality, it was in fact the net coupling strength, or a measure proportional to it, that was being evaluated. Our coupling decomposition enables us to go beyond this by quantifying the coupling strengths of the individual components of the net coupling. In figure 5, we summarize the changes of coupling strength induced by anaesthesia. The different effect on the separate coupling components is evident in the delta–alpha relationships shown in the top row of figure 5a–c. The net coupling with sevoflurane is significantly different from the awake and propofol states (figure 5a); for direct coupling all the states are different (figure 5b), while the indirect coupling for propofol was significantly the smallest (figure 5c). Note also that direct coupling is the dominant component of the net coupling. For the theta–gamma interaction, it is only the indirect coupling that differs between the awake and sevoflurane states (figure 5f). Anaesthesia increased significantly the net and indirect coupling strengths in alpha–gamma (figure 5g–i). This coupling is mostly defined by the indirect coupling component. The respiration–theta net coupling differed slightly between the two anaesthetics (figure 5j). Sevoflurane anaesthesia induced the strongest cardiorespiratory coupling strength, and this difference compared with other states is significant for all coupling types (figure 5m–o). Figure 5. Anaesthesia-induced changes in coupling strength. Each boxplot shows the coupling strength distribution of a specific coupling relationship for the awakeness state indicated by the letter A (awake), P (propofol) or S (sevoflurane) on the abscissa. The coupling relationships are shown on the vertical axes, with each interaction as a separate row, including δ–α shown in (a–c), θ–γ (d–f), α–γ (g–i), r–θ (j–l) and r–h (m–o). The h–θ row is omitted because there were no significant changes. The columns correspond to the net, direct and indirect coupling components, respectively. The line connectors on the tops of individual panels indicate cases where the difference between two boxplot distributions was statistically significant (for statistical procedures see §2d). (Online version in colour.)
The other useful characterization of coupling functions is their functional form. It defines the functional law or mechanism and it is a specific feature of coupling functions. To quantify the forms of a given coupling relationship, we use a correlation measure that quantifies the similarity of the forms of two coupling functions, irrespective of their coupling strengths [20]. If the similarities of form for between the intersubject pairs for some interaction is high enough, it means that there exists a common deterministic functional form which underlies the mechanisms of that interaction. From the coupling decomposition model, we can investigate, separately, the similarity of form for each individual component of the coupling functions (figure 6). Figure 6. Influence of anaesthesia on the form of the coupling functions. The similarity of functional forms is presented as the correlation coefficient ρ for the net ((a), (d) and (g)), the direct ((b), (e) and (h)) and the indirect coupling functions ((c), (f) and (i)). The columns correspond to the three coupling relationships: (a–c) δ–α; (d–f) r–θ and (g–i) r–h. The inter-subject similarity correlation boxplots are shown for the awake (A), propofol (P) and sevoflurane (S) states as indicated on the abscissa. The line connectors on the tops of individual panels indicate cases where the difference between two boxplot distributions was statistically significant. (Online version in colour.)
The similarity in form of the delta–alpha coupling functions is shown in figure 6a–c. There is a large difference due to anaesthesia in the net and direct similarity (figure 6a,b), while the indirect similarity is different only for sevoflurane (figure 6c). The similarity of form is especially high for the direct component, while very low for the indirect component. The respiration–theta interaction had relatively small similarity of its functional forms, and there is only a small significant increase for sevoflurane in the net and direct similarities (figure 6d,e). The respiration-heart interaction also had all the significant differences seen in the net coupling, but now decreased with anaesthesia (figure 6g). The similarity of these interactions is mostly due to the high direct similarity (figure 6h), where awake is different from when under the two anaesthetics. We note that anaesthesia had opposite effects on the similarity of the functional forms for delta–alpha and respiration-heart—cf. figure 6a,g. This quantitative description is consistent with the observations of the coupling functions made in figure 4. Dynamical Bayesian inference can decompose the dynamics into two parts: what is believed to be the deterministic part of the model; and a part originating from random (white) noise perturbations. The noise strength represents the level of random fluctuations relative to the frequency of the oscillation and its interactions with the other oscillations considered. So we also investigated whether and how anaesthesia affects the noise strength D of the brain and cardio-respiratory oscillations, with results as shown in figure 7. Correlated noise strengths, e.g. Dα,δ were found to be very small and not statistically different between the awakeness states, so they are not reported. Also, the noise strength for each of the intervals had (qualitatively) the same statistical difference when coupling was investigated with different intervals, e.g. Dα,αfigure 7b was the same whether δ−α interactions or α−γ interactions were inferred. Figure 7. Anaesthesia-induced changes in noise strength. Each boxplot shows the group distribution of all subjects' noise strengths for a specific oscillation interval during the three awakeness states A, P and S. Noise strengths are shown (a) for the δ oscillation interval, (b) α, (c) θ, (d) γ, (e) r and (f) h. The line connectors on the tops of individual panels indicate cases where the difference between two boxplot distributions was statistically significant. (Online version in colour.)
The results in figure 7 demonstrate that the noise strength for some rhythms was unaffected by anaesthesia, Dδ,δ in figure 7a and Dh,h in figure 7f; for other rhythms anaesthesia made a significant difference, either increasing like Dθ,θ in figure 7c, or decreasing like Dα,α in figure 7b and Dγ,γ in figure 7d, with anaesthesia relative to the awake state; or the result was statistically different in all three states, like Dr,r in figure 7e. The present investigation relies on three complementary factors: (i) general anaesthesia under either intravenous and inhalational anaesthetics; (ii) the novel methodology of cross-frequency coupling functions to probe interaction mechanisms directly; and (iii) assessment of the combined dynamics and interactions of the cortical, respiratory and cardiac oscillations. We have thus been able to analyse the coupling functions between brain activity, which involves information processing and control of the human body, on the one side, and the cardiorespiratory systems, which take care of energy transport and the supplies of nutrients and oxygen, on the other. While interactions have already been studied in lizards [52], mice [53], rats [15] and dogs [54], here we report the first insights into cardio-respiratory-cortical interactions in humans, in both the awake and anaesthetized states. Moreover, our extension of cross-frequency coupling to include the analysis of coupling functions has allowed us to investigate the interactions in greater depth by introducing the notion of the functional form, which represents a new dimension in the analysis of neuronal effective interactions. Thus, we have been able to present two quantitative dynamical properties of the phase interactions: the coupling strength and the form of the coupling function. The functional description of the couplings has enabled us to propose a coupling decomposition model that reveals the separate contributions, in turn providing deeper insight into the causality within a coupling. The model was strongly supported by the results (figures 5 and 6) where the effect of anaesthesia often differed for the individual coupling components. Coupling functions describe the underlying mechanisms that gives rise to the qualitative states of the interacting systems, e.g. the phase synchronization state which is of great importance in neuronal [37] and cardiorespiratory [20,29] interactions. By knowing the form of the coupling function, one can predict the occurrence of phase synchronization for given parameters [16]. Although the discussion was only for large-scale cross-frequency couplings, the coupling functions presented have wide implications at different scales and levels of the heavily connected brain network [35]. Thus one can also describe the functional form of the edges, and can use the coupling decomposition model to investigate the separate contributions from the nodes. One of the most prominent coupling relationships we identified is delta–alpha. It reflects how delta activity, associated with deep dreamless sleep [55], influences the alpha oscillations which are said to reduce the information processing [55,56] and play a key role in consciousness [57,58]. During the maintenance of general anaesthesia, the alpha and delta activities were increased [2,59]. The delta–alpha coupling has been linked to the coding mechanism of feedback valence information [60]. Even though the anaesthetized state differs from sleep and from the resting state generally, a strong delta–alpha link was observed during non-REM sleep [30] and recently it was suggested that delta–alpha coupling is mostly located within the frontal and the parieto-occipital regions when it is stronger during the eyes-closed state [13]. Our results are consistent with, and further extend and deepen, these findings. Namely, the form of the delta–alpha coupling functions (figures 3 and 6) indicates that the influence is direct modulation from delta to alpha, where the couplings are significantly stronger in anaesthesia than when awake. This shows that, once the subject is anaesthetized, delta activity influences the alpha oscillations by contributing to the reduction of information processing and integration. Gamma activity, associated with attention, memory and sensory processing, is known to decrease in anaesthesia [61]. In seeking to reveal the underlying mechanisms, we identified two significant couplings: theta–gamma and alpha–gamma. They have been widely studied already, mostly with phase-to-power cross-frequency couplings and higher gamma intervals, and various functional roles have been attributed to them in different states and tasks [62,63]. It has been suggested that theta–gamma coupling plays a prominent role in memory tasks, whereas alpha–gamma interactions are more important for attention processing [64]. The coupling function analysis (figure 4) indicated that these two couplings are affected differently by anaesthesia. Namely, propofol decreased and sevoflurane increased the theta–gamma coupling, while both anaesthetics increased the alpha–gamma coupling (figures 4 and 5). These two couplings evidently have different functional mechanisms. The theta–gamma couplings in anaesthesia result from the direct influence of theta on gamma, while alpha–gamma is dominantly an indirect coupling, implying that there might be a third process which influences both of the oscillations. We extended the analysis of cross-frequency neuronal couplings to include the interactions of two important parts of the cardiovascular system—the respiration and the heart [4,15]. We identified a coupling function from respiration to theta oscillations. The coupling function was of complex form, with strong direct component and relatively low intensity. The respiration–theta coupling was affected more by the sevoflurane than the propofol anaesthesia. Of special interest are the brain–heart interactions as they have been linked to cardiac arrhythmias, psychophysiological coordination and vascular dementia [65–67]. Our analysis identified a coupling function from the heart to the brain theta oscillations. The form of the coupling function was relatively complex, its intensity was not very high, and the influence was predominantly with a direct component from the heart to the theta oscillations. This coupling function was not greatly affected by the onset of anaesthesia. The origin of the cardiac–theta couplings could be linked to the haemodynamic function of the heart in providing blood, together with oxygen and other metabolic substances, to the brain. Astrocytes and other glial cells might be responsible for mediation of these processes on the neural level [68,69]. The cardiorespiratory coupling function has been extensively studied [19–21] and its direct coupling component [21] and phase resetting curve [20] have been associated with respiratory sinus arrythmia. The functional connectivity of cardiorespiratory interactions was affected in different ways by propofol and sevoflurane anaesthesia [29]. Interestingly, we found that the effect of anaesthesia on the cardiorespiratory coupling functions showed that the coupling strength increased with anaesthesia, whereas the similarity of form decreased (cf. figures 5m and 6g). This indicates that the inter-subject similarity of forms becomes more varied with anaesthesia, while maintaining stable and strong interactions—perhaps reflecting the chronotaxic nature of the cardiorespiratory interactions [70]. These alterations of the cardiorespiratory coupling functions and their links to the theta brain oscillations (figures 5 and 6) may reflect partially the onset of analgesia and the reduced perception of pain [31,32], with possible links to consciousness. Therefore, such results could have implications for the quest of quantifying analgesia in the absence of consciousness [71,72]. The noise strength analysis in figure 7 shows that anaesthesia changes, not only the deterministic couplings, but also some of the random fluctuations acting on the oscillations. The decrease of the noise level in α, γ and respiratory oscillations (figure 7b,d,e) might be a consequence of the higher determinism associated with the onset of anaesthesia which induces, e.g. order, coupling and coherence of the oscillations ([2,15] and figure 5). More puzzling is the result that the noise strength for θ oscillations increased with anaesthesia, figure 7c. It may perhaps be linked to the origin of the θ oscillation and its role in the hippocampus [73]. These results are intriguing and invite further investigation using dynamical Bayesian inference, which has clearly demonstrated its potential for studies of this kind as well as for the analysis of (biological) experiments of a stochastic nature quite generally. Unconsciousness is the most striking change in the state of a subject when general anaesthesia occurs [1,2,74]. The transition to unconsciousness and back can be traced through assessment of the cognitive EEG dynamics [9,10,75] and the recovery of consciousness has been found to differ in elderly subjects [76]. It has been noted that the standard clinical assessments of consciousness (motor, verbal and eye-opening responses [77]) are not sufficient and that there is a need for techniques which also assess the function and effective connectivity [1]. Our statistical comparisons of coupling functions in the awake and anaesthetized states demonstrate that there are significant differences, especially for the delta–alpha and alpha–gamma couplings (figures 3, 5 and 6). These coupling-induced changes of the phase advanced/delayed oscillations alter the attention and memory processes, and suppress information integration which is known for mediating the unconscious state [1]. The roles of propofol and sevoflurane in the induction of unconsciousness as a common mechanism was studied and power differences were outlined [78,75]. Our coupling function results have revealed that these anaesthetics often exhibit similar functional forms, perhaps implying similar mechanisms (figures 3 and 4), but that there are some quantitative differences (figures 5 and 6). In general, we observe similar forms of coupling function, but the strength and effect were significantly stronger for sevoflurane. This could be on account of the doses used. It is also possible that the molecular and neuronal processes associated with propofol and sevoflurane are largely similar, perhaps because both act on the same receptor (e.g. GABAA) but that there are minor differences in relation to the potassium channels affected [1,5,6]. In conclusion, coupling functions have enabled us to unveil a new perspective on how the neurophysiological mechanisms are affected by general anaesthesia. This initial application has been in a sense overwhelming in that we have identified six important and very illuminating coupling relationships. This was partly because we analysed not only neuronal oscillations, but also how the latter are affected by cardiorespiratory activity. The work has opened the door to a host of new questions and problems needing to be tackled. For example, can one apply coupling function analysis to assess spatial neuronal couplings using additional EEG electrodes, perhaps using different anaesthetics? The possibility of following time-evolving dynamics could lead to new insights based on studies of how the evolution of the coupling functions mechanisms lead to unconsciousness. Coupling functions can also be used to study the mechanisms of other neurophysiological perturbations, as well as to revisit known problems, states and diseases in order to reveal the underlying functional mechanisms. Needless to say, the findings and the methodology of this work also have wide implications for coupled oscillators in general, with the possibility of biomimetic [79] solutions to a diversity of difficult problems [80,23]. T.S. analysed the data, contributed to the interpretation and wrote the manuscript. S.P. analysed the time-frequency properties of the signals and helped define phases within each oscillatory interval. J.R. was a clinical group leader, managed the measurements in the Oslo University Hospital, and contributed to the physiological interpretation of the results. A.F.S. was a clinical group leader and managed the measurements in the Royal Lancaster Infirmary. P.V.E.Mc.C. helped in the execution and management of the research and editing the manuscript. A.S. conceived the study, was Coordinator of the EC FP6 grant BRACCIA and the EPSRC grant on Non-autonomous Dynamics that funded the research, was involved in discussion and interpretation of the results and edited the manuscript. The author(s) declare they have no competing interests. This work was supported by the European Union as a NEST (New and Emerging Science and Technology) Project, no. 517133, ‘Brain, Respiratory and Cardiac Causalities in Anaesthesia’ (BRACCIA), by the Engineering and Physical Sciences Research Council (EPSRC) UK (grant no. EP/100999X1) and in part by the ARRS Slovenia (program no. P20232). We thank the entire BRACCIA team for excellent collaboration and in particular to J. Petrovčič for designing the Cardio & Brain Signal Conditioning Unit, to B. Musizza for creating a LabView software platform to collect the data and to T. Draegni, S. A. Landswerk, P. Kvandal, D. A. Kenwright, L. W. Sheppard and M. Entwistle for their help in collecting the data. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. © 2016 The Authors. Published by the Royal Society under the terms of the Creative Commons Attribution License http://creativecommons.org/licenses/by/4.0/, which permits unrestricted use, provided the original author and source are credited. References
Page 7The autonomic nervous system (ANS) and the central nervous system (CNS) are strictly interconnected through anatomical and functional links, and influence each other continuously [1–4]. As an example, cortical and subcortical brain areas including the amygdala, insular cortex and pregenual anterior cingulate cortex play crucial homeostatic–interoceptive functions involving ANS dynamics [1,5]. Moreover, such cingulate cortex and amygdala activities are involved in regulating the sympathovagal balance [3,4]. On the other hand, changes in ANS signalling relevantly affect the CNS, both in physiological and in pathological conditions [2,4,6–9]. Dysfunctions of the ANS were found in acute and chronic stressful conditions [10–12], insomnia [13,14], epilepsy [15,16], parkinsonisms [17,18], psychosomatic disorders [19] and schizophrenia, anxiety and mood disorders [6,20–24], which are typically considered CNS-related conditions. Moreover, vagal nerve stimulation has been shown as an effective treatment for major depression [25,26], while relaxation techniques based on cardio-feedback are used for managing negative emotions and psychological symptoms [27,28]. A typical and paradigmatic brain–heart interaction occurs during an emotional experience. Human emotions involve several areas for their perception and processing. The prefrontal cortex and amygdala specifically represent the essence of two specific pathways. The prefrontal cortex encodes the affective elicitations longer than 6 s transmitting the related information to other areas of the Central Autonomic Network [29], whereas the amygdala encodes the briefly presented stimuli. Indeed, previous studies on emotions were mainly carried out investigating brain and neurovegetative activities [30–33]. As an example, estimates of vagal activity alone predicted neural responses during subjective rating of fearful faces [33], whereas heart rate (HR) increases were predicted by the level of activity of interconnected brain regions, including the amygdala, insula, anterior cingulate and brainstem during visual perception of emotional facial expression [34]. Concerning pathological mental states, mood disorders were linked to Takotsubo cardiomyopathy [35], which is one of the brain–heart disorders, whereas depressive states were associated with a functional disconnection between rostral anterior cingulate cortex and autonomic brainstem nuclei [36]. The scientific debate on the physiological origin of emotions is still open: whether they originate from the peripheral reactivity of the ANS, or from specific areas of the brain, or from both. As Damasio stated, ‘emotions are the most complex expression of homeostatic regulatory systems’. He hypothesized that emotions (or emotional memories) can modify our behaviour through conscious or unconscious signals [37], p. 86. Note that the latter belongs to ANS signalling whose role is to generate re-entry vegetative information to pre-existing cortical maps [37,30]. Although it is reasonable to hypothesize that emotions and emotional reactivity strongly affect brain–heart coupling, how such a brain–heart dynamics is further modulated by the specific kind of an emotional stimulus is still unknown. Considering emotions as continuous traits, each state can be described and mapped in a multidimensional space, portraying different psychophysiological and neurobiological underpinnings [30,38,39]. According to the circumplex model of affect (CMA) [40], emotions can be mapped in two dimensions through a combination of arousal and valence levels [30]. Valence refers to the pleasantness or unpleasantness of an emotion, whereas arousal refers to the intensity of the emotional stimuli, expressed in terms of degree of activation from low to high. Importantly, CMA assumes that these dimensions are orthogonal, thus with no mutual influence (or interaction) among them. At a peripheral level, the study of emotional responses is especially related to the analysis of heart rate variability (HRV) [30–32,41,42]. This is justified by the fact that oscillations of HRV above 0.15 Hz (i.e. the high-frequency band) are exclusively mediated by vagal activity [41,42], and oscillations below 0.15 Hz (i.e. low-frequency band) are mediated by both vagal and sympathetic activities [43]. At a central level, emotions have mainly been studied through functional magnetic resonance imaging, and continuous electroencephalographic (EEG) and evoked related potentials recordings [31,44,45]. Previous studies investigated the coupled brain–heart dynamics during healthy and pathological emotional responses (see reviews in [46,47]), highlighting connections in the vagally mediated regulation of physiological, affective and cognitive processes. As a general approach, previous studies have tried to link the EEG power in specific bands to HRV measures. Although significant correlations were found for the α (8–12 Hz) [48–52], β (13–30 Hz) [49,50,53] and γ (>30 Hz) bands [53,54], the psychophysiological meaning of such associations is still ill-defined. For instance, complexity of HRV series was used to predict changes in the EEG α band after stress [48]. However, physiological correlates of HRV complexity are still unknown. On the other hand, the link between the EEG θ band (4–8 Hz) and ANS activity is quite consistent. Specifically, in healthy controls, both sympathetic- and parasympathetic-related parameters were correlated with EEG θ power in temporal areas [55]. Moreover, the frontal θ power has been demonstrated to be sensitive to emotions [56–60]. Unpleasant music evoked a significant decrease of HR associated with an increase of frontal midline θ power [61], whereas θ event-related synchronization were found to occur in frontal regions of the brain during the earliest phases of affective auditory stimuli processing [58]. In response to negative emotional patterns, EEG activity in the θ band was associated with the right prefrontal cortex activity, following an increase in the sympathetic response [62]. Furthermore, a positive correlation between HRV high-frequency power and EEG frontal midline θ power was found during meditation [63]. The same kind of correlation was found between HRV low-frequency power and EEG posterior θ power during biofeedback task [64]. More in general, the EEG θ power was associated with a general emotional response and states of relaxation and internal attention [63], whereas alterations of EEG dynamics in the θ band were found in case of ANS dysfunctions [55]. Cardiac and brain dynamics were also quantitatively assessed during sleep in the frame of dynamical information theory [65], highlighting the role of EEG low-frequency bands in the ‘from-brain-to-heart’ information transfer. Limitations of the above-mentioned literature can be summarized into two categories: (i) correlation measures quantifying brain–heart coupling were performed at a groupwise level, considering few sample measures for each subject and, thus, totally disregarding intra-recording time-varying brain–heart dynamics; (ii) previously proposed correlation measures have considered only linear couplings, thus disregarding intrinsic nonlinear brain–heart interactions. To overcome these limitations, in this study we provide a unique insight into the brain–heart dynamics during emotion perception in healthy subjects, showing experimental results using high-resolution EEG signals (128 channels) and instantaneous HR estimates. We present here a novel approach to study brain–heart interactions, quantifying the linear and nonlinear brain–heart coupling mechanisms through the calculation of the maximal information coefficient (MIC) index [66], a statistical method for detecting associations between pairs of variables. Importantly, MIC calculations were performed at a single-subject level, between time-varying estimates of high-resolution EEG power spectra and instantaneous HR, during visual emotional elicitation. Twenty-two healthy volunteers were emotionally elicited through passive viewing of pictures taken from the International Affective Picture System (IAPS) [67], associated with 25 different combinations of arousal and valence levels, including neutral elicitations. More in detail, EEG data were processed using short-time Fourier transform representations in order to obtain time-varying maps of cortical activation, whereas the associated instantaneous cardiovascular dynamics was estimated through inhomogeneous point-process models of RR interval series [68], which were gathered from the electrocardiogram (ECG). The use of inhomogeneous point-process on heartbeat dynamics allows to obtain instantaneous time domain and spectral estimates, which can be considered as covariate measures of brain–heart interaction during emotional processing. Details on the inhomogeneous point-process modelling can be found in [68–70]. Briefly, we model the probability function of the next heartbeat given the past R-events. The probability function is fully parametrized to model its first-order moment. Importantly, as the probability function is defined at each moment in time, the parameter estimation is performed instantaneously. In particular, the linear terms allow for instantaneous time domain and spectral estimation. Recently, using point-process modelling, we reported on how to recognize emotional valence swings (positive or negative), as well as two levels of arousal (low-medium and medium-high), using heartbeat data only, being also able to instantaneously assess the subject's state even in short-time events (less than 10 s) [70]. In other words, emotional stimuli with high/low arousing and high/low valence levels produce changes in ANS dynamics, through both sympathetic and parasympathetic pathways, that can be tracked by a multidimensional representation estimated in continuous time by the proposed point-process model [70]. In the rest of the paper, as a proof of concept of the proposed methodology, we particularly focus on the experimental results gathered analysing EEG oscillations in the θ band, and its coupling with instantaneous heartbeat measures. Although we are aware that oscillations in other EEG frequency bands may be coupled with heartbeat dynamics, we focused on these low-frequency oscillations because the emotional processes consistently elicit changes in the θ band, together with the ANS responses [55–65]. However, for the sake of completeness, experimental results related to all of the EEG frequency bands, including α, β and γ bands, are reported in the electronic supplementary material. The recording paradigm related to this work has been previously described in [70]. As mentioned in the Introduction, we adopted a common dimensional model which uses multiple dimensions to categorize the emotions, the CMAs [40]. The CMA used in our experiment takes into account two main dimensions conceptualized by the terms of valence and arousal. Accordingly, we employed visual stimuli belonging to an international standardized database (IAPS) [67] having a specific emotional rating expressed in terms of valence and arousal. The IAPS is one of the most frequently cited tools in the area of affective stimulation and consists of a set of 944 images with emotional ratings obtained using the self-assessment manikin (the arousal scales ranging from 0 to 10, and valence scales ranging from −2 to 2). An homogeneous population of 22 healthy subjects (aged from 21 to 24), not suffering from cardiovascular, neurological and mental disorders, was recruited to participate in the experiment. All subjects were naive to the purpose of the experiment. The experimental protocol for this study was approved by the ethics committee of the University of Pisa and informed consent was obtained from all participants involved in the experiment. All participants were screened using the Patient Health QuestionnaireTM, and only the ones with a score lower than 5 were included in this study [71]. A general overview of the experimental protocol and analysis is shown in figure 1. The affective elicitation was performed by visualizing the IAPS pictures onto a PC monitor. Experimental protocol began with a resting session of 5 min with the eyes closed (session B). Then, the slideshow started, comprising nine sessions, alternating neutral sessions (from N1 to N5) and arousal sessions (from A1 to A4) (figure 1). One-minute resting-state sessions (from R1 to R8) were in between each neutral/arousal session. Neutral sessions consisted of six images having valence range ( Figure 1. Sequence scheme over time of image presentation in terms of arousal and valence levels. The y-axis relates to the official IAPS score, whereas the x-axis relates to the time. After the first 5 min baseline (b) acquisition, the neutral sessions (Nx) alternate with resting-state sessions Rx and the arousal ones Ax. Along the time, the red line followsthe four arousal sessions having increasing intensity of activation. The dotted green line indicates the valence levels, distinguishing negative and positive levels within an arousing session. (Online version in colour.)
For the whole experimental session, high-resolution 128 channels EEG and ECG signals were acquired through the Geodesic EEG Systems 300 from Electrical Geodesics, Inc. The sampling frequency was set at 1 kHz. The average of mastoid signals was used as reference. This monitoring device allowed for a rapid application and comfortable fit of the cuff. All available EEG signals were taken into account for MIC analysis between every HRV time-varying feature. However, for the sake of conciseness, signals exclusively gathered from electrodes placed on the scalp at standardized positions Fp1, Fp2, F7, F3, Fz, F4, F8, T7, C3, C4, Cz, T8, P7, P3, Pz, P4, P8, O1, O2, according to the International Standard System 10–20, were considered for post hoc statistical analyses. EEG signals were filtered by a sixth-order infinite impulse response bandpass filter with cut-off frequencies of 1–45 Hz. The Matlab toolbox EEGLAB [73] was used for the entire processing of the EEG data. EEG spectral analysis was performed using discrete fast Fourier transform to estimate the power spectra in 4 s moving windows, with 75% overlapping, within the classical frequency bandwidths: θ (4–8 Hz), α (8–14 Hz), β (14–30 Hz), γ (30–40 Hz). Frequency band δ (less than 4 Hz) was not taken into account in this study because its related to deeper stages of sleep. EEG signal pre-processing consisted in three main steps: removal of head/body movement-related artefacts; removal of eye movement artefacts; and ‘bad’ channels identification. Concerning head/body and eye artefacts, we have simply excluded all contaminated epochs from further analyses. This choice was a cautious approach aimed to avoid any effects of artefact-related residuals following the artefact-removal procedures, which could alter true EEG power and synchronization estimates. Looking for head/body movement-related artefacts, all EEG channels were analysed in order to find synchronous, sudden increases in signal amplitude. We classified EEG epochs with amplitude exceeding the threshold of the 95th percentile of the signal amplitude distribution as epochs likely to contain movement or muscular artefacts. After confirmatory visual inspection, such EEG epochs were discarded. Eye movement artefacts were detected by computing a moving-window cross-correlation between the frontal EEG channels and electro-oculogram: high values of cross-correlation were marked as ocular artefacts. We considered cross-correlation values as significantly high if greater than a specific threshold value. Such a value was derived by computing the same moving-window cross-correlation between phase-randomized surrogated [74] frontal EEG channels and the electro-oculogram. Furthermore, we considered only artefacts producing fluctuations greater than 50 μV on frontal EEG channels, lasting at least 70 ms. Epochs marked as artefact-corrupted were tagged and, after visual inspection, definitively discarded [75]. ‘Bad’ channels identification refers to the detection of low-quality EEG signal, frequent unexpected events and presence of high-frequency noise [73]. To this aim, for each channel, we calculated the second-, third- and fourth-central moments and identified the ‘bad’ channels as the outliers present in such a three-dimensional space. Good channels, in fact, usually cluster together, whereas the bad ones drift apart in different directions according to their artefactual nature (for example, channels highly contaminated by the power-line have lower kurtosis than other channels). For each dimension of this space, channels distant more than twice the interquartile range from the cluster centroid were classified as artefactual and, after visual inspection, were discarded. The ECG signal was analysed off-line to extract the RR intervals [41]. Firstly, ECG was pre-filtered through a moving average filter in order to extract and subtract the baseline. Then, a QRS complex detection algorithm was used. We adopted the automatic algorithm developed by Pan–Tompkins [76]. This algorithm allowed us to extract each QRS complex and to detect the corresponding R-peak. Erroneous and ectopic beats were corrected by a previously developed algorithm, based on the point-process modelling [77]. Starting from the RR interval series, instantaneous time and frequency domain features were estimated through point-process modelling [68,69,78,79]. The point-process framework primarily defines the probability of having a heartbeat event at each moment in time. A parametric formulation of the probability function allows for a systematic, parsimonious estimation of the parameter vector in a recursive way and at any desired time resolution. Instantaneous cardiovascular indices can then be derived from the parameters in order to quantify important features as related to cardiovascular control dynamics. Mathematically, let (0,T] denote the observation interval and 0≤u1<⋯<uk<uk+1<⋯<uK≤T the times of the events. For t∈(0,T], let Given a set of R-wave events 2.1 whereHere, we model 2.2 Since The model goodness-of-fit is based on the Kolmogorov–Smirnov (KS) test and associated KS statistics (see details in [68,69]). Autocorrelation plots were considered to test the independence of the model-transformed intervals [68,69]. Once the order p is determined, the initial model coefficients were estimated by the method of least squares [68,69]. Our framework allows for a quantitative characterization of heartbeat dynamics based on instantaneous time- and frequency-domain estimations. The time-domain characterization is based on the first- and the second-order moments of the underlying probability structure. Namely, given the time-varying parameter set ξ(t), the instantaneous estimates of mean 2.3 2.4 2.5 The linear power spectrum estimation reveals the linear mechanisms governing the heartbeat dynamics in the frequency domain. In particular, given the model of 2.6 where H1 represents the Fourier transform of the γ1 terms. By integrating equation (2.6) in each frequency band, we compute the indices within the low frequency (LF = 0.04–0.15 Hz) and high frequency (HF = 0.14–0.45 Hz) ranges, along with their ratio (LF/HF).In order to quantify the coupling between two variables, we calculated the MIC [66]. This index, in fact, is able to quantify linear and nonlinear couplings occurring between two variables over time, x and y [66]. MIC relies on the fact that, if two variables are somehow coupled, then a grid can be drawn on the scatterplot of the two variables. Formally, let Gx×y indicate all the possible partitions with x rows and y columns of the scatterplot for the ordered pairs of two vectors, and Ig the mutual information for one specific partition with x×y grids that are applied to the ordered samples of the two vectors. MIC is defined as the maximal value of mx×y over the ordered pairs (x,y), with x≤n and y≤n, where n is the length of the vectors: 2.7 In practice, it is possible to compute the estimation as Statistical analyses were performed on MIC values, considering data gathered from all of the subjects, in order to perform (i) comparison between resting state sessions; (ii) comparison between neutral sessions; (iii) comparison between (whole) arousal sessions; (iv) comparison between pleasant and unpleasant stimuli of each arousal sessions; and (v) comparison between pleasant/unpleasant stimuli among different arousal levels. These statistical comparisons were performed, for each EEG signal, between each pair of time-varying EEG-PSD, calculated in the θ, α, β and γ frequency bands, and each of the time-varying HRV features. All HRV features were instantaneously calculated with a δ=5 ms temporal resolution, and then averaged each within a 1 s sliding time window, in order to achieve temporal correspondence with the time-varying EEG-PSD series. Before performing the statistical analysis, we implemented the Lilliefors (a KS-based approach) test to check whether the data were normally distributed. As most of the samples taken into account did not show a normal distribution, non-parametric rank-based statistical analyses were carried out. Moreover, central tendency and dispersion of samples were expressed in terms of median and median absolute deviation, respectively. In particular, according to the number of sample groups under comparison, non-parametric Wilcoxon and Friedman tests for paired data were applied, considering the null hypothesis of equal medians among samples. Statistical significance was set considering p-values <0.05. Results from the statistical analysis on MIC values are shown as topographic colourmaps. After applying the EEG preprocessing steps (see §2b), all recordings showed more than 90% of artefact-free epochs. Moreover, among all subjects, up to eight EEG channels were discarded after the ‘bad’ channels identification procedure (on average, five channels were removed). Concerning the application of the point-process modelling on RR series, excellent results were achieved in terms of goodness-of-fit. To this extent, optimal model order was found to be p=7. KS distances were as low as 0.0328±0.0052 and were never above 0.051. Moreover, the independence test performed through autocorrelation plots was verified for all subjects [69], showing that in 21 out of 22 recordings, more than 98% of the autocorrelation samples fell within 95% CIs. A total of 16 out of 22 recordings showed KS plots within 95% CIs. Figure 2 shows p-value topographic maps resulting from the statistical comparison of MIC values between resting-state sessions, neutral sessions and arousal sessions, considering each pair of time-varying HRV features and time-varying EEG-PSD calculated in the θ band. Differences between resting-state sessions were found in coupling with μRR, σHR and HF only. Differences in all of the couplings between EEG θ band and HRV features, instead, were found between all neutral and between all arousing sessions. In particular, major differences between neutral sessions occurred in the prefrontal cortex, mainly due to significant changes occurring in the neutral session 5, whereas major differences between arousal sessions occurred in the prefrontal cortex and parietal lobes, mainly due to significant changes occurring in the intermediate arousal session 3. Note that, in the experimental protocol timeline, arousal level 3 follows neutral session 5. Similar EEG–HRV coupling maps, calculated for the EEG α, β and γ bands, are reported in electronic supplementary material, figures S1–S3. All MIC values, expressed as median and its absolute deviation, gathered during the first 5 min baselines, are reported in electronic supplementary material, tables S1–S4 , for each EEG frequency band. Likewise, MIC values of resting-state sessions are in electronic supplementary material, tables S5–S8, MIC values of arousal sessions in electronic supplementary material, tables S9–S12, and MIC values of neutral sessions in electronic supplementary material, tables S12–S16. Figure 2. p-Value topographic maps resulting from the statistical comparison (Friedman non-parametric tests) of MIC values between resting-state sessions (a), neutral sessions (b) and arousal sessions (c), considering each pair of time-varying HRV features and time-varying EEG–PSD calculated in the θ band. Blue regions are associated with no significant difference between sessions, whereas green/yellow/red activations are associated with significant differencesof the brain–heart coupling in at least two of the considered sessions. Each p-value topographic map is reported along plots of significantly different MIC values, and significant differences among sessions. Such significant differences (sig.) result from multiple comparison analysis considering a Bonferroni correction.
Figure 3 shows p-value topographic maps resulting from the statistical comparison (Friedman non-parametric tests) of MIC values between all of the positive, and all of the negative elicitation sessions, considering each pair of time-varying HRV features and time-varying EEG-PSD calculated in the θ band. Significant arousal-dependent differences were found between positive elicitation sessions, especially affecting the EEG θ–HRV-LF coupling in the left temporal region. This was mainly due to significant changes occurring in the intermediate arousal session 3. Negative elicitation sessions showed arousal-dependent differences, occurring exclusively on the coupling between EEG θ–μRR and EEG θ–HRV LF/HF ratio in the prefrontal cortex region. Similar EEG–HRV coupling maps, calculated for the EEG α, β and γ bands, are reported in the electronic supplementary material, figures S10–S12. Figure 3. p-Value topographic maps resulting from the statistical comparison (Friedman non-parametric test) of MIC values between all the positive elicitation sessions (a), and all the negative elicitation sessions (b). Results are shown for each HRV feature and the EEG θ frequency band, considering each arousal level. Blue regions are associated with no significant difference between sessions,whereas green/yellow/red activations are associated with significant differences of the brain–heart coupling in at least one of the considered sessions.
The electronic supplementary material, figures S4–S5, shows p-value topographic maps of MIC values between positive/negative elicitation sessions, across each arousal level, whereas electronic supplementary material, figures S6–S9, show actual MIC values calculated on each pair of time-varying HRV features and time-varying EEG-PSD, calculated in all of the EEG frequency bands, for positive and negative elicitation sessions, among all of the arousal levels Figure 4 shows, for each arousal level, p-value topographic maps resulting from the statistical comparison (Wilcoxon non-parametric test) of MIC values between positive and negative elicitation. Results show that, at arousal levels 1, 2 and 3, positive emotional pictures increased the brain–heart coupling with respect to the negative ones, as estimated through MIC values between EEG θ power and time-domain features μRR, σRR and σHR, and the parasympathetic component (HF) of HRV. At arousal level 4, brain–heart dynamics switches to opposite MIC trends, i.e. negative emotional pictures increased the brain–heart coupling with respect to the positive ones, as estimated through MIC values between EEG θ power and HRV time-domain features. Moreover, at arousal levels 1, 2 and 3, negative emotional pictures significantly increased the brain–heart coupling with respect to the positive ones as estimated through MIC values between EEG θ power and HRV frequency-domain features. It is noteworthy that the strong effect of negative pictures at arousal level 3 on of MIC values calculated between EEG θ power and HRV-LF. Interestingly, at arousal level 4, such a EEG θ power and HRV–LF coupling seems to switch to EEG θ power and frequency-domain feature HF, though with less spatial activation. Similar EEG–HRV coupling maps, calculated for the EEG α, β and γ bands, are reported in electronic supplementary material, figures S13–S14. Figure 4. p-Value topographic maps resulting from the statistical comparison (Wilcoxon non-parametric test) of MIC values between positive and negative elicitation. Results are shown for each HRV feature and the EEG θ frequency band, considering each arousal level. Green regions are associated with no significant difference between positiveand negative elicitations, whereas red/blue activations are associated with significant increase of the brain–heart coupling during the positive/negative elicitation sessions.
We studied brain–heart dynamics using MIC values calculated between time-varying series of EEG power spectra and instantaneous heartbeat dynamics during visual emotional elicitation. Twenty-two healthy volunteers were emotionally elicited through passive viewing of IAPS pictures, covering 25 different combinations of arousal and valence levels. The proposed methodological approach and experimental protocol are, to our knowledge, of great novelty in the current literature because (i) we considered linear and nonlinear brain–heart couplings, (ii) we considered all possible arousal-, valence-dependent brain–heart couplings, and (iii) such coupling measures were calculated with subject-specific, time-varying features. Furthermore, experimental results were performed using high-resolution EEG signals (128 channels). At a group level, results were shown as p-value topographic maps and related multiple comparisons, highlighting brain regions whose activity significantly correlated with heartbeat dynamics. Importantly, from a methodological point of view, the study of brain–heart dynamics performed on uneven heartbeat samples, taking into account short-time emotional elicitation (less than 10 s), was possible thanks to the use of the point-process paradigm [68]. Through this approach, we obtained instantaneous time domain and spectral cardiovascular estimates, which are known to track ANS changes due to emotional elicitation and mood changes [70,80]. Accordingly, we hypothesized that different emotional stimuli would differently affect ANS–CNS signalling. As a proof of concept of the proposed methodology, we focused on EEG oscillations in the θ band as emotional processes consistently elicit changes in the EEG θ power, regulating ANS responses accordingly [55–65]. As expected, we found that the prefrontal cortex plays a crucial role in brain–heart coupling modulation during visual emotional elicitation. This cortex was involved in the switching mechanisms between neutral and arousing elicitation, especially during negative elicitation sessions. In particular, a strong coupling between prefrontal cortex activity and heartbeat dynamics was found at intermediate arousal (arousal level 3). In such an arousing elicitation, we also found a significant EEG θ–HRV-LF coupling in the left temporal region, especially due to images with negative valence. Furthermore, MIC differences between negative and positive valence (figure 4), which are visible at the lowest level of arousal (i.e. arousal level 1), changed both in sign and in spatial location at higher levels of arousal (level 2 and 3) and disappear at the highest level of arousal (level 4). Note that this behaviour was consistently found across all EEG frequency bands (see figure 4 and electronic supplementary material, figures S4–S14). This suggests the following conclusion: (i) the assumption of having orthogonal dimensions in the CMA model, as associated with arousal and valence dimensions, has to be reconsidered. This is also in agreement with previous findings suggesting that the effect of emotional valence on affective picture perception is modulated by levels of arousal at both early and late stages of brain processing [81]; (ii) increasing arousing elicitations seems to mitigate, or even to washout, the impact of valence on brain–heart dynamics in response to visual elicitation. Positive emotions elicited greater MIC than negative ones for the total RR variability (σRR and σHR) and for the HRV parasympathetic component HF across arousal levels 1, 2 and 3. For σRR despite this behaviour (higher brain–heart coupling for positive emotions) is diffuse over the scalp, the topology of significance involves central regions of the right hemisphere. Concerning HRV–HF, MIC significance shifts from left (arousal 2) to right (arousal 3) hemisphere. Negative emotions elicited a significantly higher brain–heart reaction with respect to the positive ones, as indicated by a higher MIC over large cortical areas for HRV–LF power. This was especially evident between EEG oscillations in θ band, at arousal 3, in the areas related to visuospatial attention. Interestingly, such a strong interaction vanishes at the highest arousal level (arousal 4), being replaced by a stronger (parasympathetic) HF activity associated with EEG θ power increase in left frontal regions and in right parietal regions. Importantly, this patterns of activations seem in line with a classical attentional-bradicardic reaction to hyper-aroused images with negative emotional content [82]. We believe that the vanishing effect hereby observed at high arousing elicitation has to be related to the processing of emotions at a CNS level exclusively. At a ANS level, in fact, through the same experimental protocol, our previous findings confirm arousal-specific patterns of skin conductance [83,84], and HRV and respiratory dynamics [85], allowing for a four-class discrimination of all of the arousing visual elicitations. Statistically significant differences between sessions of the same type should be linked to the quantification of the effect of (pleasant/unpleasant) arousing stimuli to subsequent neutral elicitations and resting-state sessions. Interestingly, while we found no variations in brain–heart dynamics between resting-state sessions, we found significant differences between neutral sessions. In particular, major differences between neutral sessions occurred in the prefrontal cortex, mainly due to significant changes occurring in the neutral session 5. Note that, in the experimental protocol timeline, arousal level 3 (i.e. the elicitation with higher changes in brain–heart dynamics) follows neutral session 5. In this view, an early/late effect can be associated to an arousing visual elicitation. We also remark that brain–heart coupling, as estimated through our MIC-based method, is significantly lower during non-elicitation states than the ones with elicitations. This can be easily seen from electronic supplementary material, tables S1–S4, showing MIC values calculated during the first 5 min baseline acquisition of the experimental protocol. It is worthwhile noting that stronger associations of parasympathetic activity with frontal and parietal cortices could be related to the attentional noradrenergic network, which was previously identified in fMRI studies [86]. This network extends to both hemispheres, whose stronger connectivity is derived from the right posterior parietal cortices [86] as indicated in our results. More in general, at a speculation level, we believe that part of our results may be explained by both noradrenergic and dopaminergic signalling modulating the brain–heart coupling, along with their interaction. Indeed, for the dopaminergic system, the nucleus accumbens integrates affective inputs from the amygdala and the prefrontal cortex, and its role in salience processing, signals novelty and contextual deviance has been previously identified [87]. Moreover, prefrontal feedbacks to limbic structures are able to modulate subcortical excitability and, thus, change the brain–heart coupling in different regions. On the other hand, the noradrenergic modulation on prefrontal feedbacks to the limbic output could explain the observed different levels of brain–heart coupling as a function of arousal [88]. Our findings are consistent with Scherer's theory, which argues that synchronization of oscillatory physiological systems is fundamental to emotion [89], and with Porges's polivagal theory [90], which points out how afferent feedback from the heart to the brain through the vagus nerve and nucleus tractus solitarius could play a regulatory role in emotional response. As this study should be intended is a proof-of-concept of a novel methodology to investigate brain–heart coupling, limitations should also be mentioned. First, although we did check that subjects were actually looking at pictures throughout the experimental protocol, we could not control for attention and habituation effects since the task was a complete passive one. However, although we had a repetition of five neutral sessions throughout the elicitation, we did include different images in such sessions, minimizing the repetition effects on emotional processing. Note that our previous endeavours demonstrate that no significant changes occur in ANS dynamics among neutral IAPS elicitation sessions [91,72]. Moreover, self-assessment scores of elicited IAPS images after the experiment were not taken into account in this study. We relied on the standardization of the IAPS images, which had been performed on a large number of healthy subjects [67], ensuring highly consistent results in terms of valence and arousal ratings. However, we cannot exclude that individual differences in valence and arousal perceptions, with respect to the elicited ones, occurred. Furthermore, we are aware that brain–heart dynamics might be further mediated by psychological factors like mood, anxiety or personality traits that we did not take into account in this study. Future works should account for all these aspects in order to achieve a more detailed and precise brain–heart model, also accounting for the evaluation of brain–heart coupling directionality, as well as for the brain–brain, and ANS–ANS interactions [65]. MIC, in fact, is unable to assess how much the past samples in one series affect its future values or the future of the other time series (like in Granger causality measures). However, is important to point out that these issues has a limited impact on our results. We were interested in assessing the coupling between EEG oscillations and heartbeat dynamics, tracking their changes despite external elicitation. In other words, despite the individual emotional processing of each subject, we were able to identify significant changes between θ oscillations and HRV metrics. The effect of slight difference between the window duration for the estimation of EEG (4 s) and HRV parameters (about 7 s due to the point-process model order p=7) on experimental results should also be investigated. Finally, although in this study we performed a high-resolution EEG recording with 128 channels, our conclusion involving the prefrontal cortex activity have limited physiological interpretation due to limits in the spatial resolution. In future studies, precise electrode localization, paired with individual morphological magnetic resonance imaging, would be acquired in order to have a reliable increase of spatial resolution and thus to perform analysis at a cortical surface levels such as Loreta-based approaches. To conclude, we demonstrated that a point-to-point linear and nonlinear correlation measure between instantaneous heartbeat dynamics and EEG time-varying spectra may be a feasible method to understand the coupling between ANS and CNS. We suggest that EEG oscillations in the θ band are the most promising metrics to be used in such an evaluation, especially because this band have a significant role in monitoring the attentional significance of emotions [92]. However, it is possible to hypothesize that different EEG bands couple with different aspects of ANS dynamics. As a matter of fact, in the electronic supplementary material, we were able to show distinct pattern of coupling between different EEG bands and heartbeat dynamics, although they were less supported by psychological and neurobiological bases. To this extent, further studies should effectively focus on other specific bands than θ, also including a larger number of subjects in order to take into account other covariates such as personality traits and subjective IAPS rating, which may influence central and peripheral signalling. Moreover, quantification of the exact amount of nonlinear coupling occurring in brain–heart dynamics should also be performed. Further studies should also address whether any early/late phenomena as a response of affective stimuli can be identified in brain–heart dynamics. Moreover, it is important to investigate this coupling in psychiatric as well as neurological disorders which have been shown to involve both cognition and ANS dysfunction [93], and whether its alterations could be a specific marker of mental disorders. According to our findings, we believe that the analysis of EEG θ rhythms will provide a significant decision support tool for managing mental health, also considering the role of insular and pregenual cingulate cortices in psychiatric disorders including mood disorders, panic disorders, obsessive-compulsive disorders, eating disorders and schizophrenia [1,5,46,47,90,94]. The experimental protocol for this study was approved by the ethics committee of the University of Pisa and informed consent was obtained from all participants involved in the experiment. We declare we have no competing interests. The research leading to these results has received funding from the European Commission Horizon 2020 Programme under grant agreement no. 689691 “NEVERMIND”. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 8Schizophrenia is considered to be one of the most severe mental disorders in the world. It is associated with higher cardiac mortality rates, an approximately 15–20 year shorter life expectancy, and up to triple the risk of attaining cardiovascular disease (CVD) compared with the general population, independent of age groups [1–3]. A particular cause for concern is that the mortality gap between the general population and schizophrenia patients seems to have increased during recent decades [4]. Suicide and accidents account for only a part of excess mortality, whereas a substantial proportion is due to physical illness [5]. The largest single cause of death in schizophrenia patients leading to an increased mortality rate is due to CVD, with CVD mortality ranging from 40% to 50% [6]. Causal factors for patients with this condition are still being discussed, and have not been fully clarified until now. However, possible complicating factors are related to lifestyle, the lack of physical activity, smoking, obesity, poor diet, substance abuse, diabetes, hypertension, the cardiac side effects of antipsychotics and the imbalanced autonomic nervous system (ANS) during acute psychosis [1,6,7]. Two important differences from other patient populations suffering from primary cardiac conditions (e.g. myocardial infarction, cardiomyopathy) and which present signs of cardiac autonomic dysfunction (CAD) need to be taken into account. The first difference is the fact that severe CAD is not initially caused by major structural or functional alterations of the heart in schizophrenia patients. Moreover, it seems to be associated with an altered brain–heart interaction influenced by a lack of cortical inhibitory control over sympathoexcitatory subcortical regions [8,9]. The second difference from patient populations suffering from primary cardiac conditions is caused by the relative ‘longevity’ of patients with schizophrenia when compared with more frequent shorter survival rates of cardiac patients. Previous studies have indicated that people with lower heart rate variability (HRV), as seen in the case of schizophrenia, exhibit effective behavioural responses (e.g. faster response times and higher accuracy rates) on executive cognitive tasks as well as exhibited flexible and adaptive emotional [10,11]. Thayer & Lane [12] proposed the neurovisceral integration model, which suggests that neural networks implicated in emotional and cognitive self-regulation are also involved in the control of cardiac autonomic activity. Frontal, cingulate and subcortical brain regions have been hypothesized to play a critical role in such self-regulatory functions through top-down control from the frontal cortex over subcortical regions involved in reward and emotion, such as the amygdala [13]. A recent meta-analysis [14] revealed that resting HRV is tied to the functioning of frontal–subcortical circuits. Higher resting HRV is associated with the effective functioning of frontal-top-down control over subcortical brain regions that support flexible and adaptive responses to environmental demands [12]. It is noteworthy that a disruption of frontal–subcortical circuits has been associated with a wide range of psychopathologies, including schizophrenia [15]. Cognitive impairment is thus known to be a universal and core symptom of schizophrenia. This impairment critically influences treatment response, a patient's insight into the illness and the patient's employment status, ability to communicate, social relationships and living status [16]. Cardiovascular adjustments owing to a shift in central–autonomic control and remodelling of the heart are most prominent features of exercising [17,18]. It has been suggested that a reduced sympathetic modulation and an increased parasympathetic dominance may be caused by adaptations of peripheral and central regulatory systems [19]. Cardiovascular centres in the brainstem work through various cardiovascular reflex mechanisms such as the baroreflex, the chemoreflex and the cardiopulmonary reflex [20]. The efferent sympathetic reflex component is determined by neurons in the caudal and rostral ventrolateral medulla. These neurons contribute to the maintenance of blood pressure and heart rate by signalling to the intermediolateral column of the spinal cord. Two further medullary areas contain preganglionic parasympathetic neurons: these are the nucleus ambiguous and the dorsal motor nucleus of the vagus nerve. Both mediate the efferent parasympathetic components of the above-described reflexes [21,22]. These preganglionic parasympathetic neurons might show either cardiac- or respiration-related activity. Brain stem nuclei and pathways receive modulatory inputs from supramedullary centres such as the insula, thalamus, hypothalamus, amygdala, parietal and cingulate regions, or from the medial prefrontal region [23]. Studies have shown the involvement of these brain areas in the autonomic regulation at rest and during cognitive or emotional strains by means of functional brain imaging [24,25]. In the recent meta-analysis performed by Beissner et al. [26], it was found that largely divergent brain networks were associated with sympathetic and parasympathetic activity. It was revealed that autonomic regulation involves mainly the ventromedial prefrontal cortex, the perigenual anterior, the dorsal anterior cingulate, the posterior cingulate cortex and the insular cortices, in addition to the amygdala. It has been assumed that various autonomic function processes are generated by a network of interactions showing specificity for task and autonomic division. Psychopathological states such as anxiety, depression, post-traumatic stress disorder and schizophrenia are associated with prefrontal hypoactivity, and a lack of inhibitory neural processes reflected by a poor habituation to novel neutral stimuli, a pre-attentive bias for threat information, deficits in working memory and executive function, and poor affective information processing and regulation [27]. For healthy adults, Beissner et al. suggested that asymmetric frontal electroencephalogram (EEG) responses to emotional arousal in the form of positive and negative emotions may elicit different patterns of cardiovascular reactivity [28]. In summary, different studies using both pharmacological and neuroimaging approaches provided evidence that activity of the prefrontal cortex is associated with vagally mediated HRV [29]. Owing to this assumption that cardiovascular regulation mainly involves the prefrontal lobe, our investigation will focus on differences within this brain region with respect to heart rate and blood pressure. Thus, the aim of this study was to investigate the short-term instantaneous central–autonomic coupling (CAC) by determining the strength and direction, as well as the underlying structural patterns of this particular coupling in patients suffering from paranoid schizophrenia when compared with healthy subjects. To this end, we applied the coupling approaches high-resolution joint symbolic dynamics(HRJSD) and the normalized short-time partial-directed coherence (NSTPDC). In particular, we strived to determine whether significantly different correlations existed between changes in EEG activity at the frontal lobe and changes in heart rate, as well as in these subjects' systolic blood pressure. In this regard, it must be considered that the interactions between the central nervous system (CNS) and ANS can be assumed to be a feedback–feed-forward system that supports flexible and adaptive responses to environmental demands. This study may improve the understanding of pathophysiological processes of the central–autonomic network found in paranoid schizophrenia patients. In this study, 17 patients with paranoid schizophrenia (SZ; two females, 37.5±10.4 years) and 17 healthy subjects (CO; four females, 37.7±13.1 years), who were matched according to age and gender, were enrolled. The diagnosis of paranoid schizophrenia was established when patients fulfilled DSM-IV criteria (Diagnostic and Statistical Manual of Mental Disorders, fourth edition). Psychotic symptoms were quantified using the positive and negative syndrome scale. Patients had been treated with depot antipsychotic medication (77% being atypical neuroleptics and 23% being a mixture of antidepressant and atypical neuroleptics). Thoroughly performed interview and clinical investigations were performed for CO to exclude any potential psychiatric or other diseases, as well as to double-check for any interfering medication. The structured clinical interview and a personality inventory (Freiburger Persönlichkeitsinventar) were also applied to CO to detect personality traits and any disorders that might influence autonomic function. Furthermore, all subjects were asked to relax and to breathe normally to avoid hyperventilation. No further breathing instructions were given. Subjects were explicitly asked not to talk during the recording. All participants (SZ and CO) provided their written informed consent to a protocol approved by the local ethics committee of the Jena University Hospital. This study complies with the Declaration of Helsinki. From all schizophrenic patients and healthy subjects, a three-channel short-term electrocardiogram (500 Hz), non-invasive continuous blood pressure (200 Hz) and a 64-channel EEG were recorded for 15 min under resting conditions. Participants remained seated and their eyes were closed during this recording. The EEG was acquired using 64 active Ag/AgCl electrodes, and transmitted using a BrainAmp Amplifier (Brain Products, Germany, sampling rate 500 Hz, AFZ: ground, FCZ: reference). The electrodes were positioned according to the extended 10–20 system using an electrode cap. The impedance levels (less than 25 kΩ) for all electrodes were checked following attachment of the electrode cap to each participant's scalp. The arterial blood pressure was recorded using the volume-clamp photoplethysmographical blood pressure device Portapres model-2 (TNO Biomedical Instrumentation, The Netherlands). Investigations were performed between 14.00 and 18.00 in a quiet room that was kept comfortably warm (22–24°C) and began after subjects had rested in a supine position for 10 min. Subjects were asked to close their eyes, relax and breathe normally to avoid hyperventilation. The following time series with respect to autonomous regulation was automatically extracted from the raw data records (figure 1) using in-house software (programming environment Delphi v. 3 and Matlab
Figure 1. A visualization example of analysed raw data records and their extracted time series. Raw data are, from top to bottom: electrocardiogram, non-invasive continuous blood pressure, electroencephalogram (EEG) and EEG spectral band component (delta). RR(i) represents the beat-to-beat intervals; SYS(i) represents the maximum systolic blood pressure amplitude values over time in relation to the previous R-peak; EEG(i) and EEGδ(i) specified to time intervals of the EEG raw data (electrode: F4); and the EEG delta band time intervals (electrode: F4) in relation to RR(i).
Figure 2. Examples of extracted time series. Time series from top to bottom are: tachograms (BBI); systograms (SYS); the mean power PEEG(i) of the BBI-related EEG intervals EEG(i); and the mean power PEEGδ(i) of the BBI-related EEG-delta band intervals EEGδ(i) from a control subject (a), and a patient with paranoid schizophrenia (b). Note the typical lower variability in BBI sequences and the lower mean power PEEG and PEEGδ in the patient suffering from paranoid schizophrenia. The plot scaling of the PEEGδ(i) time series was different to guarantee for a higher degree of visual clarity.
EEG recordings were band-pass filtered (0.05–60 Hz, Butterworth filter, order=3) in order to remove slow drifts resulting from slow body movements or sweating, and to prevent higher-frequency content from additional components. For EEG analyses, artefact-free time series were used, being determined by visual inspection and via automatic classification using the Brain Products software Analyzer v. 2.0. Based on the EEG raw data recordings, new time series consisting of the EEG spectral band components as delta (0.5–3.5 Hz), theta (3.5–7.5 Hz), alpha (7.5–12.5 Hz), alpha1 (7.5–9.5 Hz), alpha2 (9.5–12.5 Hz), beta (12.5–25 Hz), beta1 (12.5–17.5 Hz), beta2 (17.5–25 Hz) and gamma (greater than 25–60 Hz) activity were derived (Butterworth filter, order=3) for each electrode (figure 1). These recordings (EEG raw data and the nine associated EEG spectral bands) were used for further analyses. For example, for the EEG channel Fp1, we obtained the 10 time series EEGFp1, EEGFp1δ, EEGFp1θ, EEGFp1α, EEGFp1α1, EEGFp1α2, EEGFp1β, EEGFp1β1, EEGFp1β2 and PFp1γ. In relation to the extracted RR-intervals (figure 1) from the EEG raw data and the EEG spectral band recordings, the corresponding time intervals were extracted as EEG(i) (ms) and EEGband(i) (ms) (figure 1). Within each RR-interval RR(i), with i (i=1: R−1) as the successive number of R-peaks (R), the mean power PEEG(i) of EEG(i) and the mean power PEEGband(i) of EEGband(i) were calculated to obtain new time series for signal analyses (figure 2): 2.1 where T represents the number of samples within RR(i) or EEG(i), respectively, t(i) represents the current point in time of RR(i) and fs represents the sampling frequency. New types of time series were thus derived as PEEG, PEEGδ, PEEGθ, PEEGα, PEEGα1, PEEGα2, PEEGβ, PEEGβ1, PEEGβ2 and PEEGγ (μV2). All extracted time series (autonomous, central) were filtered by applying an adaptive variance estimation algorithm [30] to remove and interpolate seldom occurring ventricular premature beats and artefacts (e.g. movement, electrode noise and extraordinary peaks) to obtain normal-to-normal beat time series (NN). To obtain synchronized time series, BBI, SYS, PEEG and PEEGband were resampled using a linear interpolation method (2 Hz). For all EEG analyses and CAC analyses, three brain regions with the corresponding EEG channels were analysed, namely
For the conducting of CAC analyses, the autonomic time series (BBI and SYS) from each subject were coupled/combined with all possible EEG channels, and subsequently averaged. For example, when coupling heart rate (BBI) with the raw EEG (right frontal area), we obtained 12 different coupling combinations of BBI-PEEG(Fp2),BBI-PEEG(AF4),BBI-PEEG(AF8),BBI-PEEG(F2),…,BBI-PEEG(FT10). These 12 combinations were analysed using different coupling approaches (see NSTPDC, HRJSD). Furthermore, for each combination, the methods-related indices were derived and averaged for each subject. For the quantification of the EEG, the power within the delta (Pδ: 0.5–3.5 Hz), theta (Pθ: 3.5–7.5 Hz), alpha (Pα: 7.5–12.5 Hz), alpha1 (Pα1: 7.5–9.5 Hz), alpha2 (Pα2: 9.5–12.5 Hz), beta (Pβ: 12.5–25 Hz), beta1 (Pβ1: 12.5–17.5 Hz), beta2 (Pβ2: 17.5–25 Hz) and gamma (Pγ: 25–60 Hz) bands and the whole power P within the entire frequency band (0.5–60 Hz) were calculated from the EEG raw data recordings applying the Welch method [31] to estimate the power spectral density function (window length = 5 s, overlap = 50%). HRV and blood pressure variability (BPV) were quantified by calculating standard parameters from the time domain [32,33] as
In this study, we used the dual sequence method [34] to estimate the spontaneous baroreflex sensitivity (BRS) which is based on the sequence technique. To this end, both a minimum change of a 1 mmHg increase/decrease in SYS, and 5 ms in BBI were defined as inclusion criteria for a spontaneous baroreflex-related cardiovascular oscillation. The slopes of the regression lines between SYS and BBI sequences were taken as an index for local BRS (ms mmHg−1). We analysed two kinds of BBI responses: (i) bradycardic (an increase in SYS being associated with an increase in BBI) and (ii) tachycardic fluctuations (a decrease in SYS being associated with a decrease in BBI). As a result, two indices were derived:
The tool of HRJSD was originally introduced [35,36] to quantify the effects of antipsychotics and their anticholinergic effects on nonlinear cardiovascular couplings in acute schizophrenia via the use of symbols. The idea of HRJSD is to classify frequent deterministic patterns lasting three beats (represented by symbols). The HRJSD approach enables the classification and characterization of short-term regulatory bivariate coupling patterns that are dominating the interaction generated by the ANS. In this study, we applied HRJSD as a promising tool to analyse CAC. HRJSD was applied to determine short-term bivariate coupling patterns between EEG and ANS-related (heart rate, systolic blood pressure) time series. HRJSD based on symbolization permits a coarse-grain quantitative assessment of short-term dynamics of time series/biosignals. Therefore, the direct analysis of successive signal amplitudes is based on discrete states (represented by symbols). In short, HRJSD works by transforming the two investigated time series (BBI and or SYS and PEEG) into symbol sequences based on their signal amplitudes using a given alphabet A={0,1,2}. The bivariate sample vector X of the two time series (e.g. BBI and PEEG) with xBBI and xPEEG is transformed into a bivariate symbol vector S, where n are the nth beat-to-beat values of BBI and PEEG, respectively. 2.2 and2.3 The bivariate symbol vector S is defined using the following definitions:2.4 and2.5 Increasing values were coded as ‘2’, decreasing values were coded as ‘0’ and unchanging (with no or little variability) values were coded as ‘1’. Afterwards, S was subdivided into short words (sequences of symbols) wk of length k=3. In this study, an adapted threshold l to the individual physiological dynamic variability lBBI and lSYS and lPEEG equal to 25% of the s.d. of the BBI and PEEG time series was applied. The derived different word types from the BBI (wBBI) and PEEG (wPEEG) time series (word types ranging from: 000,001,…,221,222) were sorted into a normalized 27×27 vector matrix Wn ranging from word type (000,000)T to (222,222)T. These single-word types wBBI,wPEEG (where the total number of all word type combinations is 27×27=729) were subsequently grouped into eight pattern families wf, whereby the sum of probabilities of all single-word family occurrences p(wf) was normalized to 1. The eight pattern families (E0, E1, E2, LU1, LD1, LA1, P, V) describe different aspects of autonomic modulation (strong and weak increase or decrease, no variability or alternations) of the BBI and PEEG time series. These families were then sorted into an 8×8 pattern family density matrix Wf, resulting in 64 CAC patterns. These pattern families were defined as follows (figure 3):
Figure 3. Definition of eight HRJSD pattern families (BBI, beat-to-beat intervals; HR, heart rate; SYS, systolic blood pressure; PEEG, mean power of BBI-related EEG intervals).
Using the word distribution density matrix Wf, the normalized joint probability of each word occurrence was calculated and analysed. Furthermore, from the matrix Wf, the sum of each (n=8) column cfPEEG (cfE0, cfE1, cfE2, cfLU1, cfLD1, cfLA1, cfP, cfV) and the sum of each (n=8) row rfBBI (rfE0, rfE1, rfE2, rfLU1, rfLD1, rfLA1, rfP, rfV) were calculated. This corresponds to the univariate analysis of the symbolic dynamics of the first and second signals used for the HRJSD matrix construction. To quantify the complexity of the CAC, the Shannon entropy (HRJSDShannon) within Wfwas estimated using the following equation: 2.6 In this study, the causal CAC between BBI and SYS with PEEG were investigated with respect to a bivariate dynamic system.To quantify the causal coupling direction and strength between the central and autonomic time series, the NSTPDC [37,38] approach was applied, based on an m-dimensional MAR process with model order p to determine Granger causality in the frequency domain. To this end, the NSTPDC is based on the time-variant partial-directed coherence approach (tvPDC, πxy( f,n)), providing information about the partial correlative short-time interaction properties of non-stationary signals, with f as the frequency and n the number of windows [39]. To quantify the coupling direction between two time series x and y (e.g. BBI and PEEG: with xBBI and yPEEG) a coupling factor (CF) was introduced. CF was obtained by dividing the mean value πxBBIyPEEG( f,n) by the mean value of πyPEEGxBBI( f,n): 2.7 These results were normalized to become a specific set of values leading to the normalized factor (NF). max (2.8 NF allows for the determination of the direction of the causal connections between the investigated time series (xBBI and yPEEG) as a function of frequency f. NF takes the following values: NF={−2,−1,0,1,2}. Strong unidirectional coupling is indicated if NF is −2 or 2 (where −2 denotes yPEEG as the driver), bidirectional coupling if NF=−1 or 1 (−1 denotes yPEEG as the driver) and an equal influence in both directions and/or no coupling if NF=0.For determining the coupling strength between two time series xBBI and yPEEG the areas (ABBI→PEEG, APEEG→BBI (arb. units)) generated in space by CF were estimated for each window using a trapezoidal numerical integration function for approximation. ABBI→PEEG and APEEG→BBI range between 0 and 1 [0,1]. Hereby, 1 indicates that all causal influences originating from one time series x are directed towards (arrows: →) the second one y (Ax→y=1). In this study, NSTPDC indices were calculated by applying a window (Hamming) of lengths l, where l=80 samples, and shifting the window by 20 samples per each iteration step (60 samples overlap between each window). In order to take advantage of the aspect of stationarity and scale-invariance for NSTPDC analyses, a normalization (zero mean and unit variance) of the time series BBI, SYS and PEEG/PEEGband was performed [37]. Therefore, each sample i of the BBI and PEEG time series x={xi,i=1,…N} and y={yi,i=1,…N} with N as the maximal number of samples i (temporal index) was first normalized by subtracting the mean of 2.9 In this study, the causal CAC between BBI, SYS and PEEG, as well as BBI, SYS and PEEGband were investigated with respect to a multivariate dynamic system.The non-parametric exact two-tailed Mann–Whitney U-test (SPSS 21.0) was performed to evaluate differences in central and autonomic standard indices, as well as differences in CAC indices between SZ and CO persons. The significance level was set to p<0.01 (Bonferroni–Holm adjustment: p<0.00045, n=100 indices). In order to facilitate comparisons with other findings within the field where published papers dealt almost exclusively with mean values, it was decided that this paper would present all results as mean values and s.d. All indices in the tables (results section) were also presented in medians and interquartile ranges (see also electronic supplementary material). Considering EEG-related standard frequency domain spectral component indices, we found that when comparing SZ with CO persons in terms of the frontal area, the left frontal area and the right frontal area, highly significant (p<0.00045) differences between both groups were apparent (table 1). Thereby, a significant decrease in the mean power of all spectral bands (delta Pδ to gamma Pγ) and in the whole power P was obviously present for SZ in comparison with CO. In both groups, a lower power was shown in the left frontal area when compared with the right frontal area.
HRV analysis revealed highly significant differences in meanNN (p<0.00045) and rmssd (p<0.01) between both groups (table 2). Thereby, SZ patients were characterized by having reduced mean basic beat-to-beat intervals (meanNN) and HRV (sdNN, rmssd) in comparison with CO subjects. Analyses of basic systolic and diastolic blood pressure indices (BPV_sys, BPV_dia) revealed, however, no significant differences between SZ and CO (table 2).
BRS measures bslope and tslope revealed significant differences (p<0.01) between SZ and CO, namely a significant decrease in the BRS measures bslope and tslope was shown (table 2). HRJSD analyses revealed highly significant (p<0.00045) differences between SZ and CO in all eight CAC pattern families (PEEG/BBI, 8×8=64) for the entire frontal area, the left frontal area and the right frontal area (table 3). The patterns were characterized by decreased absolute values in SZ if the central pattern family PEEG-E0, PEEG-E1, PEEG-E2, PEEG-LU1, PEEG-LD1, PEEG-LA1, PEEG-P and PEEG-V was coupled with BBI-E0, BBI-E1, BBI-E2, BBI-LU1 and BBI-LD1. SZ values significantly increased if the central pattern family was coupled with BBI-LA1, BBI-P and BBI-V. Thereby, the central family patterns PEEG-E0 and PEEG-E2 significantly decreased, and PEEG-LA1 significantly increased in SZ, when compared with CO. The cardiac family patterns BBI-E0 and BBI-E2 highly significantly decreased, and BBI-LA1, BBI-P and BBI-V highly significantly increased in SZ, when compared with CO (table 4). The index HRJSDShannon did not reveal any significant differences between SZ and CO, regardless of the investigated frontal area (see also electronic supplementary material, table S1).
In contrast to the central–cardiac results, a higher number of significantly different central–vascular coupling pattern families (PEEG/SYS, 8×8=64) for the entire frontal area, left frontal area and right frontal area were found (table 3). In SZ patients, the absolute values of central–vascular coupling pattern families significantly decreased if the central pattern family PEEG-E0, PEEG-E1, PEEG-E2, PEEG-LU1, PEEG-LD1, PEEG-LA1, PEEG-P and PEEG-V were coupled with SYS-E1. They significantly increased if the central pattern family PEEG-E0, PEEG-E1, PEEG-E2, PEEG-LU1, PEEG-LD1, PEEG-LA1, PEEG-P and PEEG-V were coupled with SYS-E0, SYS-E2, SYS-LU1, SYS-LD1, SYS-LA1, SYS-P and SYS-V (table 3). In addition, the absolute values of central family patterns PEEG-E0 and PEEG-E2 significantly decreased, and PEEG-LA1 significantly increased in SZ when compared with CO. The values of vascular family patterns SYS-E0, SYS-E2, SYS-LU1, SYS-LD1, SYS-LA1, SYS-P and SYS-V highly significantly (p<0.00045) increased in SZ when compared with CO (table 4). HRJSDShannon values significantly increased in SZ patients when compared with CO for the entire frontal area (CO: 2.7±0.9, SZ: 3.0±1.2, p<0.00045), the left frontal area (CO: 2.7±0.9, SZ: 3.0±1.2, p<0.00045) and the right frontal area (CO: 2.6±0.9, SZ: 3.0±1.2, p<0.01) (see also the electronic supplementary material, table S1). NSTPDC results revealed a significant NF between SZ (NF: −0.48±0.81) and CO (NF: −0.66±0.52). With regard to SZ, the mean NF was nearly −0.5, indicating a reduced bidirectional coupling from SYS→BBI, the driver being SYS and BBI the target variable. Additionally, ASYS→BBI as marker for the coupling strength was highly significantly different between SZ (0.39±0.16) and CO (0.43±0.14) (p<0.00045). In this case, SYS influenced BBI (baroreflex loop) to the extent that a decrease in coupling strength was found. This indicates a weaker causal influence of SYS on BBI (table 5).
When considering the coupling of BBI with PEEG, only the NF revealed a significant (p<0.01) difference between SZ and CO in the entire frontal area. Thereby, mean NF was nearly −0.8, pointing to bidirectional coupling from PEEG→BBI, with the driver being PEEG, and BBI the target variable (table 5). When coupling the vascular system (SYS) with the central activity (PEEG), it becomes apparent that the NF was highly significantly (p<0.00045) different between SZ and CO for the entire frontal area, the left frontal area and the right frontal area. Here, CO showed mean NF values of nearly 0, indicating an equal coupling influence in both directions. On the contrary, SZ showed mean NF values of −0.6 to −0.8, pointing to bidirectional coupling from PEEG→SYS (with driver PEEG). These results were supported by ASYS→PEEG and APEEG→SYS for CO, revealing nearly the same values for both directions. Both indices ASYS→PEEG and APEEG→SYS were highly significantly (p<0.00045) different between SZ and CO for the whole frontal area, the left frontal area and the right frontal area. When SYS influences PEEG (SYS→PEEG), the coupling strength was significantly reduced in SZ in comparison with CO. However, when PEEG influences SYS (PEEG→SYS), the coupling strength was significantly increased in SZ when compared with CO (table 5 and figure 4). Figure 4. Averaged NSTPDC plots for central–autonomic coupling analyses for (a) healthy subjects and (b) schizophrenic patients. Arrows indicate the causal coupling direction from one time series to another, e.g.
For all spectral bands, NSTPDC results revealed significantly (p<0.00045) decreased coupling strengths from the EEG spectral power bands towards BBI (APEEGband→BBI) for the whole frontal area, the left frontal area and the right frontal area in SZ when compared with CO. Thereby, in PEEGγ, the strongest influence of central γ activity towards BBI was found for both SZ and CO. With regard to the coupling direction of BBI towards PEEGband, we found in the coupling strengths for the whole frontal area only in PEEGβ, PEEGβ1, PEEGδ and PEEGθ significant increases in ABBI→PEEGband for SZ in comparison with CO. Regarding the left frontal area, only in PEEGβ1 and PEEGθ were significant increases in the coupling strengths (ABBI→PEEGband) found for SZ, when compared with CO. Regarding the right frontal area, in PEEGα1, PEEGβ1, PEEGδ and PEEGθ significant increases in the coupling strengths (ABBI→PEEGband) could be shown for SZ (figure 5). Figure 5. Visualization of significant differences between patients suffering from paranoid schizophrenia (SZ) and healthy subjects (CO) with respect to the coupling strength (NSTPDC) between autonomic activity (BBI, SYS) and central spectral activity (PEEGband) for (a) the whole frontal area, (b) the left frontal area and (c) the right frontal area. Arrows indicate the coupling direction, where black solid lines indicate the direction from central spectral activity towards autonomic target variables. Grey dashed lines indicate the direction from the autonomic variables towards central spectral activity. Note that all arrows were highly significantly (p<0.00045) different between SZ and CO; otherwise, the arrows were indicated by *(p<0.01). BBI, beat-to-beat intervals; SYS, maximum systolic blood pressure amplitude values over time; PEEGband, the mean power in BBI-related EEG spectral band intervals.
Considering the NF values for all couplings between BBI and PEEGband, SZ revealed generally increased NF values in comparison with CO. For the entire frontal area, left frontal area and right frontal area, the coupling directions were bidirectional, with BBI acting as the driver (electronic supplementary material, table S3). Only for PEEGδ and PEEGγ was equal influence present in both directions (see also electronic supplementary material, table S2). For the coupling of SYS with spectral power bands, we found for the direction of EEG spectral power bands towards SYS (APEEGband→SYS) an opposite behaviour, as shown by PEEGband→BBI (where significant coupling strengths were found here for all bands and all frontal areas; electronic supplementary material, table S3). Regarding the entire frontal area (PEEGα,PEEGα2) and the left frontal area (PEEGα,PEEGα1,PEEGα2) significant increases in APEEGband→SYS for SZ in comparison with CO were found. Again, in PEEGγ, the strongest influence of central γ activity towards SYS could be found for both SZ and CO. In the opposite direction, namely from SYS towards PEEGband, we found for the frontal area in all spectral bands (PEEGband) highly significant decreased coupling strengths in SZ when compared with CO. For the left frontal area besides PEEGγ and PEEGθ and for the right frontal area besides PEEGδ, PEEGγ and PEEGθ for all other bands showed significantly different coupling strengths between SZ and CO. With respect to NF, there was generally a bidirectionality found for EEG spectral power bands (PEEGδ, PEEGθ, PEEGβ, PEEGβ1, PEEGβ2), with SYS acting as the driver. Furthermore, there was a unidirectional coupling for PEEGα, PEEGα1 and PEEGα2 and an equal influence in both directions for PEEGγ for the entire frontal area, as well as for the left and right frontal areas. In the α bands and β2 band, NF was decreased in SZ when compared with CO (see also electronic supplementary material, table S2). Our study has revealed highly significantly increased heart rates, reduced HRV, decreased BRS, a reduced EEG activity (power) independent from frequency range and frontal area or left hemisphere area, as well as altered CAC in patients with schizophrenia, when compared with healthy subjects. In particular, CAC revealed that the influence of central activity towards SYS was more strongly pronounced than that towards BBI in SZ patients, when compared with CO subjects. In addition, CAC patterns in SZ were mainly characterized by a larger amount of increased short-term alternating and strongly decreasing central patterns (PEEG-LA1, PEEG-E0), as well as more widely distributed alternating and fluctuating heart rate patterns (LA1, P, V). SZ patients' CAC patterns were also characterized by a larger amount of increased short-term strong/weak increasing/decreasing, alternating and fluctuating SYS pattern families (E0, E2, LU1, LD1, LA1, P, V), as well as invariable SYS (E1), in comparison with CO subjects. Our findings in relation to HRV are in accordance with other studies that have revealed an altered autonomic tone in treated schizophrenic patients. Psychiatrists attributed the increased heart rates (meanNN↓) of schizophrenia patients to their antipsychotic treatment. This assumption is only somewhat correct, because treatment with clozapine, for instance, is associated with a reduced vagal function and increased heart rates [40]. It was shown that clozapine, quetiapine and amisulpride revealed stronger anticholinergic side effects of cholinergic or adrenergic receptors on ANS modulation than olanzapine [41]. In addition, Agelink et al. [42] investigated the effects of atypical antipsychotics on autonomic neurocardiac function, showing that only amisulpride did not significantly alter HRV owing to its lack of activity at cholinergic or adrenergic receptors. Bär et al. [43] found a reduction in the complexity of heart rate regulation after olanzapine treatment, measured by compression entropy. Thus, the authors suggested a decreased cardiac vagal function that could increase the risk for cardiac mortality. Although we found some indication of increased sympathetic modulation, their results seemed to be restricted to heart rate regulation and not to blood pressure [44]. Thus, a cardiac dysfunction in SZ does not reflect a simple stress-induced arousal, but rather chronic and distinct changes of heart rate and respiratory regulation [37]. When considering BRS, we found significantly reduced tachycardic (tslope) and bradycardic (bslope) slopes in accordance with other studies investigating unmedicated patients [35,45]. The decrease of efferent vagal activity and the inhibition of baroreflex vagal bradycardia in SZ might be caused by stress owing to psychotic experiences or to the psychosis itself, a process that allows the organism under physiological conditions to adjust to demanding environmental stress [40]. The results of central activity (via EEG frequency analyses) showed a highly significantly reduced EEG activation (power) in all frequency bands from the frontal lobe, being much more pronounced in the right frontal hemisphere when compared with CO. Recently, MacCrimmon et al. [46] investigated the effects of the atypical antipsychotic clozapine among 64 SZ patients. They found that clozapine augments power globally in the δ and θ bands, but this effect was more pronounced over frontal areas. The authors could demonstrate a significant clozapine-induced α topographic shift frontally and to the right. They suggested further investigations of subcortical structures in an attempt to better understand the diverse aetiologies and optimal treatments of the schizophrenias. Small et al. [47] investigated chronic treatment-resistant patients in relation to placebo, haloperidol, chlorpromazine and clozapine treatment. They found increased frontal δ activity particularly with clozapine and chlorpromazine treatment. Nagase et al. [48] investigated 12 medicated SZ patients, finding that α2 power and slow-wave power were reduced when compared with the neuroleptic-naive state. They concluded that the reduction in α power may occur from the early stage of the disease and progress even further, even though the patients are medicated and clinically improved. Kemali et al. [49] found that after acute treatment, patients showed a significant decrease of δ and an increase of θ2, β1 and β2. After 28 days of haloperidol treatment, similar changes were observed for δ, together with an increase of α1, and a decrease of fast β. Light et al. [50] found that schizophrenia patients have frequency-specific deficits in the generation and maintenance of coherent γ-range oscillations, reflecting a fundamental degradation of the basic integrated neural network activity. In general, γ responses in schizophrenic patients are not necessarily weakened. Depending on the status of the schizophrenic behaviour (negative or positive symptoms) and depending on the difficulty of the applied paradigm, an increase of γ activity may also be observed. Thus, the oscillatory dynamics in schizophrenia also depict the unstable behaviour of electrophysiology in this disease [51]. Patients who were treated with clozapine and olanzapine revealed most prominent changes in the anterior cingulate and medial frontal cortex and a decrease in fast frequency activities in the occipital cortex. These results suggest a compensatory mechanism in the neurobiological substrate for schizophrenia [52]. Unfortunately, at the moment, comparative studies between medicated and unmedicated patients are not available in the literature. This makes it difficult to assess the effectiveness of medication and the effect of central activity in schizophrenia patients. It was shown in many studies on medicated and non-medicated patients that the γ response is lower in SZ patients when compared with healthy subjects [51,53]. Nevertheless, it is strongly justified, based on available literature, to conclude that the δ excess (and to a lesser extent the θ excess) is a strong and bona fide biological marker for schizophrenia, as well as the fact that changes in EEG patterns are not medication-induced [54]. Cardiovascular coupling results based on NSTPDC analysis revealed a reduced bidirectional coupling and a reduced coupling strength from SYS→BBI in SZ, supporting the decreased BRS results. Due to the fact that SYS were not significantly changed in SZ, independent of medication usage when compared with CO, it can be assumed that the primary BRS changes in SZ were a result of impaired heart rate regulation. It was shown that during stressful conditions such as mental stress (supposedly present in SZ), the arterial baroreflex was generally inhibited. From the point of view that central mechanisms are involved in BRS regulation, central sites proven to elicit the facilitation are the medial prefrontal cortex, the preoptic/anterior hypothalamus, the ventrolateral part of the periaqueductal grey matter and the nucleus raphe magnus [55]. Central–vascular coupling analyses demonstrated that the coupling strength was highly significantly reduced in SZ for the direction SYS towards central activity (ASYS→PEEG). For the opposite direction from central activity towards SYS, the coupling strength (APEEG→SYS) was highly significantly increased in SZ when compared with CO. Central–vascular coupling in SZ pointed to a bidirectional one with the central driver (PEEG→SYS), whereas the direction for CO was equal in both directions ( Central–vascular coupling by HRJSD in SZ was dominated mainly by highly variable SYS patterns in combination with all other eight central pattern families. This was demonstrated by highly significantly decreased SYS-E1 and highly significantly increased SYS-E0, SYS-E2, SYS-LU1, SYS-LD1, SYS-LA1, SYS-P and SYS-V. It seems to be that the blood pressure regulation is more complex and mainly influences the central–vascular coupling pattern in SZ. Furthermore, it could be shown that central–vascular coupling is strongly affected by reduced BPV (SYS-E1) and short-term strong/weak, increasing/decreasing, alternating and fluctuating vascular family patterns (SYS-E0, SYS-E2, SYS-LU1, SYS-LD1, SYS-LA1, SYS-P, SYS-V), in combination with central activity. We could also found that the complexity of the central–vascular coupling is significantly increased in SZ when compared with CO. Central–cardiac coupling was characterized as bidirectional, with central driving mechanisms (PEEG→BBI) towards autonomic system (BBI) in SZ and CO. For both groups, central activity is much stronger towards the autonomic system than in the opposite direction. However, the results indicated that, for SZ, this closed-loop interaction does not work well owing to the known significant heart rate changes for those patients [37,56]. It is presumed that lesions within the CNS may result in profound alterations in cardiac regulation and may even result in potentially fatal cardiac arrhythmias or sudden cardiac death (owing to cardiovascular dysfunctions) [57]. The possibility that greater levels of cerebral dysfunction are associated with an increasing severity of cardiovascular dysfunction [57] were determined by the overall changes in heart rate and blood pressure in SZ patients, thus posing an increased risk of sudden cardiac death for these patients. It has been assumed that an abnormal interplay between frontocingulate and subcortical brain areas can lead to abnormal autonomic arousal, being expressed as a functional disconnection in autonomic and central systems when patients with paranoid schizophrenia process threat-related signals [9]. Thus, the supposition is that paranoid cognition may reflect an internally generated cycle of misattribution regarding incoming fear signals owing to a breakdown in the regulation of these systems resulting in an altered brain–heart interaction, influenced by a lack of cortical inhibitory control over sympathoexcitatory subcortical regions [9]. Thayer et al. showed that resting HRV is tied to the functioning of frontal–subcortical circuits, in the way that a higher resting HRV is associated with the effective functioning of frontal-top-down control over subcortical brain regions that support flexible and adaptive responses to environmental demands [12]. It is worth noting that the disruption of frontal–subcortical circuits has been associated with a wide range of psychopathologies, including SZ [15]. HRJSD results demonstrated that central–cardiac coupling in SZ was mainly characterized by a larger amount of decreased short-term strong/weak, increasing/decreasing central pattern families (PEEG-E0, PEEG-E1, PEEG-E2, PEEG-LU1, PEEG-LD1) and an increased alternating and fluctuating of central pattern families (PEEG-LA1, PEEG-P and PEEG-V). This means that central activity is much more variable and more random, with weaker rhythmic oscillatory components. Moreover, fast alterations of increased and subsequently decreased (BBI-P), fast alterations of decreased and subsequently increased (BBI-V) and alternating (BBI-LA1) of heart rate patterns were increased for SZ compared with CO, indicating a more random central–cardiac coupling with weaker rhythmic components of cardiac cycle intervals in relation to central activity in SZ. Schulz et al. [35] could also demonstrate that autonomic regulation in medicated SZ patients seems to be partly dominated by an increasing amount of invariable heart rate patterns (BBI-E1, BBI-LA1, BBI-V), in combination with alternating SYS (SYS-E2, SYS-LU1, SYS-P). These results suggest an impairment of the baroreflex control feedback loop. They assumed that these effects are probably related to the anticholinergic effects of the antipsychotic treatment. Being a unique feature of the HRJSD approach in contrast to other coupling approaches, we could clearly identify different altered central–autonomic physiological regulatory patterns generated by the interplay of the CNS and the ANS in patients with schizophrenia. One of our results to be highlighted is the finding that, in schizophrenic patients, the central activity had a much stronger variability and higher degree of randomness with less rhythmic oscillatory components than the central activity in healthy controls. From the aspect of biomedical signal processing based on symbolic analysis, the HRJSD approach, based on a redundancy reduction strategy and grouping of single-word types into eight pattern families, enables a detailed description and quantification of bivariate couplings. As a further unique feature in contrast to the classical JSD approach and other coupling approaches [58,59], HRJSD emphasizes a clear characterization of how the couplings are composed, with regard to the different regulatory aspects of the CNS and ANS. To summarize, the applied HRJSD approach creates a bridge between univariate and bivariate symbolic analyses, allowing the quantification and classification of deterministic regulatory bivariate coupling patterns depending on the experimental conditions at hand [35]. It was stated by Williams et al. [9]: ‘that paranoid schizophrenia is characterized by a specific disjunction of arousal and amygdala-prefrontal circuits that leads to impaired processing of significant, particularly threat-related, signals. The pattern of excessive arousal but reduced amygdala activity in paranoid patients points to a dysregulation in the normal cycle of mutual feedback between amygdala function and somatic state (autonomic activity). The concomitant lack of ‘with-arousal’ medial prefrontal engagement suggests that this region cannot undertake its usual role in regulating amygdala-autonomic function, leading to a perseveration and exacerbation of arousal responses’. The precise mechanism of ANS dysregulation in SZ still remains unclear, complicated by the large number of cortical, subcortical and brainstem structures that coordinate autonomic function. However, it was suggested that reduced HRV may represent evidence of an inhibitory deficit that is mirrored by impaired cognitive and behavioural inhibition [60,61]. Moreover, it was shown that the prefrontal cortex could play a critical role in ANS dysregulation and that SZ patients are characterized by a decrease in their prefrontal cortex activity and concomitant deficits in executive function and inhibition [60,61]. Thayer & Lane [60] also proposed that ANS dysregulation is driven by the failure of the prefrontal cortex to inhibit the amygdala-mediating cardiovascular and autonomic responses to stress. They further [14] emphasized the importance of the medial prefrontal cortex as the ‘core integration’ system owing to its assumed critical role in the representation of internally and externally generated information, as well as its integrative function to regulate behaviour and to adapt peripheral physiology. Furthermore, when integrating findings from research concerning the central and cardiovascular effects of increasing emotional intensity, the suggestion is posed that changes in functional central system activation and changes in heart rate and blood pressure are related. Especially noteworthy is the proof that the left and right central hemispheres are specialized for parasympathetic and sympathetic control of cardiovascular functioning [57]. Tanida et al. could show that mental arithmetic (MA) task induced larger activity in the right prefrontal cortex than that in the left prefrontal cortex in subjects with high heart rate increases, suggesting that the right prefrontal cortex activity during MA task has a greater role in the central regulation of heart rate owing to virtue of decreasing parasympathetic effects or increasing sympathetic effects [62]. The relationship between autonomic functioning and cognitive performance in patients with schizophrenia is still exclusive. Mathewson et al. [63] are the first to show associations between autonomic regulatory capacity and neuropsychological performance in patients with schizophrenia. In particular, susceptibility to perseveration in patients was associated with faster heart rates at rest and reduced vagal modulation. Moreover, the authors suggest that the executive function deficits in schizophrenia and autonomic deficiencies reported in this population should be investigated jointly. Unfortunately, neuronal effects of cardiac autonomic dysfunction have not yet been investigated in schizophrenia patients. However, the assumption exists/it has been assumed that an abnormal interplay between frontocingulate and subcortical brain areas can lead to abnormal autonomic arousal [9]. Considering central–cardiac coupling and central–vascular coupling with respect to central spectral power bands, the strongest influence of cerebral γ activity towards BBI and SYS was found for both SZ (here reduced) and CO, independent of the brain hemisphere. This highlights the role of γ activity in SZ and was also demonstrated in multiple studies [51,64]. It was shown that γ and β activity is most augmented in SZ over frontal and temporal brain regions, reflecting a genetic liability for schizophrenia [65]. It was suggested that impaired neural oscillation (e.g. a reduction in amplitude and altered phase synchronization in all frequency bands with emphasis on the β and γ band activity) in schizophrenia patients can be considered a marker for a functional disconnectivity between different brain areas and for dysfunctional cortical networks [66]. Moreover, studies also showed that the parasympathetic and sympathetic nervous systems are lateralized to the left and right central hemispheres, respectively. Furthermore, researchers have proposed that increasing levels of central activation within the left hemisphere are associated with increasing parasympathetic tone [57]. Thus, the differences in central activity in SZ between the two hemispheres would determine the overall changes in heart rate and blood pressure [57]. The central–cardiac and central–vascular coupling directions with respect to central spectral power bands were characterized as bidirectional with BBI and SYS acting as the driver in each frequency band. This may suggest that the autonomous system provides feedback information towards the different central oscillatory components (with the exception of γ). All these components considered together as the whole central activity provide, in turn, feed-forward information to the ANS. A limitation of this study is that SZ patients received antipsychotic treatment. However, on the one hand, it is very difficult to recruit unmedicated patients for such investigations due to the fact that the patients are very instable in their psychotic states and are also less cooperative. While antipsychotic drugs provide a basic therapeutic tool for the treatment of schizophrenia and other psychotic conditions, their effectiveness is associated with a series of unresolved questions. It is not clear, for example, which of the neurobiological mechanisms (beyond D 2 receptor-blocking) is the final therapeutic target responsible for the beneficial effect on distorted information processing in schizophrenia, and for subsequent elimination or reduction of psychotic symptoms [52]. On the other hand, some studies exist [46,48,51,52,64] where the effect of various atypical antipsychotic drugs in schizophrenic patients via quantitative EEG analysis methods are compared. They revealed that in clozapine and olanzapine an increase in slow frequency bands were found, both in routinely treated patients and in healthy volunteers after a single dose and that risperidone produces fewer changes than clozapine and olanzapine [52]. In further research, we will also investigate the couplings of central activity and respiration as well as of electrodermal activity coupling since it was clearly hallmarked that these autonomic variables were significantly altered in SZ [36,37,67]. However, in this study, we did not find any significant differences in respiratory-related variables (electronic supplementary material, table S4). At the moment, we have focused our investigation on short-term instantaneous CAC in the frontal area. However, it is not clear if these CAC are changing over longer time (time-variant) and if they are possibly concentrated in specific frontal areas (clusters). Therefore, time-variant analyses and topographical EEG cluster analyses (some EEG electrodes were formed to a cluster) are necessary. Moreover, to solve these open questions, in addition, functional magnetic resonance imaging (fMRI) data analyses are necessary. However, the advantage of using the EEG technique instead of fMRI is that EEG has a very good temporal resolution. This allows a very good characterization of the central–autonomic network (CAN). The EEG covers many cortical regions of the CAN. In comparison, fMRI has a very good spatial resolution with a very slow response (shown by the BOLD signal). Thus, fMRI does not seem to be the best technique to characterize ANS alterations owing to its slow signal response (BOLD). Therefore, EEG seems to be the more adequate technique in order to establish an association/connection between central activity (cortical EEG) and autonomic function (cardiovascular). However, the EEG is unable to investigate anatomical connectivity owing to its poor spatial resolution. This study is the starting point for an enhanced understanding of the complex brain–heart network between central activity and cardiovascular regulation in SZ patients. We are currently just beginning to understand the interrelationship between the autonomic system in psychotic states, the central networks and control mechanisms in SZ patients. Moreover, we hypothesize that an improvement in autonomic functioning could be achieved through physical fitness (sport intervention), for example, which, in turn, would lead to an improvement in these patients' clinical symptoms. Additional studies are nevertheless needed. To conclude, the scientific impact of this study marks a further step towards a more comprehensive understanding of the interplay of neuronal and autonomic regulatory processes in schizophrenia patients. At the same time, greater insight has now been provided regarding the complex relationship between psychotic stages and (cardiac) autonomic activity. This study may thus contribute to an optimal selection of therapy strategies that would lead to more successful therapy plans for these patients in the future. Results supporting this article have been uploaded as part of the electronic supplementary material (see tables S1–S4). An exemplary dataset of one schizophrenic patient and one healthy subject supporting this article has been uploaded as part of the electronic supplementary material, in the file Schulz_etal_RSTA-2015–0178.zip. All the used data are the property of the University Hospital Jena, Germany and it is strictly forbidden to pass on or distribute all the data online. However, those interested in using the data should contact the authors to prepare an appropriate cooperation. S.S., calculation, analyses and interpretation of central–autonomic indices, manuscript writing. K.-J.B., study design, patient recruitment and study supervision. M.B., study design and measurement of physiological biosignals. A.V., supervision of central–autonomic indices, result interpretation. All authors contributed to and approved the final manuscript. We declare we have no competing interests. This work was partly supported by grants from the Thuringian Ministry of Economy, Labour and Technology, and European Social Fund (ESF) 2014 KN 0018, as well as by grants from the Federal Ministry for Economic Affairs and Energy (BMWI) KF 2447308KJ4. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 9Cardiovascular and cerebrovascular systems are regulated by several control mechanisms aiming at avoiding that physiological variables assume risky values [1]. Among these control mechanisms, cardiac baroreflex and cerebral autoregulation play a relevant role. Cardiac baroreflex is a short-term regulatory reflex that adjusts heart period (HP) in response to changes in arterial pressure (AP) [2]. The baroreflex sensitivity, an index quantifying the magnitude of HP changes driven by a unit variation of systolic AP (SAP), is a very important clinical marker to predict mortality in specific cohorts of patients, e.g. after myocardial infarction [3]. Cerebral autoregulation is a homeostatic mechanism responsible for maintaining mean cerebral blood flow relatively constant, despite the changes in mean AP (MAP) [4–7]. Cerebral autoregulation has been studied for decades but only the advent of transcranial Doppler ultrasound technique, providing a non-invasive estimation of the velocity of the blood in the middle cerebral arteries, allowed the assessment of a variable linked to the cerebral blood flow (CBF), i.e. the CBF velocity (CBFV), with a time resolution similar to that of continuous AP recordings [8,9]. The application of simple procedures leading to the evaluation of the CBFV response to a drop of AP through the inflation of large thigh cuffs [4,10] (i.e. thigh cuff manoeuvre) or a forced exhalation against a closed airway [11–13] (i.e. the Valsalva manoeuvre), and the direct assessment of spontaneous CBFV fluctuations [14–16] has enlarged the possibility of studying the relation between mean CBFV (MCBFV) and MAP series. Respiration is known to influence both baroreflex control and cerebral autoregulation, and this influence is nonlinear, because the effect of respiration on both regulatory systems depends on the respiratory phase [6,7,17–20]. Control breathing experiments proved the important role of respiration in modulating HP–SAP [7,17–19] and MCBFV–MAP interactions [7,17]. These studies pointed out that, on the one hand, respiration must be taken into account when investigating baroreflex control and cerebral autoregulation and, on the other hand, respiration might modulate their interactions. Symbolic analysis (SA) is an emerging branch of signal processing allowing the classification of relevant patterns, while discarding insignificant details [21–33]. SA methods appear to be particularly suited for the assessment of the HP–SAP and MCBFV–MAP joint nonlinear interactions and the inherent nonlinear influences of respiration because they provide strategies for the construction of HP–SAP and MCBFV–MAP joint patterns without presuming linearity, assumed by, for example, cross-correlation or coherence function, and allowing a joint analysis [28–33] gated by respiration [33]. The aim of this study was to apply joint SA (JSA) and joint conditional SA (JCSA) to assess whether respiration influences cardiovascular and cerebrovascular control systems, as described by HP–SAP and MCBFV–MAP symbolic indexes, respectively, and can modulate the crosstalk between them. We hypothesize that, if respiration was able to affect cardiovascular and cerebrovascular controls in a nonlinear way, JCSA would provide different results from JSA. The HP–SAP and MCBFV–MAP variability interactions were assessed according to the JSA and JCSA approaches set in [33] in a population of individuals experiencing recurrent episodes of postural syncope (SYNC) [34]. Syncope was evoked by a prolonged exposure to an orthostatic challenge (i.e. head-up tilt test at 60°). Results were compared with those derived from age- and gender-matched healthy subjects never experiencing postural syncope (non-SYNC) and undergoing the same postural challenge. The contemporaneous evaluation of indexes describing a cardiovascular systemic control and a cerebral homeostatic regulation allowed us to check whether heart–brain interactions were present, and this presence depended on the experimental condition and/or population. We exploited the JSA approach described in [33]. Briefly, given two synchronously recorded time series x={x(i), i=1,…,N} and y={y(i), i=1,…,N}, where i is the sample counter and N is the series length, they are first transformed via uniform binning procedure over ξ=6 bins into a sequence of integers from 0 to ξ−1. Then, L=3 consecutive symbols are grouped together to form patterns. Each pattern shares two symbols with the adjacent one. Thus, the number of patterns in each series is N−L+1. Patterns were classified into four classes [25] according to the shape of the pattern: (i) 0 variations (0V), i.e. all the symbols are equal, (ii) 1 variation (1V), i.e. two consecutive symbols are equal and the third one has a different value, (iii) 2 like variations (2LV), i.e. the pattern looks like an ascending or descending ramp, and (iv) 2 unlike variations (2UV), i.e. the pattern looks like a peak or a valley. From the two series of patterns, we build a series of joint patterns formed by associating one pattern of x and one of y. The two patterns are separated in time by a latency τ, thus the number of joint patterns is N−L−τ+1. Among all possible combinations between patterns of x and y (i.e. 16 families), we consider only joint schemes where the pattern family built over x is equal to that over y. These patterns, referred to as coordinated patterns [33], can be subdivided into four classes labelled as 0V–0V, 1V–1V, 2LV–2LV and 2UV–2UV and their percentage inside the coordinated family can be evaluated (i.e. 0V–0V%, 1V–1V%, 2LV–2LV% and 2UV–2UV%). We remark that the 0V–0V patterns describe coordinated behaviours at the slowest time scale, whereas the 2UV–2UV family those at the fastest time scale. The 1V–1V and 2LV–2LV patterns typify the association among the two series at time scales faster than the 0V–0V ones, but slower than the 2UV–2UV patterns with the interactions described by the 1V–1V class occurring at time scales slower than those illustrated by the 2LV–2LV patterns. In addition, because the direction of the changes does not matter, both in-phase and out-of-phase matched behaviours are accounted for. We exploited the JCSA approach described in [33]. Briefly, given the coordinated patterns defined in §2a (i.e. 0V–0V, 1V–1V, 2LV–2LV and 2UV–2UV), they can be conditioned on the respiratory phase. 0V–0V, 1V–1V, 2LV–2LV and 2UV–2UV patterns whose symbols are associated with events all occurring in the inspiratory (INSP) phase are classified as 0V–0V|INSP, 1V–1V|INSP, 2LV–2LV|INSP and 2UV–2UV|INSP. Analogously, we define as 0V–0V|EXP, 1V–1V|EXP, 2LV–2LV|EXP and 2UV–2UV|EXP the joint patterns whose values are all linkable to the expiratory (EXP) phase. The percentages of the patterns belonging to each class inside the family of the coordinated patterns in the INSP phase are labelled as 0V–0V%|INSP, 1V–1V%|INSP, 2LV–2LV%|INSP and 2UV–2UV%|INSP and those in the EXP phase as 0V–0V%|EXP, 1V–1V%|EXP, 2LV–2LV%|EXP and 2UV–2UV%|EXP. Thirteen SYNC subjects (age: 28±9 years, min.=18 years, max.=44 years; five males) with previous history of unexplained syncope (more than three events during the foregoing year) were enrolled in this study together with 13 non-SYNC healthy subjects (age: 27±8 years, min.=18 years, max.=44 years; five males) [34]. The two groups had similar age and gender composition. The study took place at the Neurology Division of Sacro Cuore Hospital, Negrar, Italy, adhered to the principles of the Declaration of Helsinki for medical research involving humans and was approved by the local ethical committee. Subjects avoided the intake of caffeine or alcohol containing beverage for 24 h before the experiment. All of them signed a written informed consent before performing the experiment. The protocol consisted of 10 min of recording at rest in supine position (REST) followed by head-up tilt test (TILT). TILT was performed in a controlled environment, with subjects laying on the tilt table supported by two belts at the level of thigh and waist and with both feet touching the footrest of the table. The tilt table inclination was 60°. The maximum duration of the TILT session was 40 min. All SYNC subjects experienced presyncope signs before the end of the TILT session and exhibited spontaneous recovery after returning to the supine position. Subjects returned to the supine condition as soon as presyncope signs were observed. None of the non-SYNC subjects experienced presyncope symptoms before the end of the TILT session. Data are available through the corresponding author's ResearchGate profile (https://www.researchgate.net/profile/Alberto_Porta). Electrocardiogram (lead II) was acquired together with AP measured at the level of middle finger through a photopletysmographic device (Finapres Medical Systems, Ohmenda, The Netherlands). CBFV and respiration were measured at the level of the middle cerebral artery by means of a transcranial Doppler ultrasonographic device (Multi-Dop T2, Dwl, San Juan Capistrano, CA) and through a thoracic impedance belt, respectively. Signals were synchronously acquired at a sampling rate of 1000 Hz and stored in a personal computer for off-line analysis. CBFV and respiratory signals were low-pass filtered with a sixth-order Butterworth filter with cut-off frequency of 10 Hz. Attention was paid to avoid phase distortion. From the raw signals, cardiovascular and cerebrovascular variability series were extracted. HP was approximated as the time distance between the ith and the (i+1)th R-wave peaks on the electrocardiogram, where i is the cardiac beat counter. The application of the parabolic interpolation over the R-wave peak allowed the minimization of the jitters in the R-wave apex location. The ith SAP (i.e. SAP(i)) was measured as the maximum of AP signal inside the ith HP (i.e. HP(i)). The ith diastolic AP value (i.e. DAP(i)) was taken as the minimum of AP between the occurrences of SAP(i) and SAP(i+1). We computed MAP values by integrating AP between the occurrences of DAP(i−1) and DAP(i) and, then, by dividing the result by the duration of the ith diastolic interval (i.e. the time distance between the occurrences of DAP(i−1) and DAP(i)). We calculated the MCBFV values by integrating CBFV between the diastolic values (i.e. the minima of the CBFV close to the occurrences of DAP(i−1) and DAP(i)) and, then, by dividing the result by the time distance between the two diastolic values. The peaks and troughs of the respiratory signal were automatically detected, thus defining the INSP and EXP phases as the trough-to-peak and peak-to-trough periods, respectively. The series HP={HP(i), i=1,…,N}, SAP={SAP(i), i=1,…,N}, DAP={DAP(i), i=1…,N}, MAP={MAP(i), i=1,…,N} and MCBFV={MCBFV (i), i=1,…,N}, where N is the total series length, were computed. Sequences of N=250 consecutive synchronous values were selected from each HP, SAP, MAP and MCBFV series. The length of the series allows one to focus on short-term regulatory mechanisms [35]. The beginning of the TILT epoch started 5 min after the onset of the head-up tilt manoeuvre. The rationale of this choice is to avoid the initial transient adjustment of the cardiac variables, thus limiting the influence of non-stationarities over the analysis, and to explore cardiovascular and cerebrovascular response to the manoeuvre before the occurrence of presyncope signs. Selection of the sequences was made at random at REST and in the first 10 min of TILT. Attention was paid to select periods of analysis in which the power spectrum of the respiratory signal featured a dominant peak, thus facilitating the automatic detection of INSP and EXP phases. The detected onset and offset of the INSP and EXP phases were verified by an operator and eventually adjusted. The HP, SAP and DAP series were manually checked for values coming from ectopic beats or misdetections and these values were eventually corrected through cubic spline interpolation. Corrections did not exceed the 5% of the overall length of the sequence considered for analysis. If evident non-stationarities of the mean and the variance were present, the random selection was carried out again. Stationarity of the selected sequences was finally checked according to [36]. Mean and variance of HP, SAP, MAP and MCBFV variability series were extracted, indicated as μHP, μSAP, μMAP, μMCBFV and The null hypothesis of Gaussianity of the distribution of the parameters was tested according to Kolmogorov–Smirnov test. If it was rejected, then the values of the indexes plus 1 were log-transformed before applying any additional statistical test. The addition of 1 allowed us to map 0 again to 0 after the log-transformation. Two-way repeated measures analysis of variance was used to check the significance of the differences between non-SYNC and SYNC groups within the same experimental condition and between experimental conditions (i.e. REST and TILT) within the same group (one factor repetition, Holm–Sidak test for multiple comparisons). Correlation analysis was computed to test the association between symbolic indexes derived from HP–SAP and MCBFV–MAP analyses in non-SYNC and SYNC subjects at REST and during TILT. The Spearman rank correlation coefficient, ρ, and the probability of type I error, p, were computed. Statistical analysis was carried out using a commercial statistical program (SigmaPlot, v. 11.0, Systat Software, Inc., Chicago, IL). A value of p<0.05 was always considered statistically significant. Table 1 shows results relevant to time domain parameters in terms of mean and variance extracted from the considered variability series (i.e. HP, SAP, MAP and MCBFV) at REST and during TILT in non-SYNC and SYNC subjects. As to the mean, regardless of the group (i.e. SYNC or non-SYNC), μHP and μMCBFV significantly decreased during TILT, whereas μSAP and μMAP remained stable. As to the variance, regardless of the group (i.e. SYNC or non-SYNC),
The grouped bar graphs of figure 1 show the results relevant to the rate of occurrence of the 0V–0V joint pattern family (i.e. 0V–0V%) as derived from JSA and JCSA computed over HP and SAP series in figure 1a–c and over MAP and MCBFV series in figure 1d–f. Findings relevant to JSA are shown in figure 1a,d, whereas those relevant to JCSA are depicted in figure 1b,c,e,f divided into those conditioned on INSP (figure 1b,e) and EXP (figure 1c,f) phases. The 0V–0V% index is reported as mean plus standard deviation as a function of the experimental condition (i.e. REST and TILT) in both non-SYNC (black bars) and SYNC (white bars) subjects. When 0V–0V% was assessed over HP and SAP series, regardless of the respiratory phase (figure 1a), 0V–0V% was able to detect differences between conditions but unable to find differences between groups. Indeed, 0V–0V% increased during TILT in non-SYNC subjects, whereas SYNC and non-SYNC groups could not be distinguished both at REST and during TILT (figure 1a). When 0V–0V% was assessed over MCBFV and MAP series interactions and regardless of the respiratory phase (figure 1d), 0V–0V% was unable to detect differences between either conditions or groups. Similar conclusions could be drawn when the HP–SAP variability interactions were conditioned on the respiratory phase (figure 1b,c). Conversely, MCBFV–MAP analysis carried out after conditioning on the respiratory phase detected differences between experimental conditions with the same group: indeed, 0V–0V% in the EXP phase significantly increased during TILT in SYNC individuals (figure 1f), whereas it remained stable in the INSP phase (figure 1e). Figure 1. Grouped bar graphs report 0V–0V% assessed over HP–SAP (a–c) and MCBFV–MAP (d–f) patterns as a function of the experimental condition (i.e. REST and TILT) in both non-SYNC (black bars) and SYNC (white bars) subjects. The analysis was unconditioned on respiration (a,d) and conditioned on respiratory INSP (b,e) and EXP (c,f) phases. Results are reported as mean plus standard deviation. Asterisk indicates p<0.05 between experimental conditions within the same group.
Figure 2 has the same structure as figure 1 but it is relevant to the rate of occurrence of the 1V–1V joint pattern family (i.e. 1V–1V%). When 1V–1V% was evaluated over HP and SAP series regardless of the respiratory phase (figure 2a), 1V–1V% was able to detect differences between conditions and groups. Indeed, 1V–1V% decreased during TILT in non-SYNC and it was larger in SYNC group than in non-SYNC one during TILT (figure 2a). When 1V–1V% was assessed over MCBFV and MAP series and regardless of the respiratory phase (figure 2d), 1V–1V% was unable to detect differences between either conditions or groups. Similar conclusions could be drawn when 1V–1V% was assessed over HP and SAP series after conditioning on the EXP phase (figure 2c) and over MCBFV and MAP series after conditioning to both INSP and EXP phases (figure 2e,f). Conversely, when the 1V–1V% was computed over HP and SAP series during the INP phase a significant difference between conditions was detected in non-SYNC with 1V–1V% dropping during TILT (figure 2b). Figure 2. Grouped bar graphs report 1V–1V% assessed over HP–SAP (a–c) and MCBFV–MAP (d–f) patterns as a function of the experimental condition (i.e. REST and TILT) in both non-SYNC (black bars) and SYNC (white bars) subjects. The analysis was unconditioned on respiration (a,d) and conditioned on respiratory INSP (b,e) and EXP (c,f) phases. Results are reported as mean plus standard deviation. Asterisk and section symbol indicate p<0.05 between experimental conditions within the same group and between groups within the same experimental condition, respectively.
Figure 3 has the same structure as figures 1 and 2 but it is relevant to the rate of occurrence of the 2LV–2LV joint pattern family (i.e. 2LV–2LV%). The absolute value of 2LV–2LV% was quite small, thus suggesting that this family was unlikely. No difference between conditions or groups was detected, regardless of the type of analysis (i.e. JSA or JCSA) and variability interactions (i.e. HP–SAP or MCBFV–MAP analysis). Figure 3. Grouped bar graphs report 2LV–2LV% assessed over HP–SAP (a–c) and MCBFV–MAP (d–f) patterns as a function of the experimental condition (i.e. REST and TILT) in both non-SYNC (black bars) and SYNC (white bars) subjects. The analysis was unconditioned on respiration (a,d) and conditioned on respiratory INSP (b,e) and EXP (c,f) phases. Results are reported as mean plus standard deviation.
Figure 4 has the same structure as figures 1–3 but it is relevant to the rate of occurrence of the 2UV–2UV joint pattern family (i.e. 2UV–2UV%). When 2UV–2UV% was assessed over HP and SAP series, regardless of the respiratory phase (figure 4a), 2UV–2UV% was able to detect differences between conditions but it was unable to find differences between groups. Indeed, 2UV–2UV% decreased during TILT in SYNC subjects, whereas SYNC and non-SYNC individuals could not be distinguished both at REST and during TILT (figure 4a). When 2UV–2UV% was assessed over MCBFV and MAP series and regardless of the respiratory phase (figure 4d), 2UV–2UV% was able to detect differences between both conditions and groups. Indeed, at REST 2UV–2UV% was larger in SYNC group than in non-SYNC one and it decreased significantly during TILT in SYNC individuals (figure 4d). When the 2UV–2UV% indexes assessing the HP–SAP variability interactions were conditioned on the respiratory phase, they were not able to detect either differences between conditions or groups (figure 4b,c). Conversely, 2UV–2UV% assessing the MCBFV–MAP variability interactions in the EXP phase showed that SYNC subjects were significantly different from non-SYNC individuals during TILT (figure 4f) with 2UV–2UV% significantly higher in SYNC subjects. Figure 4. Grouped bar graphs report 2UV–2UV% assessed over HP–SAP (a–c) and MCBFV–MAP (d–f) patterns as a function of the experimental condition (i.e. REST and TILT) in both non-SYNC (black bars) and SYNC (white bars) subjects. The analysis was unconditioned on respiration (a,d) and conditioned on respiratory INSP (b,e) and EXP (c,f) phases. Results are reported as mean plus standard deviation. Asterisk and section symbol indicate p<0.05 between experimental conditions within the same group and between groups within the same experimental condition, respectively.
Table 2 shows results relevant to the correlation analysis between corresponding JSA indexes derived from HP–SAP and MCBFV–MAP analyses as a function of the experimental condition (i.e. REST and TILT) in both groups (i.e. non-SYNC and SYNC). The results of correlation analysis over JCSA parameters are also given. Table 2 reports the Spearman rank correlation coefficient, ρ, and the type I error probability, p. A significant correlation with p<0.05 is marked with an asterisk. It can be observed that the JSA parameters were not significantly correlated. The result held, regardless of condition and group. Remarkably, when JSA was conditioned on the respiratory phase a significant correlation was detected in both non-SYNC and SYNC groups, but the scenario was completely different in the two populations. Indeed, in non-SYNC subjects, several JCSA indexes were significantly correlated and this situation occurred in both respiratory phases and in both experimental conditions. Conversely, in SYNC individuals, solely the percentages of the 1V–1V patterns derived from HP–SAP and MCBFV–MAP analyses were significantly correlated and this situation occurred exclusively at REST and during the INSP phase. It is worth noting that, when significant, the correlation coefficient is always positive.
This study investigates the effect of respiration on cardiovascular and cerebrovascular control systems and the ability of respiration to modulate the interactions between them in a population experiencing recurrent postural syncope. The main findings can be summarized as follows: (i) time domain analysis of cardiovascular and cerebrovascular parameters was not able to differentiate healthy subjects from pathological individuals, whereas JSA could, (ii) a nonlinear influence of respiration on cardiovascular and cerebrovascular control systems was detectable, and (iii) respiration modulated the degree of association between cardiovascular and cerebrovascular control systems and this modulation depended on the experimental condition and population. Under the hypothesis of no interactions or linear interactions of respiration, we expect that (i) the results of JCSA in the INSP and EXP phases would be comparable and similar to those derived from JSA unconditioned on respiration and (ii) the degree of coordination between cardiovascular and cerebrovascular control systems, as measured from the correlation coefficient between HP–SAP and MCBFV–MAP markers, computed in the INSP and EXP phases would be comparable and similar to that calculated regardless of the respiratory phase. The violation of the above-mentioned first condition allows us to detect nonlinear effects of respiration on cardiovascular and cerebrovascular control systems and the infringement of the above-mentioned second condition indicates nonlinear influences of respiration on the interaction between cardiovascular and cerebrovascular control systems. The proposed approach allows one to test both these conditions. TILT provoked the expected changes of time domain parameters. More specifically, given that TILT leads to a sympathetic activation mainly driven by the drop of central blood volume [27,37–39], HP significantly decreased and SAP variance significantly increased in both non-SYNC and SYNC groups. In addition, the reduction of MCBFV during TILT in both populations is in agreement with the literature [40,41], being the likely consequence of the cerebral vasoconstriction associated with the challenge in both groups. The reduction of HP variance during TILT, observed exclusively in SYNC individuals, suggests an accentuated sympathetic activation and/or vagal withdrawal in this group compared with non-SYNC subjects. Unfortunately, time domain analysis based on the computation of mean and variance of cardiovascular and cerebrovascular variables did not allow the direct distinction of the two groups. Conversely, JSA unconditioned on respiration separated not only experimental conditions within the same group, like the time domain parameters, but also groups within the same experimental condition. For example, we confirmed that the percentage of 0V–0V joint pattern describing the HP–SAP variability interactions increased in non-SYNC subjects during TILT [33] (figure 1a), whereas that of 1V–1V family decreased [33] (figure 2a), thus suggesting that sympathetic activation induced by the postural challenge increased the strength of the HP–SAP coupling at slow time scales but provoked the HP–SAP uncoupling at faster ones probably in relation to the sympathetic activation and vagal withdrawal associated with the stressor. Remarkably, the reduction of the percentage of 1V–1V patterns describing the HP–SAP variability interactions during TILT was less marked in SYNC subjects, leading to the separation between the two groups during TILT (figure 2a). This finding, in addition to the negligible increase of the percentage of 0V–0V patterns in SYNC subjects (figure 1a), allows us to speculate that individuals who will undergo postural syncope at the end of the head-up tilt test might fail to modulate the coordination of the HP and SAP dynamics in response to a postural challenge especially at slower time scales. Conversely, this ability is over-expressed at fastest time scales: indeed, solely in SYNC subjects, the percentage of 2UV–2UV patterns describing the HP–SAP variability interactions significantly decreased during TILT (figure 4a). In addition, when JSA was carried out over MCBFV and MAP series, indexes derived from classification of the joint patterns were able to distinguish both experimental conditions and groups. Indeed, the percentage of 2UV–2UV patterns derived from the MCBFV–MAP analysis decreased during TILT in SYNC individuals, and at REST it separated SYNC from non-SYNC subjects (figure 4d). We speculate that sympathetic activation and vagal withdrawal associated with TILT were more effective in decoupling MCBFV and MAP series at fastest time scales in SYNC subjects, mainly because the degree of the MCBFV–MAP coupling at these time scales at REST was stronger in this group than in the non-SYNC one. Because indexes derived from JCSA exhibited the same trends as those derived from JSA, it might appear that nonlinear influences of respiration over HP–SAP and MCBFV–MAP variability interactions are irrelevant. For example, the grouped bar graphs relevant to the percentage of 0V–0V patterns derived from the HP–SAP analysis in INSP (figure 1b) and EXP (figure 1c) phases are comparable and similar to that showing the same index, regardless of the respiratory phase (figure 1a). The comparison of the results of JSA and JCSA relevant to 1V–1V, 2LV–2LV and 2UV–2UV patterns confirmed this impression (figures 2–4). However, a more careful observation of the grouped bar graphs suggests that nonlinear influences of respiration over HP–SAP and MCBFV–MAP variability interactions are present. For example, the percentage of 0V–0V patterns assessed over the MCBFV and MAP series in the EXP phase increased during TILT in SYNC individuals (figure 1f), whereas the same parameter remained steady in the INSP phase (figure 1e) or regardless of the respiratory phase (figure 1d). Another example of the nonlinear effect of respiration was provided by the percentage of 1V–1V patterns derived from the HP–SAP analysis: indeed, the drop in non-SYNC subjects during TILT was more marked in the INSP phase (figure 2b) than in the EXP one (figure 2c) or regardless of the respiratory phase (figure 2a). The correlation analysis between a marker describing the HP–SAP variability interactions and the same index assessing the MCBFV–MAP ones is used to check whether cardiovascular and cerebrovascular control systems interact with each other. The correlation coefficient is taken as an indicator of the degree of coordination between them. When correlation analysis was performed over JSA indexes unconditioned on respiration, cardiovascular and cerebrovascular control systems appeared to work independently. This result held, regardless of the experimental condition and group. Conversely, when JCSA indexes were considered, the opposite conclusion was drawn, thus suggesting JSA might smear influences of respiration by mixing INSP and EXP phases. This finding suggests a possible role of respiration in modulating the crosstalk between different physiological systems. Even more importantly, this modulating capability of respiration depends on the experimental condition and population. Indeed, while in non-SYNC subjects the degree of coordination between cardiovascular and cerebrovascular control systems was significant both at REST and during TILT, a completely different scenario was detected in the SYNC group. Indeed, in SYNC individuals, cardiovascular and cerebrovascular regulatory systems appeared to be coupled only at REST. This finding stresses that in SYNC subjects respiration can modulate the degree of interactions between the two control systems but its ability appeared to be impaired during TILT, thus possibly contributing to the development of postural syncope. Future studies should check whether countermeasures focused on the respiratory drive might be helpful in reversing this trend and preventing postural syncope. Interestingly, coordination between cardiovascular and cerebrovascular control systems in non-SYNC individuals occurred more likely between patterns characterized by slow time scales (i.e. the 0V–0V and 1V–1V schemes) than between those more directly influenced by respiration (i.e. the 2LV–2LV and 2UV–2UV schemes). Therefore, it seems that in both non-SYNC and SYNC subjects respiration might have the possibility to modulate the crosstalk between different control systems at time scales completely different from the dominant time scale of its action, thus stressing again the nonlinear nature of the phenomenon. The study is based on a respiratory signal recorded with a thoracic belt and a min–max procedure delineating the onset and the offset of the respiratory phases. We advocate, on the one hand, the contemporaneous recording of the respiratory activity according to different modalities directly assessing respiratory flow and/or volume to check whether conclusions of this study might depend on the type of signal transduction, and, on the other hand, the test of alternative methods for the delineation of the respiratory phases excluding the typical apnoeic phase at the end of the EXP phase. We also promote studies testing systematically respiratory patterns, alternative to spontaneous breathing, with the final aim to classify them according to their influence on cardiovascular and cerebrovascular control systems. A more systematic approach could improve our knowledge of the ability of the respiratory drive to interfere with physiological control mechanisms and modify key regulatory parameters. In addition, because some overlap exists between the information contained in the symbolic categories, we suggest also to perform specific studies aiming at quantifying possible relations among the percentages of symbolic categories. Because, in principle, a possible link between heart rate asymmetry (i.e. heart rate decelerates more rapidly than it accelerates) [42] and the observed differences between JCSA indexes might exist, we encourage future studies correlating results obtained from JCSA with markers describing heart rate asymmetry [42–44]. The study stresses that different regulatory systems involving systemic cardiovascular and cerebral homeostatic variables can interact with each other and it suggests that the degree of their interaction can be modulated by respiration. Because the detected influences of respiration vary according to the experimental condition and group, the findings support a more systematic use of a physiological input, in large part under voluntary control, such as respiration, to adjust the degree of coordination between cardiovascular and cerebrovascular regulatory systems with the final aim to favour specific control behaviours selected among others to improve the quality of life in pathological subjects and flexibility in coping with stressors in healthy individuals. In this context, the proposed analysis framework could be a viable tool to quantify the outcome of the application of countermeasures, focused on the respiratory function and specifically designed to interfere with cardiovascular and cerebrovascular regulatory systems, in both pathological and healthy individuals. In addition, the study encourages the joint monitoring of indexes derived from different regulatory systems along the brain–heart axis to achieve a more integrated view on the state of the organism. Data are available through the corresponding author's ResearchGate profile (https://www.researchgate.net/profile/Alberto_Porta). V.B.: analysis of the data, interpretation of the data, drafting the manuscript, critical revision of the work, final approval of the version to be published; A.M., B.D.M.: analysis of the data, final approval of the version to be published; G.R.: acquisition of the data, critical revision of the work, final approval of the version to be published; G.N., L.F.: critical revision of the work, final approval of the version to be published; A.P.: conception and design of the work, interpretation of the data, drafting the manuscript, critical revision of the work, final approval of the version to be published. We declare we have no competing interests. We received no funding for this study. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 10Following the advent of transcranial Doppler ultrasound (TCD) for measuring cerebral blood flow velocity (CBFV) with high temporal resolution [1], the study of cerebral haemodynamics has often benefited from data-based input–output predictive models that describe quantitatively how measured physiological variables of interest interact dynamically with each other. Over the past 25 years, a particular focus has been the study of cerebral autoregulation, which is defined as the ability of the cerebrovascular bed to maintain a constant cerebral blood flow in response to pressure changes [2]. Specifically, dynamic autoregulation has been studied by examining the relation between step (e.g. thigh cuff deflations) or spontaneous arterial blood pressure (ABP) changes and the corresponding CBFV variations [3–12]. Furthermore, it is well known that the cerebrovascular bed is exquisitely sensitive to arterial CO2 changes; spontaneous fluctuations of end-tidal CO2 (ETCO2) have been shown to be correlated to both global blood flow using TCD [13,14] as well as regional blood flow using functional magnetic resonance imaging [15]. Therefore, an approach that has gained considerable scientific traction in recent years is to quantify how spontaneous changes in ABP and ETCO2, viewed as two concurrent inputs, cause changes in the output variable of CBFV [13,16–20]. For instance, this approach has been recently used to assess the effect of orthostatic stress [17], as well as possible impairments of cerebral autoregulation or vasomotor reactivity in patients with Alzheimer disease [19] and amnestic mild cognitive impairment [20]. A key concept used in this data-based modelling approach is the set of principal dynamic modes (PDMs) that are the main characteristics of each given dynamic nonlinear system and make the representation, as well as estimation, of nonlinear Volterra-type models feasible and reliable [21]. The PDMs, furthermore, reveal the key functional features of the subject system and, consequently, facilitate model interpretation and offer unique insights into the functional mechanisms of the system. The latter benefit may prove valuable in deciphering the possible physiological causes of a given disease and, therefore, assist clinical diagnosis as well as suggest potential treatments. Last, but not least, this capability can be useful in assessing the effects of treatments or pharmaceuticals in a cohort of patients relative to appropriate controls. This multitude of potential benefits has given recent impetus to this approach, although obstacles remain in terms of data availability (owing to practical constraints) and peer acceptance owing to its relative (perceived) mathematical complexity. This paper seeks to extend the aforementioned model to include changes in heart rate (HR) as a third input and explore the concurrent dynamic effects of HR changes upon CBFV using a three-input model of cerebral haemodynamics (inputs: ABP, ETCO2, HR; output: CBFV). The rationale for this three-input/one-output modelling objective is twofold: (i) the desire to assess quantitatively/statistically the existence of dynamic effects of HR changes upon CBFV and, if present, to elucidate the particular characteristics of these dynamic effects with regard to plausible physiological hypotheses by examining the form of the obtained PDMs; (ii) the extent to which HR and the related dynamic changes in cardiac output may have direct influence on cerebral blood flow independent of changes in ABP and ETCO2. It is evident that the motivation for pursuing these objectives is the development of model-based diagnostic procedures for diseases with a significant cerebrovascular component, as well as the potential discovery of improved treatments that rely on a better understanding of the underlying physiological mechanisms. Time-series data were collected in 18 healthy subjects (nine men and nine women, age 66.8±7.4 years), who participated voluntarily in this study at the Institute for Exercise and Environmental Medicine of the University of Texas Southwestern Medical Center, Dallas, TX, USA, and signed the informed consent form that was approved by the institutional review board. The data were collected in a quiet, environmentally controlled laboratory under supine resting conditions. After 20 min of supine rest, 5 min of recordings were made at an initial sampling rate of 1 kHz. ABP was measured continuously with finger photo-plethysmography (Finapres), ETCO2 tension was obtained via a nasal cannula using capnography (Criticare Systems), and HR was monitored by electrocardiography. CBFV was measured in the middle cerebral arteries using a 2 MHz transcranial Doppler (TCD) probe (Multiflow, DWL) placed over the temporal window and fixed at a constant angle with a custom-made holder. All measurements are non-invasive, safe and comfortable for the subjects. The highly sampled data were subsequently reduced to beat-to-beat measurements using the procedures that have been outlined in our previous publications [19,20]. Figure 1 shows illustrative time-series data over 5 min for one of the subjects. Figure 1. Illustrative time-series data over 5 min for one subject, representing beat-to-beat spontaneous variations of CBFV (a), ABP (b), ETCO2 (c) and HR (d). The units are: cm s−1 for CBFV, beats per minute for HR and mmHg for ABP and ETCO2.
The beat-to-beat data of each control subject were analysed using the method of PDMs to obtain predictive models of the dynamic effects of changes in the three inputs of ABP, ETCO2 and HR upon the output CBFV. We briefly outline below the PDM-based modelling approach. For the many mathematical and technical details of PDM and Volterra-type modelling, the reader is referred to the monograph [21] and to our publications presenting its application to cerebral haemodynamics [19,20]. The PDM-based modelling methodology has its rigorous foundations in the general input–output nonlinear modelling approach initially proposed by Norbert Wiener for Gaussian white-noise inputs in 1958 [22] that was based on the general mathematical foundation of the Volterra functional expansion [23] In this study, four discrete Laguerre functions were used for each input with the Laguerre alpha parameters 0.5, 0.8 and 0.85 for ABP, HR and ETCO2, respectively, which yield the minimum average prediction error over all subjects following a search procedure over a grid of alpha values with increments 0.05. The appropriate number of Laguerre functions was selected on the basis of the Bayesian information criterion (BIC), which takes into consideration the total number of free parameters in the respective Laguerre-based Volterra model. The latter is (L+Q)!/L!Q!, where L denotes the number of Laguerre functions and Q is the nonlinear order of the Volterra model. Like all other criteria derived from statistical hypothesis testing, this criterion is a sufficient (but not necessary) condition for rejecting the null hypothesis of non-significant reduction in residual variance. Having obtained the kernel estimates for the cohort subjects, we apply the aforementioned SVD procedure on the rectangular cohort kernel matrix (simply composed of all kernel estimates) to obtain the PDMs. Three PDMs were selected in this case, using the threshold criterion of the respective singular values being larger than 10% of the maximum singular value. The obtained PDMs form a filter-bank for each input that receives the respective input signal and generates (via convolution) signals, which are subsequently transformed by the associated nonlinear functions (ANFs) that represent the nonlinear characteristics of the system for each PDM. The coefficients of the ANFs (cubic polynomials in this application) are estimated via multi-linear regression of the PDM outputs and their powers upon the output signal—because the outputs of the ANFs are summed to form the model output prediction as (also depicted schematically in figure 2) Figure 2. Block diagram of the three-input/one-output PDM-based model with three global PDMs for each input, x1: ABP, x2: ETCO2 and x3: HR. The output um,j(n) of the jth PDM for the mth input is the convolution ofthe PDM with the respective input. In this application, the ANFs are cubic polynomials: zm,j=fm,j (um,j). The model prediction is the sum of the ANF outputs (plus a constant offset). The units of each PDM are: (output units)/(input units)/second. (Online version in colour.)
As indicated in figure 2, the output of the PDM-based model is formed by additive signal components that are generated by cascaded operations of convolutions of each input signal with its respective PDMs and subsequent static nonlinear (cubic in this case) transformations by the respective ANFs. The obtained PDMs constitute the common functional basis for the representation of the model dynamics for all subjects (often termed global PDMs), resulting in compact dynamic nonlinear model representations. The ability of the global PDMs to represent the dynamics of the entire ensemble of subjects is validated by the predictive accuracy of the global PDM-based model for each subject. Three global PDMs and cubic ANFs were found to be adequate for each input of this model, according to the BIC. The use of three global PDMs for each input and cubic ANFs keeps the total number of free parameters for the three-input PDM-based model low (i.e. 27 plus an offset constant). We emphasize that the global PDMs are common for all subjects, but the estimated ANFs associated with each global PDM are subject-specific and can be used to characterize uniquely the cerebral haemodynamic characteristics of each subject. This is the basis for the potential utility of the PDM-based modelling approach for clinical diagnosis. We begin with the statistical assessment of the effect of including the third HR input and/or the transition from a linear to a nonlinear model. We pursue this by using the BIC, which balances the reduction in NMSE of the output prediction by the three-input and/or nonlinear model with the number of free parameters contained in the respective models. Under the assumption that the model prediction errors are independent and identically distributed according to a normal distribution, the BIC is defined as where
Following the procedure outlined in the Methods, we obtained the global PDMs from the cohort of 18 control subjects that are shown in figure 3 for the ABP input figure 4 for the ETCO2 input and figure 5 for the HR input (in the time and frequency domains). We observe that the global PDMs exhibit distinctive spectral characteristics in terms of resonant peaks, which may attain importance in the interpretation of the PDM-based model and the functional properties of this system (see Discussion). Specifically, we observe the following. Figure 3. The obtained three global PDMs for the ABP input of the three-input model with CBFV output in the time domain (a) and frequency domain (b): first PDM (solid),second PDM (dotted), third PDM (short dotted). The units in the ordinate axis of the time-domain PDMs are (cm s−1)/mmHg s−1.
Figure 4. The obtained three global PDMs for the ETCO2 input of the three-inputmodel with output CBFV in the time domain (a) and frequency domain (b): first PDM (solid), second PDM (dotted), third PDM (short dotted). The units in the ordinate axis of the time-domain PDMs are: (cm s−1)/mmHg s−1.
Figure 5. The obtained three global PDMs for the HR input of the three-input model with output CBFV in the time domain (a) and frequency domain (b): first PDM (solid), second PDM (dotted),third PDM (short dotted). The units in the ordinate axis of the time-domain PDMs are: (cm s−1)/(beats min−1) s−1.
For the ABP input (figure 3)
For the ETCO2 input (figure 4)
For the HR input (figure 5)
The PDMs are obtained by performing singular value decomposition on the matrix containing the first- and second-order Volterra kernels from all subjects. In order to illustrate their variability, we used the bootstrap method to obtain 90% confidence intervals [25]. The results are shown in figure 6, where it can be seen that the PDMs corresponding to ABP exhibit the least variability, whereas the PDMs corresponding to HR exhibit the most. Figure 6. Bootstrap-based 90% confidence intervals for the PDMs corresponding to all three inputs.
Unlike the PDMs, the (cubic) ANFs are estimated via least-squares regression and, therefore, the three coefficients (linear, quadratic and cubic) are statistical estimates with some probability density function. The average (cubic) ANFs defined by the average coefficient estimates for each of the PDM branches of the three-input model are shown in the model diagram of figure 2. It is evident that ETCO2 and HR have the strongest nonlinearities. The ANFs for the HR input exhibit symmetry about the origin (operating point); however, only one ANF for the ETCO2 input exhibits such symmetry (see Discussion). Our previous studies have shown that most of the nonlinear characteristics of the cerebral haemodynamics are exhibited in the low-frequency range—particularly below 0.05 Hz [13,17,19,20]—with regard to the ETCO2 input. Therefore, the nonlinearities are expected to be associated primarily with the PDMs that exhibit resonant peaks in that frequency range. This is, in fact, observed in the ANFs shown in the PDM-based model of figure 2. However, when we assess the statistical significance of the obtained estimates of the ANF coefficients using p-values (under the assumption of system stationarity), we observe significant nonlinearity only in the quadratic coefficient of the first PDM of the ETCO2 input (p=0.039), as indicated in table 2 where the average (s.d.) values of these coefficient estimates and the corresponding p-values are shown for the three PDMs of each input. It is also seen in table 2 that several linear coefficient estimates do not rise to the level of statistical significance. In the Discussion section, we consider the possible requirement of longer data records and/or larger sets of subjects with respect to the statistical significance of nonlinear estimates, as well as the possible effects of system non-stationarities on these results.
Although the PDM-based model separates the dynamics (PDMs) from the nonlinearities (ANFs), it cannot be generally assumed that the system nonlinearities are ‘separable’ in terms of the specified PDM–ANF pathways—i.e. cross-interactions may generally exist between ANF outputs of the model. For this reason, we must generally include in the PDM-based model cross-terms in the form of pair-products of ANF outputs, which have a statistically significant correlation coefficient with the output signal, based on hypothesis testing of the w-statistic which is the Fisher transformation of the correlation coefficient estimate. In this study, only two subjects out of the 18 exhibited a single statistically significant pair-product between ANF outputs of two different inputs (PDM2 of HR with PDM1 of ETCO2 in subject 9 and PDM1 of ABP with PDM3 of ETCO2 in subject 14). Two other subjects also exhibited a single statistically significant pair-product between two ANF outputs of the ETCO2 input (PDM1 with PDM2 of ETCO2 in subject 4 and PDM1 with PDM3 of ETCO2 in subject 7). Because such statistically significant cross-terms were not consistently found in most of the subjects, we have not included cross-terms in the structure of the PDM-based model used for all subjects (figure 2). To illustrate the relative contributions of the PDM–ANF branches of each input to the model prediction of the output, we show in figure 7 the separate contribution of each input (i.e. the sum of the respective three ANF outputs for each input) in the time and frequency domains, along with the total model prediction and the actual output signal for one subject. It is evident that all three inputs make significant contributions, although with different spectral characteristics. The ETCO2 contribution is more pronounced in lower frequencies and does not show the effects of respiratory sinus arrhythmia evident in the contributions of the other two inputs. The ABP contribution is the strongest (average root-mean-square (RMS) value of 1.86), with the ETCO2 contribution second (average RMS value of 1.41) and the HR contribution third (average RMS value of 0.92). Figure 7. The separate contributions of the three inputs to the model CBFV output prediction for one subject in the time domain (a) and frequency domain (b): ABP-driven contribution (i), ETCO2-driven contribution (ii), HR-driven contribution (iii). The total model prediction of the CBFV output is shown in (iv) and the actual CBFV output is shown in (v).
The PDM modelling study of cerebral haemodynamics in 18 control subjects (nine men and nine women) has shown that the inclusion of HR as a third input (along with ABP and ETCO2) in the model that predicts the CBFV output causes a statistically significant reduction in the resulting BIC value for all subjects. We note that the BIC balances the changes in prediction NMSE and the number of free parameters in the respective model. This implies that there exist contributions of HR changes to CBFV changes that are independent from the contributions of ABP and ETCO2 changes, which are included as concurrent inputs of the model. Furthermore, this finding demonstrates that the HR dynamic effects on CBFV changes can be captured and quantified, along with the concurrent effects of ABP and ETCO2, by a three-input PDM-based model that can be estimated from time-series data of each subject. It is noted that the inclusion of HR as a third input does not alter the PDMs of the other two inputs obtained by use of the two-input model. This is additional evidence of the independent influence of HR changes. An important question that arises from critical contemplation of the proposed approach is whether the obtained PDMs (being singular vectors derived from SVD of the kernel matrix, and not statistical estimates) reflect actual physiological entities or they are simply by-products of the computational method. This valid concern can be addressed through statistical analysis of the ANF coefficient estimates that are associated functionally with the PDMs, because these coefficient estimates (obtained via least-squares regression) directly define the contribution of each PDM to the model prediction. When we performed the traditional analysis of statistical significance (p-values) on the obtained ANF coefficient estimates, only a subset of these coefficients were shown to be statistically significant (table 2) despite the fact that nonlinear models were found to satisfy the BIC as discussed above. This may imply that a larger sample of subjects and/or longer data records may be required to demonstrate the statistical significance of some of these coefficient estimates (because the statistical hypothesis testing is a sufficient but not necessary condition). It may also indicate that the presence of possible system non-stationarities, which has been investigated recently [26,27], may affect the results of traditional hypothesis testing (p-values). Examination of the relative contributions of the three inputs (in terms of their average RMS values over all subjects) to the prediction of the output reveals that ABP makes the largest contribution (RMS mean = 1.78) with the ETCO2 second (RMS mean = 1.37) and the HR third (RMS mean = 0.93). The importance and independence of the contribution of HR changes to the dynamic changes in CBFV is illustrated in figure 5 for one subject, along with the relative contributions of ABP and ETCO2 changes. It is evident that the ETCO2 contribution is mainly in lower frequencies (less than 0.05 Hz), whereas the other two inputs make contributions over a similar frequency range (as illustrated in the respective spectra shown in figure 5) and exhibit clearly the effects of respiratory sinus arrhythmia—although the latter effects are about six times stronger in the ABP spectrum than in the HR spectrum. These findings demonstrate the potential of this modelling approach to explain the dynamic changes in CBFV under the combined influence of concurrent changes of all three inputs and, thereby, achieve a reliable representation of the haemodynamics of each subject that can be used potentially for clinical diagnosis of cerebrovascular disease. To the best of our knowledge, one study has considered the effects of three different inputs on CBFV. Specifically, a recent study by Panerai et al. [18] examined the combined effects of ABP and ETCO2 changes on CBFV changes in response to a motor task (elbow flexion), aiming to disentangle physiological effects from neurovascular coupling (effect of motor task on CBFV). Therefore, the effects of HR were not considered and the effect of the motor task was incorporated as a third input in their multiple-input autoregressive model using a perfect step function, i.e. an idealized signal that is rather different in nature than experimental measurements of physiological fluctuations, which exhibit broadband characteristics. With respect to the effects of ABP and ETCO2 on CBFV variability, their results overall agree with this study, in that ABP explains mainly fast CBFV variations and ETCO2 explains the slow CBFV variations (e.g. figs 6 and 3 in [18]). ETCO2 was found to have a slightly larger contribution to CBFV variability than ABP—albeit more variable across subjects and subsequently less statistically significant (table 1 in [18]). However, this could be due to the fact that there appears to be a strong correlation between the third input of their model (elbow flexion) and ABP, and also to that the mean value of ABP and CBFV increased during elbow flexion. Overall, our results regarding the effects of ETCO2 agree with previous two-input studies using Volterra models [13] and multiple coherence [28,29]. A note should be made regarding the use of NMSE as a criterion for model performance and a means for model validation. Although most investigators would accept the ability of a model to reduce the residual variance as evidence of improved model performance, it must be noted that a reduction in residual variance may be accompanied by an increase in estimation variance, which is a detriment to model performance. The balance in this delicate trade-off often depends on the statistical characteristics of the data-contaminating noise, which are generally unknown; one way to achieve a good trade-off is to use a statistical criterion such as the BIC. The PDMs of this model show promise in revealing the important dynamic characteristics of the underlying physiological mechanisms and allow the formulation of specific quantitative hypotheses about the role of the sympathetic and parasympathetic activity (corresponding to specific PDMs) in cerebral perfusion and autoregulation. Consistent with the widely held view regarding the temporal characteristics of autonomic control of the cardiovascular system, it is posited that the HR PDMs with a resonant spectral peak around 0.2 Hz may be associated with parasympathetic activity, whereas the PDMs with a resonant spectral peak around 0.12 Hz may be associated with sympathetic activity. Hamner & Tan [8] investigated the relative contributions of sympathetic, cholinergic and myogenic mechanisms to cerebral autoregulation using separate blockades of alpha-adrenergic receptors, muscarinic receptors and smooth-muscle calcium channels, respectively. Their data were collected during the application of periodic low-body negative pressure (LBNP) at a frequency of 0.03 Hz under the premise that this forcing will ‘actively engage the mechanisms of cerebral autoregulation’. The delineation of the relative contributions of the three separate blockades was sought through the statistical method of ANCOVA decomposition, combined with projection pursuit regression, and was limited to the study of the steady-state autoregulatory curve at the LBNP frequency of 0.3 Hz—unlike our study that examines the resting-state dynamics over all frequencies. Their main finding was that the primary influence on the slope of the autoregulatory curve (at 0.03 Hz) within the range of active autoregulation is exerted by sympathetic activity. This finding is consistent with our PDM model interpretation that the third PDM of the ABP-to-CBFV relationship may be associated with sympathetic influences on cerebral autoregulation in low frequencies within the region of active autoregulation (resting state), because this third PDM has larger values in the low frequencies (less than 0.05 Hz). The relative importance of the physiological mechanisms described by the PDMs is quantified by the values of the ANFs. This interpretation of the PDM-based model, if it becomes confirmed in future studies, may allow the development of model-based diagnostic tools for diseases with a significant cerebrovascular component (such as neurodegenerative diseases, stroke, diabetes, etc.) and enable the quantitative assessment of interventions or pharmacological treatments of such diseases. The obtained ANFs in the PDM-based model are subject-specific (whereas the PDMs are common to the entire cohort) and indicate that there exist significant nonlinearities in the dynamic effects of the ETCO2 and HR inputs upon the output CBFV. It is seen in figure 2 that some ANFs of the ETCO2 and HR inputs are not symmetric about the origin, indicating the asymmetric effect of ETCO2 (hypercapnia versus hypocapnia) upon CBFV. No significant nonlinearities were observed for the dynamic effects of the ABP input within the narrow range of ABP variations in the resting-state data analysed in this study. It is surmised that these data remain within the plateau region of the celebrated pressure–flow homeostatic nonlinearity (i.e. much larger variations of ABP are needed in the data before the pressure–flow homeostatic nonlinearity can be captured by the model). In figure 8, we illustrate the model-predicted CBFV response to concurrent pulse changes of HR and ABP, while keeping ETCO2 at baseline (figure 8a), and to concurrent pulse changes of ETCO2 and ABP, while keeping HR at baseline (figure 8b). The pulse strength is the half-average RMS value of the respective time-series data. The autoregulatory response to ABP changes is evident by the rapid return of CBFV to the steady state after the onset of the ABP input pulse, and it is symmetric when ETCO2 and HR are at baseline. However, the peak of the autoregulatory response is asymmetric when ETCO2 is elevated or HR is not at baseline. We also observe the anticipated increase in CBFV for increased ETCO2 (vasomotor reactivity) or increased HR, as well as the asymmetric decrease in CBFV for decreased ETCO2 or decreased HR (owing to their respective nonlinearities). The dynamic response of CBFV is faster to HR changes than to ETCO2 changes. Figure 8. The model-predicted CBFV response (i) to concurrent pulse changes of ABP (ii) and HR (iii), while keeping ETCO2 at baseline (a), and to concurrent pulse changes of ETCO2 and ABP, while keeping HR at baseline (b). The symmetric autoregulatory response to ABP changes is evident, as well as the asymmetric increase/decrease in CBFV for increases/decreases of HR or ETCO2 owing to the respective nonlinearities. (Online version in colour.)
The presented data-based modelling approach holds the promise of revealing new and valuable insights into the dynamic characteristics of cerebral haemodynamics and autoregulation in future extended studies. Its ability to yield quantitative subject-specific measures of cerebrovascular characteristics engenders the prospect of improved diagnostic procedures for diseases with a significant cerebrovascular component. To the best of our knowledge, it is the first study to demonstrate that, in addition to the well-established effects of ABP spontaneous fluctuations on CBFV, there are direct cardiac effects on the latter, as shown by the contribution of HR. Therefore, it is a potentially valuable tool in advancing our understanding of heart–brain interactions. V.Z.M. designed and directed the data analysis/modelling work and wrote the first draft of the paper. G.D.M. contributed significantly to the design of the analysis/modelling study and made critical contributions to the manuscript, including the preparation and submission of the final draft of the paper. D.C.S. performed most of the data analysis/processing and took care of the proper presentation of the results, including all figures and tables. R.Z. provided the experimental data and offered valuable physiological/scientific guidance in the design of the analysis/modelling study and the interpretation of the results. We declare we have no competing interests. This work was supported in part by the Biomedical Simulations Resource at the University of Southern California under NIH/NIBIB grant no. P41-EB001978 and NIA grant no. R01AG033106-01 to the UT-SWMC. We thankfully acknowledge Dr T. Tarumi for his contribution to the collection of the data. This work was supported in part by the Biomedical Simulations Resource at the University of Southern California under NIH/NIBIB grant P41-EB001978 and NIA R01AG033106-01 grant to the UT-SWMC. No competing interests or conflict of interest exist. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 11Cerebrovascular disease is a significant human, social and economic burden worldwide. In the USA, it is the fourth leading cause of death and a leading cause of permanent disability. The current total annual cost of cerebrovascular disease is $105 billion in the USA, and it is expected to increase to $240 billion in 2030 [1]. Despite the usefulness of global risk assessment, it is estimated that only a fraction of cerebrovascular disease risk is explained by currently observable risk factors such as hypertension, dyslipidaemia, diabetes, smoking and arterial compliance in systemic circulation [2]. Hence, there is an urgent need for novel, individualized cerebrovascular risk signatures capable of detecting additional risk factors such as intima-media thickening, atherosclerotic plaques and stiffening of cerebral arteries. Because these goals could be achieved through early, selective detection of decreased cerebrovascular compliance, the central aim of this work was to design and validate novel, non-invasive, magnetic resonance imaging (MRI)-based biomarkers to evaluate cerebrovascular compliance through direct measures of blood volume changes with high sensitivity and speed. Compliance is an important marker of the biomechanical properties of human tissues in normal as well as disease states. When considering a tissue compartment (e.g. blood, cerebrospinal fluid—CSF—or brain parenchyma), compliance can be quantified by the magnitude of change in the compartment volume owing to a given change in pressure, and is inversely related to the stiffness of the structure that bounds the compartment. In the human body, such pressure changes are commonly effected by travelling pressure waves, such as endogenous cardiac and respiratory pulse waves, and are one of the mechanisms through which brain–heart interactions occur. Ageing and pathologies such as hypertension, dyslipidaemia and diabetes mellitus can cause stiffening of the aorta and other large arteries of the human body, resulting in an increase in pulse pressure in central and peripheral arteries [2]. In turn, increased pulse pressure can induce local changes in compliance by influencing local arterial remodelling in the brain such as increased wall thickness and stiffness, the development of stenosis and plaques and/or a higher likelihood of plaque rupture. Accordingly, several studies conducted on large and small artery disease suggest that increased arterial stiffness may be predictive of cerebrovascular events, with a local increase of pulse pressure in large (e.g. carotids) as well as small (e.g. arterioles) cerebral arteries [3–5] as the underlying mechanism. In this context, arterial stiffness (or its inverse, compliance) has been characterized extensively in systemic circulation, and these studies have been crucial to the understanding of key haemodynamic parameters such as vessel wall compliance, impedance and reflection coefficients in both health and disease. The most simple, robust, reproducible and widely employed non-invasive marker of systemic compliance is the aortic pulse wave velocity, i.e. the speed at which the peak (or the ‘foot’) of the cardiac pulse wave moves between the carotid and the femoral arterial sites. Importantly, aortic pulse wave velocity has proven to be an important cardiovascular risk factor in patients with cardiovascular disease (e.g. myocardial infarction, coronary artery disease, stroke and heart failure [6]). Nevertheless, in spite of its potential usefulness as an early biomarker of cerebrovascular disease, cerebrovascular compliance is currently estimated through indirect Doppler sonography and MRI measurements (e.g. pulsatility and resistivity index) of blood velocity (rather than blood volume) changes. Such measurements are performed mainly in the carotid arteries and larger intracranial arteries. Because of limited access to intracranial tissue encased by the bony skull, grayscale, colour and spectral Doppler sonography only allow the investigation of morphological (intima-media thickness, plaque location) and functional (blood flow) abnormalities of extracranial arteries, and are usually employed to detect carotid stenosis with high sensitivity but moderate specificity [7]. Transcranial Doppler ultrasound (TCD) can measure functional bulk flow velocity changes in few large intracerebral vessels, although with reduced spatial resolution. TCD is also limited by anatomical variance in the population—10% of individuals lack the necessary ‘acoustic window’ that enables successful intracerebral vessel insonation [8]. Further, a few MRI techniques are currently used to study vascular integrity. Magnetic resonance angiography [3] can be employed to investigate structural abnormalities of extracranial and intracranial arteries in a non-invasive manner. Gated CINE-MRI [9] is a functional technique that has been mainly used to measure intracranial compliance to cardiac pulsatility in the ventricles and CSF spaces by measuring CSF velocity across the cardiac cycle [10,11]. Gated CINE-MRI has also been employed for imaging of the cardiac function [12], but is not currently employed to measure cerebrovascular compliance, most probably because of limited speed and spatial coverage (few minutes per slice to quantify bulk flow velocity in three directions). In summary, there is an unmet need for novel non-invasive functional methods to assess cerebrovascular compliance through direct measures of blood volume changes with high spatio-temporal resolution, sensitivity and spatial (e.g. whole-brain) coverage. The central aim of this work was to provide a novel MRI-based endogenous method to assess cerebrovascular and brain parenchymal compliance by exploiting the vessel wall mechanical response to cardiac and respiratory pressure waves. This was achieved as follows. First, we solved the Bloch equation for non-stationary (flowing) spins in a steady state during the MRI acquisition and constrained the sequence parameters to produce a novel endogenous MRI contrast dependent primarily on cerebral blood volume changes. The endogenous contrast for non-stationary spins was optimized to achieve an approximately sixfold gain compared with Ernst angle acquisitions. Then, a fast MRI sequence based on echo-planar imaging (EPI), implemented on high-field scanners, was employed along with dedicated processing to compute an average cardiac and respiratory MRI pulse waveform in each voxel of the brain, representing the brain response to physiological pressure waves. The MRI pulse waveform was obtained for several cerebral arteries, veins, voxels within the CSF and voxels within the brain parenchyma. Notably, in cerebral vessels, this waveform was expected to be directly related to blood volume changes in response to physiological pressure waves, and was used to derive a novel indicator of cerebrovascular compliance, the pulsatility volume index (pVI). In a group of healthy subjects, the pVI owing to the cardiac pulse wave was measured for several cerebral arteries; further, the origin of the MRI signal changes underlying the pVI was investigated, by characterizing the contribution of the signal changes at echo-time equal to zero (S0) and of The aim of this work was to solve the Bloch equation for non-stationary (i.e. flowing/moving) spins in a steady state throughout the acquisition, and constrain the sequence parameters to produce an endogenous MRI contrast: (i) dependent primarily on changes in total spin concentration (i.e. volume) and (ii) optimized for non-stationary spins under certain conditions (fast-flowing spins—see definition of regime 3 below) in order to achieve a large (sixfold) gain compared with Ernst angle acquisitions, which instead provide the maximum signal for stationary spins. In general, for a given compartment k in the brain, the gradient-echo MRI signal changes over time (M(t)) owing to the propagation of physiological pulse waves are related to both changes in the velocity (v(t)) and in the total concentration (V (t)) of moving/flowing spins (e.g. blood flow velocity and blood volume for blood compartments, respectively) as follows: 2.1 where Mxy represents the spoiled equilibrium transversal magnetization, T1 andThe MRI signal dependence on these two parameters (v(t),V (t)) was inspected in three regimes: regime 1, stationary spins (with velocity v=0 cm s−1); regime 2, non-stationary spins flowing/moving at a velocity smaller than the critical speed vC (i.e. the speed at which there is complete inflow of new spins and thus complete spin replacement in the slice, where vC=ST/TR, with ST=slice thickness); regime 3, non-stationary spins flowing at a velocity greater than or equal to the critical speed (v≥vC). Specifically, in this work, we reported on the solution of the Bloch equations under steady-state acquisitions in regimes 1 and 2 [13], explicitly investigating its dependence on spin velocity, rather than, for example, on the FA as previously shown [13]; further, we extended the solution of the Bloch equations to regime 3. In the three regimes, for a short-TR two-dimensional (slice-selective) gradient-echo imaging method, under steady-state conditions (i.e. after the delivery of p radiofrequency or RF pulses, p depending on the regime, see below), the expected Mxy, solution of the Bloch equations, is (see figure 1 for a plot of Mxy in the three regimes for blood) as follows. Figure 1. Theoretical predictions on the expected transversal magnetization Mxy for stationary (regime 1) and non-stationary spins with velocity lower (regime 2) and greater than or equal (regime 3) to the critical speed vC. We computed Mxy (equations (2.2), (2.4), (2.7) in ‘Theory’ section for regimes 1–3, respectively) for spin velocities spanning the three regimes (regime 1, velocity=0 cm s−1; regime 2, spins with velocity range= [0.36 1.8] cm s−1 obtained using vC/v in the range=[2 100]; regime 3, spins with velocity≥3.6 cm s−1 (critical speed vC= 3.6 cm s−1) obtained with a ST=1.2 mm and a TR= 33 ms). We varied the FA (range [0° 90°]) and focused on the blood compartment, assuming M0=1, and T1blood=2.5 s at 7 Tesla [14]. We show in (a) Mxy for velocities between 0 and 11 cm s−1 (at five different FA); (b) a zoomed view of (a) displaying Mxy for velocities between 0 and 0.8 cm s−1; (c) the dependence of Mxy on the FA, as usually reported in MRI textbooks [9,13] (for the sake of simplicity, Mxy in (a)–(c) is displayed for TE=0, i.e. excluding possible
(1) Stationary (v=0) spins (regime 1): 2.2 where G is a gain factor, and Mz(v=0) is the longitudinal magnetization for stationary spins equal to2.3 where q equals(2) Non-stationary spins flowing at v<vC (regime 2): 2.4 Mz(v=0) and q as defined above. To achieve the steady state, the number p of RF pulses needs to be greater than or equal to vC/v (so that the velocity segment that last entered the slice experienced 1 RF pulse, whereas the first entered segment experienced vC/v RF pulses—note that for the sake of simplicity, we assumed integer values of vC/v). Crucially, the magnetization in regime 2 depends on spin velocity, is greater than in regime 1 (Mxy(0<v<vc)>Mxy(v=0), a phenomenon usually reported as ‘flow-related enhancement’), decreasing with the spin velocity, and reducing to Mxy(v=0) for v→0; further, the optimal FA depends on the spin velocity (figure 1). Interestingly, in this regime, for FA=90°, q is equal to zero and Mxy(0<v<vc) increases linearly with the spin velocity v:2.5 which after replacing equation (2.3) in equation (2.5) (with q=0) reduces to2.6 (3) Non-stationary spins flowing at v≥vC (regime 3):2.7 For v=vC, this equation can be derived considering that spins flowing/moving at vC experience only one RF pulse (all spins at the vC that enter a slice at one RF pulse leave the slice by the next RF pulse); as a result, the transverse magnetization relates only to the projection of M0 in the transverse plane after the application of the RF pulse, and toIn summary, the theory predicts that pulsatile MRI signal changes during physiological cycles are expected to be related to both v(t) and V (t) in regime 2, and, crucially, primarily to V (t) (mostly irrespective of pulsatile changes in v(t), as well as of TR and T1) in regime 3 (see Mxy in figure 1): 2.8 and2.9 We therefore hypothesized to produce an endogenous MRI contrast in several blood compartments, which would be dependent primarily on cerebral blood volume (rather than velocity) changes owing to physiological pulse waves by constraining the sequence parameters (TR, ST) to obtain a critical speed (e.g. approx. 3.6 cm s−1, vertical line in figure 1) lower than for instance the (diastolic) blood velocity in large and middle cerebral vessels.Further, the theory predicts that compared with Ernst angle acquisitions (approx. 10° for blood and the short TR employed) a 90° FA enables a many-fold gain (equal to 6.2 for blood) in endogenous signal of fast flowing/moving spins (regime 3), and a many-fold ‘background suppression’ (equal to 6.2 for blood) of stationary spins (regime 1) (see also figure 1); thus, we hypothesized that the use of a 90° FA would be optimal to investigate pulsatile signal changes of fast flowing/moving spins, without the need of using an exogenous contrast agent. We performed four MRI experiments in order to (i) verify theoretical predictions on the optimal use of a 90° FA to study non-stationary spins in regime 3, (ii) estimate the MRI pulse waveform in response to cardiac and respiratory effects in several brain compartments and extract an indicator of cerebrovascular compliance, the pVI, in larger cerebral arteries, (iii) investigate the origin of the MRI pulse waveform, namely S0 and possible Gradient echo (GE) single-echo EPI was performed on three subjects (3f, age 24±2 years) with the following parameters: TR/TE/FA=33 ms/18 ms/90°, inplane voxel size=1.2×1.2 mm2, ST=1.2 mm (vC=3.6 cm s−1), Nscans=625, GRAPPA factor=5, bandwidth (BW)=1666 Hz per pixel, one coronal slice, acquisition time (TA) per slice ≅20 s. To scrutinize changes in Mxy of spins in regime 3, the same slice was repeatedly acquired while varying the FA (FA∈[10° 30° 45° 65° 90°]). For each subject and FA, the signal in a manually defined region of interest (ROI; nine voxels) within the right internal carotid artery (regime 3, v≥vC) was averaged across time (after linear detrending), and then the ratio of Mxy(v≥vc) obtained with an FA=90° and 10° was calculated. Finally, we performed a similar study using a readout scheme other than EPI, and acquired on one subject (female 23 years) cardiac-gated GE CINE images (without flow-encoding gradients to quantify bulk flow) with FA=10° and 90° (in two separate acquisitions) and the following parameters: nominal TR/TE=38.12 ms/5.78 ms, voxel size=1.15×1.15 mm2, ST = 1.2 mm, GRAPPA factor=2, BW=775 Hz per pixel, 22 phases (i.e. time-points during one cardiac cycle), TA equal to 49 s and 52 s for each acquisition. GE single-echo EPI was performed on three subjects (2m/1f, age 30±6 years) with the following parameters: TR/TE/FA=51 ms/22 ms/90°, inplane voxel size=1.2×1.2 mm2, ST = 1.2 mm (vC=2.4 cm s−1), Nscans=600, BW=1645 Hz per pixel, GRAPPA factor=4. The acquisition was stepped through 25 spatially contiguous axial slices (axial slab, TA/slice approx. 30 s, total TA/slab approx. 12.5′) covering the occipital/temporal/frontal poles (including larger cerebral arteries, the sagittal sinus, and the lateral and third ventricles). To better cover the carotid arteries, similar GE single-echo EPI measures were performed on four subjects (4f, age 24±2 years) with the following parameters: TR/TE/FA=33 ms/18 ms/90°, inplane voxel size=1.2×1.2 mm2, ST = 1.2 mm (vC=3.6 cm s−1), Nscans=625, GRAPPA factor=5, BW=1666 Hz per pixel, TA/slice≅20 s, acquisition stepped through seven coronal slices (TA∼2.5′) covering the carotid arteries. Finally, to prove the feasibility of performing such measures at 3 Tesla, one subject (m, 20 years) performed the last MRI protocol on the carotid arteries at 7 Tesla as well as at 3 Tesla (3 Tesla parameters: same TR/TE/FA/Nscans/GRAPPA factor as at 7 Tesla, inplane voxel size = 1.7×1.7 mm2, ST=1.7 mm, BW=1786 Hz per pixel). In each voxel, magnitude EPI signals were converted to % signal changes by dividing the signal at each time-point by the mean signal across time; temporal drifts (third-order polynomials) were removed. To investigate which areas of the brain are mainly affected by cardiac and respiratory pressure waves, a second-order Fourier series was employed to model the effects in EPI time-courses related to the phase of the cardiac pulsation and respiration, respectively (four cardiac and four respiratory RETROICOR regressors [15]). The RETROICOR regressors, which are commonly used in fMRI [15,16], were adopted to map both the first and the second harmonics of pulsatility effects, as well as both in-phase and out-of-phase effects in our EPI data. The EPI signal variance explained (%) by cardiac and respiratory RETROICOR regressors was computed as the R2-value adjusted for the degrees of freedom, multiplied by 100 (as in [16]). To analyse respiratory effects in MRI data, EPI time-courses were low-pass filtered (cut-off frequency 0.6 Hz). To analyse cardiac effects, EPI time-courses were band-pass filtered between 0.9 and 1.4 Hz. For each voxel, and for each (cardiac or respiratory) peak detected in physiological recordings, we considered a window of N. wave cycles Nc=3 (e.g. approx. 3 s and approx. 13 s, respectively, for cardiac and respiratory effects) of the EPI time-course; we then averaged this window across temporally consecutive peaks to obtain an average cardiac and an average respiratory MRI waveform (the number of consecutive averaged peaks, Nav, was the maximum achievable in the TA). With the aim of computing the pVI on the carotid arteries, we manually segmented the carotid arteries, and from each average MRI waveform and each voxel of the carotid arteries, we obtained voxel-wise estimates of the signal at systole (S(tsystole)) and at diastole (S(tdiastole)) as the maximum and minimum signal of the average MRI waveform, respectively. We then computed the average of S(tsystole)/S(tdiastole) on an ROI, namely on a voxel-by-voxel basis (pVIvoxel, ROI=single voxel), across a vessel cross section (pVIcross section, ROI=vessel cross section) or across an entire segmented vessel (pVIsegment, ROI=entire vessel): 3.1 GE multi-echo EPI was performed on four subjects (1m/3f, age 26±3) using the following parameters: TR/TEs/FA=69 ms/[18 40] ms/90°, three coronal slices, inplane voxel size=1.5×1.5 mm2, ST=1.5 mm (vC=2.2 cm s−1), Nscans=300, GRAPPA factor=4, BW=2056 Hz per pixel, acquisition stepping through three spatially contiguous coronal slices covering the carotid arteries, with TA/slice≅20 s. To investigate the reproducibility of the results, we repeated the GE multi-echo acquisition for a different temporal resolution (and vC=0.7 cm s−1), with the acquisition parameters as above except for an approximately three times longer TR (TR=205 ms), Nscans=100 (to have a similar TA as the shorter TR acquisition) and three spatially contiguous slices covering the carotid arteries acquired with the traditional scheme (slice loop inside the repetition loop). For both acquisitions (shorter and longer TR), for each voxel and repetition, To study the dynamics of the pVI in response to changes in cerebrovascular viscoelastic properties, three subjects, who participated in experiment 3 (1m/2f, age 27±3), also performed a self-paced breath-holding task (normal breathing for approx. 30 s, then breath-hold for approx. 30 s, and then normal breathing again) during the acquisition of GE multi-echo EPI with the following parameters: TR/TEs/FA=69 ms/[18 40] ms/90°, inplane voxel size=1.5×1.5 mm2, ST=1.5 mm (vC=2.2 cm s−1), GRAPPA factor=4, Nscans=1350, one coronal slice covering the carotid arteries, with TA/slice≅93 s. From inspection of the respiratory recordings, we found that one subject did not correctly perform the task, and therefore the data of this subject were excluded from further analysis. For each voxel and repetition, In agreement with the theoretical predictions (figure 1), our results from experiment 1 (figure 2) for both EPI and CINE MRI showed an approximate sixfold signal gain when imaging non-stationary spins in regime 3 (flow velocity>vC, example provided for the carotid arteries) with an FA equal to 90° compared with an FA equal to the Ernst angle. Figure 2. On the optimal use of a 90° flip angle (FA) to image non-stationary spins in regime 3 (experiment 1). In (a), we show, for an example dataset, an EPI image (mean across time, TA≅20 s, vC=3.6 cm s−1) acquired with several FA. The dependence of the signal in the carotid artery (v≥vC, regime 3) with the FA, displayed in (b), shows good agreement with the predictions (equation (2.7), contrast increasing with sin(FA)). In particular, for subjects 1–3, the use of a FA=90° enabled a gain in signal of fast-flowing spins (v≥vC, regime 3) of the carotid artery equal respectively to 5.7, 6.7, 6.1 compared with FA=10°, in agreement with the expected theoretical gain equal to 6.2. Note that the use of a 90° FA is advantageous for EPI as well as for any image read-out scheme; to provide an example, we show in (c) a coronal slice (averaged across 22 time-points across the cardiac cycle) covering the carotid artery and acquired with the cardiac-gated CINE technique with 10° and 90° FA (TA=[49 52] s, respectively). The ratio of the average CINE-MRI signal measured in the carotid artery at 90° versus 10° was 5.93.
The areas of the brain mainly affected by cardiac and respiratory pulsatility are shown in figure 3a–d (experiment 2). These included: larger intracortical arteries (e.g. the carotid arteries, the anterior cerebral artery (ACA), the middle cerebral artery (MCA)), veins (notably the sagittal sinus), smaller intracortical (e.g. smaller sulcal arteries) and scalp vessels, CSF spaces (e.g. the lateral and third ventricles, the sulcal CSF), as well as—although to a smaller extent—the brain parenchyma (especially grey matter neighbouring cerebral vessels and the CSF). Examples of the raw MRI time-courses and of the MRI pulse waveform estimated in different compartments are also shown in figure 3e and f, respectively. Figure 3. Thespatial distribution of cardiac and respiratory pulsatility effects in the brain and the MRI pulse waveform (experiment 2). For an example dataset, in (a–d), we show the average EPI image across 600 time-points (TR= 51 ms) on the left column, and the variance explained by the cardiac and respiratory RETROICOR regressors on the middle and right columns, respectively. Note that in (a) we display an axial slice (out of 25 slices) of the acquired axial slab (TA∼12.5′) showing pulsatility effects in the lateral ventricles (top–red–arrow), scalp arteries (bottom left–blue–arrow), sagittal sinus (bottom middle–yellow–arrow) and brain parenchyma (bottom right–green–arrow); in (b,c), we show a reformatted sagittal and coronal view of the axial slab (note the good spatial coverage across slices, i.e. no visible head motion) displaying pulsatility effects in the ACA and the MCA, respectively (orange arrows); further, in (d), we show a coronal image of a coronal slab (TA∼2.5′) centred on the carotid arteries (orange arrows). The signal time-course (only 20 s) from a scalp vessel (bottom left–blue–arrow in (a)) and its amplitude spectrum are plotted in (e) (left and right panels, respectively): note the presence of respiratory (approx. 0.24 Hz), cardiac (approx. 0.98 Hz, and higher harmonics) components, and also of interaction terms (approx. 0.98±0.24 Hz, and higher harmonics), the latter probably related to the respiratory sinus arrhythmia. Finally, in (f), we show the average cardiac (left) and respiratory (right) MRI pulse waveforms in three voxels (blue dashed-dot line, scalp vessel; red dashed line, lateral ventricle; green solid line, grey matter, see arrows in (a)). The MRI pulse waveforms were used to estimate the proposed biomarker of compliance, the pulsatility volume index (pVI). (Online version in colour.)
The MRI pulse waveform was employed to compute a novel indicator of cerebrovascular compliance, the pVI, which, in this work, was mainly investigated and preliminarily validated for the carotid arteries. For an example dataset, in figure 4a–c, we show respectively the computed pVIvoxel (maximum intensity projection across slices), pVIcross section and pVIsegment (mean±s.e. across subjects) for the carotid arteries. Figure 4. The pulsatility volume index (pVI) of the carotid arteries (experiment 2). The cardiac pVI of the carotid arteries was evaluated from the cardiac MRI pulse waveform at different spatial scales: (a) at the voxel level (pVIvoxel, example dataset displayed), (b) across their cross section (pVIcross section, example dataset displayed, cross section defined perpendicular to the z-axis indicated by a white arrow) and (c) across each entire segmented artery (pVIsegment, mean s.e. across subjects). Note that in (b) the pVIcross section of the segmented left carotid arteries (displayed on the left part of the central panel) is shown on the left panel, and the pVIcross section of the segmented right carotid arteries (displayed on the right part of the central panel) is shown on the right panel. Interestingly, pVIvoxel was lower for voxels inside or containing the arterial lumen (note in (a) a darker stripe inside the arteries, indicated by three red arrows), and higher for voxelscovering the arterial wall, as expected for a blood volume effect. Further, pVI varied across the arterial axis (see pVIcross section in (b)), between arteries (see pVIsegment in (c), *p<0.02), and displayed a trend of left/right asymmetry. (Online version in colour.)
In the electronic supplementary material, we provide movies of the MRI pulse waveform in the brain (Nc=3, repeatedly displayed four times, for a total of 12 cardiac cycles displayed), in order to demonstrate specific pulsatility effects in several brain compartments such as the carotid arteries, ACA, MCA, scalp arteries, sagittal sinus, as well as CSF spaces (e.g. note that the CSF is being pushed out of the lateral ventricles into the third ventricle during the cardiac cycle). In order to characterize the pVI, we scrutinized the origin of the MRI signal changes related to the phase of the physiological cycles, specifically the presence of possible Figure 5. On the origin of the MRI pulse waveform, scrutinizing S0 and possible
In order to provide a preliminary validation of the pVI as a biomarker of cerebrovascular compliance, we investigated the dynamics of the pVI, underlying transient changes in cerebrovascular viscoelastic properties during a breath-holding challenge. As shown in figure 6, the pVI displayed a significant increase during breath-holding, and decreased a few seconds before and after breath-holding. Figure 6. ThepVI dynamics during a breath-holding challenge (experiment 4). The dynamics of pVIsegment of the carotid arteries during a breath-holding challenge are shown for two subjects and E1, E2, S0 and
Finally, in the electronic supplementary material, we show the feasibility of imaging pulsatility effects in the carotid arteries at 3 Tesla, with a preliminary comparison with 7 Tesla images, obtained in the same subject and with a similar MRI acquisition scheme. Brain–heart interactions occur through several mechanisms and along several pathways, including neuronal autonomic regulation and feedback, hormonal regulation and mechanical coupling. In this work, we focused on mechanical coupling, specifically the influence of cardiac as well as respiratory activity on the mechanical and viscoelastic properties of intracranial tissue. Perturbation of these properties has been implicated in the pathophysiology underlying cerebrovascular disease (e.g. atherosclerosis, stroke and aneurysm), which is the second leading cause of mortality worldwide, and may also be involved in neurodegenerative disorders, sleep disorders, as well as in migraine. The latter is a common neurological disease, which has often been associated with neurovascular disease and increased CSF pressure [19–21], and might thus be related to the impairment of brain–heart mechanical interactions. We designed and characterized a novel indicator of cerebrovascular and brain parenchymal compliance, the pVI, based on fast MRI of cardiac and pressure waves propagating in the brain. Here, we first discuss the benefits of the employed MRI sequence in terms of speed and sensitivity. We then discuss the spatial distribution in the brain of cardiac and respiratory effects and the possible mechanism underlying these two effects. Third, we discuss the pVI and its characterization by multi-echo imaging aimed at investigating the origin of the observed contrast, and its preliminary validation as a compliance indicator by the use of breath-holding challenges. Finally, we discuss the limitations of this work and possible future work. The speed of the EPI sequence implemented in this work to image pulsatility effects in the brain compares favourably against commonly used gated sequences such as CINE MRI (2.5 times faster in our experiments) and phase contrast imaging (approx. 7.5 times faster, results not shown) for the following reasons: first, it does not require gating, rather it exploits the use of simultaneous physiological recording; second, it does not seek to quantify bulk-flow velocity changes (which require the use of flow-encoding gradients in three different directions), instead it provides a signal primarily dependent on blood volume changes. We expect to further boost the speed of this EPI sequence by a factor of 3 by employing simultaneous multi-slice imaging [22], and to achieve an almost full brain coverage (9 cm thick slab) with an approximately 1 mm isotropic resolution in 10 min (e.g. considering a TA of 20 s per three simultaneously acquired slices). Our theoretical predictions and experimental results demonstrate an approximately sixfold gain in endogenous signal of fast flowing (regime 3) spins using a 90° FA compared with the Ernst angle (approx. 10° in our experiments, which instead maximizes the signal of stationary spins), as well as a concurrent significant background suppression. This work suggests that the use of exogenous contrast (e.g. gadolinium) is not needed to map cardiac and respiratory pulsatility effects for spins in regime 3 at 7 Tesla, and our preliminary results at 3 Tesla are also encouraging in this sense. For the critical speed employed in our experiments (between 2.2 and 3.6 cm s−1 depending on the experiment), we expect to have spins in regime 3 for larger arteries (spin velocity>100 cm s−1, e.g. carotid arteries), middle arteries (spin velocity approx. 50–100 cm s−1, e.g. middle cerebral arteries), arterioles (spin velocity approx. 5–10 cm s−1) and larger veins (spin velocity approx. 10 cm s−1, e.g. basal vein of Rosenthal). This will not be the case for capillaries (spin velocity approx. 0.5–2 mm s−1), nor the ventricular, sulcal and brain parenchymal interstitial CSF (CSF bulk flow velocity is in the range 1–4 mm s−1, with the maximum bulk flow velocity of 2 cm s−1 in the cerebral aqueduct [10]), whose spins will belong to regime 2. Cardiac and respiratory pulsatility MRI effects were mainly visible in arteries, veins and CSF spaces, as also observed in previous work [15]. Notably, in our data, we were able to observe respiratory pulsatility effects with high detail and sensitivity through the use of high spatial resolution (including the use of thin slices, which decreased the contribution of other respiratory effects, e.g. off-resonance effects in the brain owing to chest motion), an ultra-high-field scanner and optimized sequence parameters (e.g. the FA). Interestingly, the spatial distribution and amount of variance explained by cardiac pulsatility effects differed from those of respiratory effects, as the former were more visible and stronger in arteries, whereas the latter where more visible in the CSF. This might be explained in the light of the different mechanisms underlying cardiac and respiratory pulsatility. The propagation of the cardiac pressure waves occurs directly along the cerebral arterial tree. Instead, respiratory pulsatility is expected to originate mainly from the mechanical coupling of (thin-walled large neck and abdominal) veins with changes in the thoracic pressure during respiration [23,24] rather than of large (thick-walled and low compliant) arteries, as expected for cardiac pulsatility. In turn, the propagation of the respiratory pressure wave in vertebral veins might induce pulsatility mechanisms in the CSF in the spinal cord and retrogradely in CSF spaces in the brain. Therefore, in contrast to cardiac pulsatility, respiratory pulsatility might be more easily detectable in cerebral veins and the CSF than in arteries, and also might travel retrogradely in veins/spinal cord with respect to the direction of blood/CSF flow. Previous work with Doppler ultrasound methods [25] and gated CINE MRI methods [10] mainly focuses on measuring blood/CSF velocity during the cardiac cycle only; further, to probe brain tissue elasticity, other techniques (MR elastography) employ external mechanical vibrations at high frequency (range 25–100 Hz [26,27]). The possibility, demonstrated by our results, to map the effects of two different endogenous (cardiac and respiratory) pressure waves enables one to probe brain tissue compliance at two physiological modes of vibration, hence possibly providing complementary information on cerebrovascular and brain parenchymal compliance (for instance because of different compliance mechanisms at different frequencies, namely approx. 1 Hz and approx. 0.25 Hz for cardiac and respiratory effects, respectively). Our investigation of pulsatility effects was mainly conducted at 7 Tesla. Nevertheless, we also reported preliminary results at 3 Tesla, which demonstrate the feasibility of imaging pulsatility effects in the carotid arteries: future work will focus on optimizing the acquisition parameters (e.g. the TE) for 3 Tesla imaging, and on increasing the spatial resolution (in this paper, using a lower resolution at 3 Tesla when compared with 7 Tesla most probably caused a contrast decrease owing to partial volume effects). We evaluated the proposed novel biomarker of compliance, the pVI, in response to the cardiac pressure wave, in larger cerebral arteries in a small group of subjects. Interestingly, the pVI varied across each arterial segment and across arteries, displaying a trend towards a left/right asymmetry, and a significant increase between the external and the common carotid artery. Future work is planned to confirm these findings in a larger control group cohort and possibly examine the pVI in patient populations (e.g. high blood pressure or atherosclerosis). In addition, we sought to characterize the MRI signal changes underlying the pVI. Considering that, on the basis of our simulations (see Theory section), blood volume changes in response to pulse pressure waves were expected to produce S0 signal changes, in the carotid arteries, we investigated the occurrence of S0 effects related to the phase of the cardiac cycle by the use of multi-echo EPI; we also explored the possible presence (if any) of Finally, we employed a breath-holding task to perform a preliminary validation of the pVI, where pVI was shown to have utility as a compliance biomarker able to detect changes in cerebral vessel viscoelastic properties. In the carotid arteries, the pVI increased during breath-holding owing to the increased blood volume mediated by hypercapnia associated with breath-hold [30] (i.e. elevated CO2, associated with a measured decrease in the RVT regressor). Interestingly, the pVI transiently decreased before and after hypercapnia, and occurred with a measured small increase in the RVT, possibly indicating a hypocapnic adjustment performed by the subject to prepare for and compensate after the breath-holding challenge. Further, by inspection of multi-echo signals, we noted that the pVI changed dynamically in single-echo signals, as well as in S0 signals, demonstrating its positive performance as a compliance biomarker for both future single-echo and multi-echo experiments. A more marked change in pVI in second-echo rather than first-echo signals was detected, but this was not associated with significant changes in the pVI in In this work, we have shown the feasibility of mapping the cardiac and respiratory MRI pulse waveform in several brain compartments (large and middle arteries, large veins, CSF compartments, brain parenchyma), and focused the characterization and validation of the pVI on the carotid arteries and for cardiac pulsatility only. Future work will also extend these results to other brain compartments and for respiratory pulsatility. The fast MRI of physiological pulsatility proposed here suffers from artefacts common to EPI, e.g. spatial distortion, motion and ghosting. Future work will address the impact of these artefacts on the estimation of the MRI pulse waveform and of the pVI, and on possible correction strategies (for instance, the use of field maps or images acquired with a reversed phase-encoding direction to correct for spatial distortions as is customary for fMRI). Nevertheless, we expect the pVI to be less affected by artefacts at the ROI (e.g. pVIcross section, pVIsegment) than at the voxel level (e.g. pVIvoxel). Head motion will require special consideration: first, because the current two-dimensional acquisition scheme prevents a proper three-dimensional motion correction, this issue could be circumvented in the future by adopting pseudo-volumar acquisition, enabled by the use of simultaneous multi-slice imaging [22], or prospective motion correction methods; second, because the presence of local motion and deformations in high-contrast areas of the brain owing to pulsatility effects (e.g. in ‘bright’ larger arteries and CSF owing to blood volume and motion effects during the physiological cycles) are effects of interest and should not be removed by motion correction procedures. Our work demonstrates the feasibility of imaging the cardiac and respiratory pulsatility in the brain with high sensitivity and speed by the use of a fast EPI-based MRI sequence and of optimized sequence parameters, based on the solution of the Bloch equations for non-stationary spins. Our MRI sequence might prove useful in future work for fast imaging of pulsatility effects in the brain, to investigate stiffness of walls and tissue that bound flowing/moving fluids in the brain, as well as stagnancy of cerebrovascular fluids. We expect that our MRI sequence might also be employed for fast imaging of pulsatility effects in scalp vessels (neatly visible in our data), aimed at improving models of ballistocardiogram artefacts in simultaneous EEG–fMRI studies [31]. Lastly, we foresee the possible use of the pVI, preliminarily characterized and validated in this work, in future studies of brain–heart interactions, as a more direct indicator of blood volume changes (compared with previous measures of indirect blood velocity changes) underlying cerebrovascular compliance in cerebrovascular disease (e.g. larger and smaller vessel disease), ageing and other disorders (e.g. neurodegenerative disorders, including Alzheimer's disease and multiple sclerosis, sleep disorders and migraine). Data are made available in the electronic supplementary material as movies showing pulsatility effects in the brain over time. Drs Bianciardi, Toschi, Polimeni, Evans and Keil acquired the data. Data analysis was carried out by Dr Bianciardi. All the authors substantially contributed to the conception of this work and to data interpretation, drafted the article and approved the final version to be published. We have no competing interests. Dr Evans is currently an employee and stockholder of Biogen; his participation in the study and interpretation of the findings were not supported by nor have a financial interest of Biogen. This work was supported by these sources of funding: NIH NIBIB P41-RR014075 and R01-EB000790, NIH NIMH K23MH086619, and NIH NCRR S10-RR023043 and S10-RR023401. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 12The autonomic nervous system (ANS) owes its name to John Newport Langley [1], a British physiologist widely regarded as the father of autonomic neurophysiology. As later communicated by Langley himself, ‘it was a local autonomy that [he] had in mind. The word “Autonomic” does suggest a much greater degree of independence from the central nervous system that in fact exists’ [2]. Hence, since its discovery and despite its nomenclature, the ANS has never been considered or treated as completely independent from the central nervous system (CNS). More than a century after its discovery and initial description, there is still substantial debate about the nature as well as functioning of the control pathways through which the CNS interacts with the ANS. While functional magnetic resonance imaging (fMRI) based on the blood-oxygen-level dependent (BOLD) effect [3] allows access to non-invasive, spatially and temporally resolved functional information about brain activity at rest as well as during task execution, direct non-invasive in vivo recordings of ANS activity are not accessible. In this context, the analysis of heart rate variability (HRV) has emerged as a valid signal-processing approach to the estimation of ANS outflow, which also allows us to partially disentangle sympathetic and parasympathetic activity. In particular, band-specific spectral powers in the RR interval time series (i.e. the series of intervals between two consecutive ‘R’ peaks in the electrocardiogram (ECG)) are thought to provide a reliable static estimate of the parasympathetic (high frequency (HF) 0.15–0.40 Hz) activity, as well as mixed information about sympathetic and parasympathetic tones (low frequency (LF) 0.04–0.15 Hz) [4]. While both time- and frequency-domain RR analysis have been widely employed to assess the strength of these two efferent ANS pathways, only a few techniques (e.g. the pseudo Wigner–Ville transform [5] or wavelet decompositions [6,7]) have allowed a time-resolved estimation of ANS activity, which is mandatory to allow joint analysis with brain functional imaging. In a pioneering paper, Critchley and co-workers [8] jointly analysed the fMRI BOLD signal and HRV indices calculated from concurrently acquired ECG signals, demonstrating a relationship between the dorsal anterior cingulate cortex and sympathetic modulation. Furthermore, using a sliding window approach to RR series analysis, Chang et al. [9] presented the first joint fMRI/HRV analysis on a resting-state dataset, investigating the role of the amygdala and dorsal anterior cingulate cortex (which are involved in the central mediation of vigilance and emotional arousal) in modulating undirected brain–brain resting-state functional connectivity. In order to improve the reliability of time-resolved ANS activity estimation, we have proposed a novel recursive algorithm to estimate instantaneous heart rate (HR) and HRV through a probabilistic point-process (PP) framework [10–13]. Contrary to classical time–frequency decomposition approaches, this framework does not require interpolation of RR series, is inspired by a physiological integrate-and-fire model and provides an instantaneous probabilistic estimate of the occurrence of the next R peak at arbitrary time-resolution from which a number of linear and higher-order HRV-related estimates can be obtained [3,14–16]. The dynamical estimates of ANS activity provided by this framework have been employed in conjunction with BOLD fMRI to obtain in vivo, non-invasive information about the functional CNS correlates of the ANS. In this context, using a dynamic grip task Napadow et al. [17] demonstrated that fMRI BOLD activity in several cortical and subcortical brain regions exhibits significant correlation with cardiovagal activity, as assessed by the HF component of HRV. In a similar set-up, Sclocco et al. [18] demonstrated that the activity of a number of brain regions, such as the insula, regions belonging to the default mode network and those belonging to visual motion areas, were strongly associated with both sympathetic and parasympathetic activity during nauseogenic visual stimulation. This body of literature has helped to characterize a so-called central autonomic network (CAN) (see for instance [19]) whose existence was first postulated in several animal models [20,21]. In this context, in a large meta-analysis of recent neuroimaging studies, Thayer et al. [22] showed that cerebral blood flow as measured with fMRI or positron emission tomography (PET) is associated with HRV and identified a number of brain areas whose association with HRV is modulated by emotional or cognitive/motor tasks. In another recent meta-analysis, Beissner et al. [23] studied the CAN and its constituents using an activation-likelihood framework. They showed that the CAN involves some core constituents such as the amygdala, right anterior and left posterior insula and midcingulate cortices in a task-specific manner. In addition, the CAN also includes several other brain regions as components of the ANS regulatory network, such as the insula, ventromedial prefrontal cortex, hippocampal formation, mediodorsal thalamus, hypothalamus, posterior cingulate cortex, lateral temporal cortices and bilateral dorsal anterior insula. To date, all studies investigating the functional links between the CNS and ANS have employed concepts of undirected associations (e.g. correlations) and did not employ ultra-high-field technology. The directed, causal influences within the brain regions comprising the CAN as well as between the CAN and HRV-related metrics of ANS activity have not yet been investigated. Given that CAN regions interact tightly among themselves, the study of bivariate, undirected correlations may not be sufficient (or may introduce spurious results) when aiming to disentangle the central pathways of ANS modulation. Arguably the most influential contribution to the definition of causality was proposed by Clive W. J. Granger [24]. In brief, a dynamical system is considered to be Granger-causing (G-causing) if observations from the past of that dynamical system allow better predictions of the future of another dynamical system, compared with predictions based on the past of the latter alone. This general definition allowed detection of time-domain causality between time series with a certain degree of confidence [25]. Following the seminal work of Geweke [26], who proposed a measure of directed feedback between time series, the idea of Granger causality (GC) has been mostly associated with its implementation in the context of autoregressive (AR) models [27] and its expansion to multivariate systems through the use of vector AR (VAR) models [28]. Additionally, when GC is implemented in terms of AR and VAR models, it is also possible to model the influence of a variable over another one by including instantaneous (zero temporal lag) influences [15,29,30]. Along with the classical time-domain approach to GC, a number of spectral methods for estimating causality have been proposed in the recent literature. In one of the earliest developments of spectral GC, directed coherence (DC) was used to characterize brain causality between two electroencephalogram (EEG) signals [31]. The concept was further generalized to the multivariate case by Baccalá et al. [32], through the idea of multivariate DC, and in the seminal paper by Kaminski & Blinowska [33], where the concept of directed transfer function (DTF) was introduced. Many other spectral methods for GC estimation have been developed, including the partial directed coherence (PDC) [34], generalized PDC (gPDC) [35], information PDC (iPDC) and information DTF (iDTF) [36]. Baccalá et al. [37] presented a unified theoretical framework that encompasses all aforementioned spectral domain techniques. Interestingly, the theoretical development of spectral methods for studying causality has been fuelled by its potential application to the analysis of EEG signals, due to the robust body of biomedical literature relating band-specific EEG powers with distinct physiological functions and states [38]. Goebel et al. [39] and Harrison et al. [40] pioneered the use of VAR models for the analysis of fMRI data, and GC techniques are now routinely employed in fMRI analysis [41,42] (see also the review by Seth et al. [43]). While fMRI offers exceptional local specificity in quantifying dynamical brain activity (i.e. a spatial resolution or 3 mm or less, resulting in 104–105 signals for every fMRI acquisition), compared with EEG data fMRI signals are typically collected with low sampling frequency (usually 1 Hz or less), short data lengths (of the order of minutes or tens of minutes) and low signal-to-noise ratio (SNR) [44], possibly limiting the applicability of VAR-based GC methods. A valid workaround is to obtain locally averaged BOLD time series according to predefined brain atlases (see [45] for a review). Another very recent approach to overcome the dimensionality problem is to employ principal component analysis (PCA) or independent components analysis (ICA) prior to fitting the VAR models. The feasibility of the ICA-based reduction in the context of time-varying BOLD activity has been demonstrated in [46] and has been used to introduce the concept of large-scale GC (lsGC). Also, in the context of EEG analysis ICA allowed the estimation of large-scale, time-varying PDC [47,48]. A proof-of-concept application of GC after PCA in fMRI has been performed in [49], and in [50] lsGC was used to characterize network modules. A potentially limiting factor in applications of GC in fMRI is the very nature of the BOLD signal, which stems from a convolution product between the neuronal activation and an unknown haemodynamic response function (HRF) reflecting the local homeostatic adjustment of blood flow at the tissue level [51]. Characterization of the relationship between neuronal activation and the BOLD signal is far from complete (see [52] and [53] for reviews), and poses practical difficulties due to inter-subject variability [54], is tissue-dependent both in human [55] and in animal models [56], and may also be time-dependent [57] as well as task-dependent [58] (although inter-subject variability seems to dominate all other sources of fluctuations [59]). In this context, some authors have suggested addressing the HRF-related confound through a blind data-driven deconvolution method under the assumption that neural activation is sparse and of a binary nature [60]. This method has been employed in resting-state fMRI to characterize transfer entropy between PCA components (note that transfer entropy is equivalent to GC under the assumption of Gaussian noise [61]) in computing GC through a lagged version of structural equation modelling [62]. While the assumptions behind blind deconvolution may appear arbitrary, in general they have proved useful in repeated tasks paradigms and when the neural activation can be described by a purely threshold-trigger mechanism [63]. In an effort to investigate the true influence of the HRF-related confounds, recent papers employed direct intracranial measurements of neural activity to explicitly investigate the effect of HRF convolution on GC estimates in vivo [64], at the population level [46] and with synthetic data simulations [65] using analytically solvable VAR models as well as VAR models with non-trivial topologies and realistic haemodynamics. They showed that GC is ‘extremely resilient’ against HRF convolution even when local variations in HRF parameters strongly degrade information about time precedence between neural signals. Instead, GC estimates are sensitive to signal down-sampling in the time domain and to low SNR. With respect to the problem of time resolution (which may be the most important limiting factor in GC applications of fMRI) Hemmelmann et al. [66] employed both simulated (through the balloon model [67]) and real BOLD signals to demonstrate that combining the complementary information provided by time-domain GC and frequency-domain PDC allows correct detection of GC in spite of HRF-related confounds. It should be noted that, when using causality techniques to analyse the directed interaction between brain-derived BOLD time series and external signals, an alternative approach is the convolution of the external signal with the best assumed HRF. While the problem of causality detection in the brain is clearly multivariate, in line with most classical, correlation-based connectivity studies the use of bivariate frameworks dominates the recent literature on directed intra-brain interactions. However, when quantifying brain–brain interactions, the high inter-dependence of brain-derived signals can lead to the appearance of false positives (i.e. spurious, unrealistic causal links). In this context, the importance of ‘conditioning’ GC between two brain regions (i.e. estimating their ‘true’ directed interaction while accounting for indirect influences mediated by additional brain regions) has been proposed by Guo et al. [68], and these also demonstrate the analogy between conditioned GC and the ideas behind partial correlation. Also Marinazzo et al. [69] defined a partial conditioning approach based on a priori, mutual information-based identification of the subset of signals which share the most information with the target signal under study, and demonstrated its application to the brain. While in [69] a prior downselection within conditioning variables was deemed necessary due to the limited length of typical fMRI time series, in GC analyses of the brain the widest possible set of conditioning variables (i.e. brain regions) should be included within a well-defined multivariate conditioning framework [28,41–43,65]. This is equally relevant when studying the causal links between multiple brain regions and external signals, like time-varying estimates of ANS activity. In this paper, we provide the first causality-based investigation of directed brain–heart interactions in the resting state and associate our results with estimates of directed, conditioned brain–brain resting-state networks. In order to mitigate fMRI confounds related to sensitivity, temporal resolution and physiological noise, we employed state-of-the-art, ultra-high-field fMRI combined with simultaneous multi-slice (SMS) echo-planar imaging (EPI) techniques and physiological signals acquisition. We employed time-varying estimates of heartbeat dynamics and ANS outflow obtained through our state-of-the-art probabilistic PP framework and investigated how proper conditioning of GC estimates, which accounts for the intra-subject interdependence between local, fMRI-derived brain activity across the whole brain, affected inference about directed brain–brain and directed brain–heart interactions. Estimating GC from the jth variable to the ith variable (j→i) amounts to testing the null hypothesis that knowledge of past values of j improves the prediction of (or restricts the uncertainty about) the future of i. In AR-model-based bivariate GC, we consider two models. First, the so-called restricted (order P)-AR model of the ith variable, which includes only the past values of i itself. Second, the so-called unrestricted AR model, which includes both variables i and j. The unrestricted model has order P with respect to the ith variable, and order Q with respect to the jth variable, 2.1 and2.2 where a′p and ak,p are the AR model parameters, and εt and ε′t are uncorrelated white noise processes. The restricted model represents the null hypothesis (Q=0) that should be rejected if j significantly G-causes i. The unrestricted model is the alternative hypothesis that j significantly G-causes i. Prior to model identification, optimal AR orders P and Q have to be inferred. As far as we are aware, all applications of GC to brain signals assume identical AR orders P=Q. However, given that P can only be chosen based on merit, quantified through some sort of information criterion (e.g. final prediction error of the Akaike [70], the Bayesian information criterion (BIC) [71] or the Pagano and Hartley t-test [72]) of the restricted model (and should not be varied in the unrestricted model), Q should be chosen independently while adopting the same figure of merit which has been adopted for the restricted model. Information criteria, rules of thumb and extensive searches are all strategies that have been employed in the past. One simple approach entails conducting a search of all possible Q>0 (up to a reasonable limit) to be employed as the alternative hypothesis, and Geweke [25] suggests that the number of lags Q of the independent variable should be kept smaller than the number of lags P of the dependent variable.We now extend the bivariate GC model (equations (2.1) and (2.2)) to the so-called ‘conditioned’ GC approach. If the causing variable xj is highly correlated with a set of variables (x1…xL) in order to correctly disentangle (j→i) in both the restricted and unrestricted model the influence of variables (x1…xL) should be included (i.e. the causality estimate is ‘conditioned’ to exclude the influence of the latter variables); the restricted and unrestricted conditioned models now read 2.3 and2.4 As before, given P and R as the optimal regression orders for the restricted model, the order Q is to be chosen independently.In order to accept or reject the null hypothesis, in either approach (bivariate or conditioned), the residual sum of squares (RSS) is used to compute the so-called fratio which will follow an F-distribution with Q and (Nobs−P−LR−Q) degrees of freedom [73,74], 2.5 where Nobs is the number is observations of xi, and RSSr and RSSur refer to the restricted and unrestricted models, respectively. We can then reject the null hypothesis and assume that a significant GC relation (j→i) exists if the observed fratio corresponds to a p-value defined as2.6 which is less than 0.05 [74]. This means that the reduction in the RSS resulting from the inclusion of j in the prediction of i is significantly higher than would be expected from including a non-G-causing variable in the prediction of i.Nine healthy volunteers (age 28±3 years) underwent 7 T MRI with simultaneous physiological signal acquisition under institutional review board approval at the Athinoula A. Martinos Center for Biomedical Imaging (MGH) on a Siemens 7 T whole-body scanner (Siemens Healthcare, Erlangen, Germany). We used a single-shot 2D EPI readout for 1.8 mm isotropic T2*-weighted axial images, with matrix size/generalized autocalibrating partially parallel acquisitions (GRAPPA) factor/nominal echo-spacing=136×136/2/0.57 ms and additional parameters echo time/flip angle/repetitions=26 ms/40°/300. The use of an SMS factor equal to 3 enabled whole-brain coverage (number of slices=75) and a repetition time (TR) of 1.5 s which minimized aliasing in ANS frequency bands (see below). With TR=1.5 s, the sampling frequency of the brain signal was 0.667 Hz. Physiological noise correction (high-pass filtering at 0.01 Hz and removal of second-order RETROspective Image CORrection [75] regressors) was applied using custom MATLAB scripts, while slice timing, motion correction and co-registration to Montreal Neurological Institute (MNI) space were applied to fMRI data using the Oxford Centre for Functional MRI of the Brain (FMRIB) Software Library (FSL). Successively, the average BOLD signal was extracted in 116 regions of interest (ROIs) using the automated anatomical labeling atlas [76] and a 117th region relative to the brainstem was extracted using the Harvard–Oxford subcortical atlas (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Atlases). Cardiac pulsation was recorded by a piezoelectric finger pulse sensor and respiration by a piezoelectric respiratory bellow positioned around the subject's chest (1 kHz sampling). The cardiac pulsation signal was used to detect cardiac peaks from which a peak-to-peak interval series (henceforth called the RR series) was derived and modelled as a PP [11], hence circumventing the need for RR interpolation. Briefly, assuming history dependence, the probability distribution of the waiting time t−uj until the next R-wave event ui−1 follows an inverse Gaussian model: 3.1 with the instantaneous mean RR interval defined as:3.2 whereFigure 1. Example of the fMRI BOLD ROI-wise signal (blue solid line) and HRV regressors HF (parasympathetic), LF (sympathetic+parasympathetic) and Bal (sympathovagal balance)(yellow, green and red solid lines, respectively) for one healthy volunteer. (Online version in colour.)
We computed both bivariate and globally (i.e. to all other brain regions) conditioned within-brain (i.e. brain–brain) causality between locally averaged brain signals, using equations (2.1) and (2.2) and equations (2.3) and (2.4), respectively, For the bivariate GC model, the AR orders were Pb=5 and Pc=5 for both the causing variable and the caused variable, corresponding to the minimum of the BIC [71] after independent optimizations as described above. For the globally conditioned GC (GCGC) model As above, we computed both bivariate and globally (i.e. to all other brain regions) conditioned brain–heart causality between locally averaged brain signals and the PP-derived dynamical band-specific powers (henceforth called regressors). To estimate bivariate GC between a brain region xi and a heart regressor Hsig (where ‘sig’ labels HF, LF and Bal) an AR model of order Pb is adopted for the brain signal, and an AR model of order Ph is adopted for the heart regressor. The AR order for the heart regressor was Ph=6, corresponding to the minimum of the BIC the restricted model. The AR order for the brain signal was Pb=5, corresponding to the minimum of the BIC for the unrestricted model given the order of the BIC for the restricted model. When evaluating the GCGC (xj→Hsig) between a brain region xj and a heart regressor Hsig both the restricted and unrestricted model (equations (2.3) and (2.4)) included the influence of all other brain variables, i.e. the restricted and unrestricted models are 3.3 and3.4 where Ck,p, C′k,p, Dk,p, D′k,p, are VAR model parameters and ηt and η′t are uncorrelated white noise processes. In the GCGC the AR model for the heart regressor was Ph=6. The AR order Pm for conditioning brain variables was set to Pm=1, corresponding to the minimum value of the BIC with respect to Hsig,ur(t). The AR model order Pq for causing variable xj was set to Pg=3 to minimize the BIC with respect to Hsig,ur(t).For each subject, both bivariate GC and GCGC between each of the 117 brain regions and every other region were computed. Additionally, bivariate GC and GCGC between each of the 117 brain regions and each heart regressor (HF, LF and Bal) were computed. For each directed connection between any two brain regions, for each connection between each brain region and every ANS regressor, and for both bivariate GC and GCGC the appropriate fratio was used to evaluate the p-value of the connection. Successively, to provide a group-wise count for each connection the number of subjects with p<0.01 and p<0.05 was counted. A measure of relative connection strength 3.5 and used for visualization purposes. All processing was performed through custom-build code in Mathematica 10.Figures 2 and 3 show the bivariate GC strengths Figure 2. Bivariate GC strength
Figure 3. Group-wise count of significant connections in bivariate GC. For each directed connection the group-wise count was computed as the number of subjects in whom that particular connection was statistically significant with p<0.01 (a) and p<0.05 (b). Rows label brain regions (GC sources) G-causing the brain regions labelled by columns (GC target). Brain regions are ordered according to Salvador et al. [79]. (Online version in colour.)
Figure 4. GCGC strengths in within-brain interactions. All nine subjects are shown. In rows and columns, respectively, causing and caused brain regions are shown. Brain regions are ordered according to Salvador et al. [79]. The colour scale shows connection strength (see equation (3.5)). (Online version in colour.)
Figure 5. Group-wise count of significant connections in GCGC. For each directed connection the number of subjects where that particular connection is statistically significant with p-value<0.01 (a) and p-value<0.05 (b) is shown. Rows label brain regions (GC sources) G-causing the brain regions labelled by columns (GC target). Brain regions are ordered according to Salvador et al. [79]. (Online version in colour.)
Figure 6. Group-wise directed resting-state network graphs as inferred with bivariate GC (a) and GCGC (b). The matrices containing the number of subjects in which each connection was statistically significant (p<0.01), i.e. figures 3 and 5, respectively, were thresholded at six subjects. Colour scale, average (across subjects) GC strength. Refer to Salvador et al. [79] forlong label names. In order to project onto anatomical space, each node of the network was positioned on the (x,y) coordinates of the centre of mass of the corresponding ROI as defined by the automated anatomical labelling atlas [76]. (Online version in colour.)
The results of studying directed brain–heart interactions using GCGC are summarized in table 1, while the results of the same analysis using the simple bivariate approach are presented in the electronic supplementary material, table S2. For each of the ANS regressors (HF, LF, Bal), the brain regions where the number of subjects in which GCGC is statistically significant (p<0.05) is shown.
Figure 7 shows the results of combining globally conditioned within-brain and brain–heart networks (i.e. condensing the data shown in figure 6b and in table 1). Brain–brain networks are taken out of anatomical space in order to improve legibility. Figure 7. Group-wise directed resting state brain–brain (cut-off at p-value<0.01) and brain–heart network graphs as inferred using the GCGC approach (right). Within-brain networks are computed and presented as in figure 6a (see the electronic supplementary material for explicit brain region names; the order of the brain region follows a frontal–caudal direction, see Salvador et al. [79]) exceptfor the spatial distribution. In this figure, brain nodes are arranged according to a spring embedding (i.e. vertices are placed so that they minimize mechanical energy when each edge corresponds to a spring). Significant causal brain–heart connections have been added according to table 1. (Online version in colour.)
To our knowledge this is the first investigation of causal brain correlates of ANS outflow. In this study, we analysed GC-based brain–brain interactions through fMRI BOLD signals, and GC-based brain–heart interactions through combined evaluation of BOLD fMRI signals and ANS-related signals. The latter were obtained by use of a PP approach, which models the waiting time between two consecutive heart beats as an inverse Gaussian distribution and represents the parameters of the distribution through an AR system. One major advantage of the PP model is that there is no need to align the heart rate signal with the sampling rate of the BOLD signal, hence avoiding any type of interpolation which would be necessary when employing any other type of time–frequency decomposition. To this end, it has been shown that resampling after classical interpolation significantly degrades the accuracy of model fitting [11]. In addition, it should be noted that the instantaneous PP model for the heartbeat incorporates physiological knowledge about the processes which underlie and modulate the heartbeat. Specifically, the model caters for the threshold behaviour of the heart's endogenous pacemaker, the history of autonomic inputs to the sinoatrial node as well as their time-dependency. Brain signals were locally averaged according to a ROI-based approach, while heart-related regressors were computed as described in [11,17,18] and resulted in three signals: HF—an estimate of parasympathetic activity; LF—an estimate of both sympathetic and parasympathetic activity; Bal=LF/HF—an estimate of sympathovagal balance. Brain–brain GC analyses were conducted in two different ways: a ‘classical’ bivariate approach, building monovariate and bivariate AR models for the restricted and unrestricted models, respectively, and a GCGC approach, consisting of a VAR system encompassing the whole state space. Each considered link (either between any two brain regions or between any brain region and an ANS regressor) was considered causal if it was found significant (according to its p-value) in a sufficient number of subjects. While correlation-based (i.e. non-causal) fMRI studies at 2 T [8] and 3 T [8,17,18] and longer (approx. 3 s) TR have provided important insight into the human brain correlates of ANS modulation, this study has tackled for the first time the problem of causal link estimation between brain activity and ANS outflow. In the brain–brain part of our analysis, within-brain bivariate GC analysis results in a great number of significant directed edges (figures 3 and 6a), with challenging physiological interpretation. Instead, using within-brain GCGC analysis results in a reduced number of significant directed edges (see figures 5 and 6b) whose physiological interpretation could be tackled based on current knowledge about correlation-based resting-state networks. The resulting graph spans most brain regions while retaining a certain degree of simplicity; given the high redundancy of brain signals even in the resting state, it is plausible that the high percentage of connections which are removed through the use of the globally conditioned approach correspond to spurious/indirect causalities which should not be interpreted as direct within-brain causal influences. In this study, using superior spatial and temporal resolution enabled by 7 T imaging (which we used in order to improve the reliability of our GC estimates), we demonstrated the existence of significant causal links between cortical as well as subcortical brain regions and ANS outflow for all three heart-related signals. While our findings cannot be directly compared with correlation-based studies, the results of the GCGC approach to studying brain–heart networks appear partially consistent with previous studies. In particular, the amygdala is seen to play a causal role (significant causal connection in at least four subjects) in the modulation of HF power (parasympathetic activity). This finding might seem counterintuitive, since the amygdala might be expected to be associated with increased heart rate, given its involvement in regulating the ‘fear’ response [80]; however, it should be noted that an increased heart rate might be induced by a reduction in the parasympathetic tone and, if the amygdala plays a role in this mechanism, this would justify the results of the amygdala having causal effects in the modulation of HF power (parasympathetic activity). This argument is consistent with a number of studies conducted in the LeDoux laboratory [81] which proposed that the amygdala may serve as a rapid (HF-related) emotional appraisal acting as a detector of potential threats and mediator of adaptive ‘fear’ responses. There is also consistent and robust evidence from animal and human research that the amygdala plays a key role in generating a series of adaptive ‘fear’ responses (e.g. fight, flight, freeze responses) which are all critically mediated via the ANS [22,23]. Our data may also partially fit with animal models of predator defence that have shown that intermediate levels of threat (e.g. when the most adaptive response would be to freeze) may cause heart rate deceleration or fear-induced bradycardia. Studies in rodents have demonstrated that this freezing behaviour and the associated increased parasympatethic tone depend on the amgydala to periacqueductal grey (PAG) projections [82]. A recent study in healthy volunteers [83] has also shown that the same neural circuits (i.e. amgydala–PAG connections) may be critically involved in humans as well to mediate fear bradycardia. Thayer et al. [22] showed how areas belonging to the prefrontal cortex (the medial and ventromedial prefrontal cortex in particular) play an inhibitory role on the amygdala, and this might explain the causal link of frontal brain areas to both HF and BAL regressors; given that our results suggest that the activity of such brain areas significantly affects sympathovagal balance, it is possible that the superior frontal gyrus (left and right) and the medial frontal cortex (left and right) play a key role in such modulation, thus indicating that the modulation of sympathovagal balance is possibly intertwined with self-awareness functions [84]. Additionally, both sympathetic and parasympathetic activities are modulated by activity in cerebellar regions, in the vermis, in the hippocampus, in parahippocampal regions and crucially in the brainstem. The brainstem contains several grey matter nuclei involved in the control of autonomic functions, including the periaqueductal grey, parabrachial complex, solitary nucleus and the dorsal motor nucleus of vagus [85]. The causal relations between the cerebellum and parasympathetic outflow are in accordance with previous studies [23]. In general, our analysis does not show a large number of brain regions directly involved with sympathetic outflow. This is compatible with the idea that, as also stated in [18], the brain regions correlated with sympathetic versus parasympathetic response differ substantially, providing grounds to posit the existence of two distinct CANs with complementary roles. Also, while finding regional concordance in around four subjects (with a maximum of six in one brain region) out of nine may point towards low statistical power, it should be noted that this does not affect the validity of the results, which were detected as significant. In this paper, we chose to use GC in the time domain as opposed to choosing one of the several frequency-domain approaches to GC, such as multivariate DC [32], DTF [33], PDC [34], gPDC [35], iPDC and iDTF [36]. Given that we used AR models containing both brain-related BOLD fMRI signals [36] (where the full bandwidth of the preprocessed signal was considered) and PP-derived ANS regressors, which represent instantaneous band-specific powers, a frequency-domain GC approach would probably have had limited advantage. Furthermore, it would have suffered from the differences in frequency domains between the two (brain and heart related) sets of signals. Additionally, the use of a frequency-domain method would have required interpolation of the unevenly sampled RR interval series, possibly introducing bias in the resulting estimates. Moreover, while estimates of parasympathetic and sympathetic activity could be readily extracted in the frequency domain, a definition of sympathovagal balance (which is considered a physiologically reliable indicator) would have been problematic in the frequency domain. Also, all VAR-based GC methods rely on a priori selection of the model lag length (i.e. the amount of past information to include in the model for the present observation), which, in most VAR applications, coincides with model order selection or, more specifically, the choice of AR order. While this choice can have a significant impact on the estimation of GC, and a number of criteria exist for this choice [86] (also see [27] for an extensive discussion), no general consensus exists on how to estimate AR orders in causality studies. In this context it should be noted that, while the development of algorithms for VAR parameter estimation has coaxed most published studies towards employing identical orders for the caused and for the causing variable, there is no theoretical reason for this choice, which may in fact contradict the basic concept underlying the idea of GC (i.e. that, given the restricted model, one should find the best possible unrestricted model based on some figure of merit). In this study, we, therefore, moved away from this common practice by performing independent, sequential order selection for the caused, causing and conditioning variables. Recent evidence points towards the viability of GC analysis in fMRI, provided the data are acquired under a high SNR and high temporal-resolution regime. In this exploratory contribution we chose a conservative, atlas-based approach to spatially averaging brain activity. While it is well known that the mutual influence of brain regions (i.e. ‘connectivity’) plays a crucial role in physiological functioning of the brain, rendering a conditioning approach necessary, a putative mutual influence between HF, LF and Bal would have an unclear physiological basis. We therefore chose to exclude the reverse causality direction from this study. Also, while we cannot exclude that our PP-based approach may introduce some extent of autocorrelation in our time-varying spectral estimates, the latter estimates are band-specific, while the possible autocorrelation would most probably be in frequencies lower than the lower bound in LF. In this sense, the PP model is not akin to more classical strategies used to obtain time-varying estimates (e.g. time–frequency decompositions) or dynamical inference approaches [87,88] nor is it similar to transfer-function-based methods [89] or directly related to mechanistic model-based approaches to analyse ANS activity [90] or the cardiovascular system, in general [91]. Our model uses a relatively large number of parameters (around 120) with respect to the number of time points available, possibly running the risk of overparametrization. In this paper, the difference in the degrees of freedom between the restricted and the unrestricted model (five in this study) is low with respect to the difference between the number of points and the degrees of freedom of the unrestricted model (which in our case is 300–127=173). Therefore, indices calculated by comparing the two VAR residuals exclusively (as opposed to the parameters) could be less affected by overparametrization. Still, the relationship between differences in model orders and the quality of GC estimates remains to be investigated in dedicated future studies involving synthetic data simulations. In this context, in future investigations possible countermeasures could also include assuming a sparse transition matrix and fitting the VAR model using a least-squares approach with a least absolute shrinkage and selection operator (LASSO)-type penalty [92–96]. Also, in this paper we have chosen not to include instantaneous causality effects in our model. While most strategies which have been proposed to include an instantaneous effect in GC involve some degree of a priori knowledge about the direction of such an effect, it has been shown that the presence of instantaneous effects, whose presence cannot be excluded in the present investigation, can have detrimental effects on GC estimates [97]. Future work should aim to use those methods which tackle the issue of instantaneous GC in an assumption-free framework [29,98,99]. While including a dimensionality reduction step such as ICA or PCA [47–50] may add robustness to our globally conditioned brain–heart networks, the dimensionality reduction step (which typically estimated hundreds of components) is itself susceptible to sample size. Given the underlying nonlinear nature of interactions which occur in physiological networks, nonlinear GC techniques or kernel-based GC approaches [100] could be implemented within a conditioning approach. While these techniques may be able to capture more subtle aspects of directed brain–brain or brain–heart interactions, their implementation would pose much more severe constraints on data length and possibly temporal resolution. Also, GC techniques defined in the frequency domain (PDC [34], gPDC [35], iPDC and iDTF [36]) may offer complementary information when combined with time-domain GC in situations of low SNR or poor time resolution [66]. It should be noted that the GCGC approach is not to be limited to the study of healthy subjects or of the resting state. Future developments therefore include the study of task-dependent changes of brain–brain and brain–heart causal connectivity as well as their changes in ANS-related stimuli (e.g. pain [101]) or disease, where the study of causal brain–brain and brain–heart links could prove a sensitive instrument both in neurodegenerative disorder and in other disorders involving ANS impairment such as Parkinson's disease [102–104], epilepsy [7,105], multiple system atrophy or certain forms of mild cognitive impairment [106,107]. While the limited number of subjects employed in this study necessarily warrants some caution in casting a strong physiological hypothesis based on our globally conditioned networks, in this feasibility study we have shown that 7 T functional imaging coupled with GCGC estimates is able to quantify directed brain–heart interactions, which can be interpreted in terms of central modulation of ANS outflow while correctly disentangling the high redundancy between locally aggregated brain signals. A.D. designed the study, performed the data analysis and wrote the manuscript. M.B. acquired, preprocessed the data and drafted the manuscript. L.P. drafted the manuscript and interpreted the results. L.L.W. acquired the data. M.G. and R.B. helped to draft the manuscript. N.T. conceived and designed the study, acquired the data, drafted the manuscript and preprocessed the data. All authors read and approved the manuscript. We declare we have no competing interests. This work was supported by the Italian Ministry of University and Research (MIUR), grant no. RBFR08VABD (awarded to N.T.) and NIH NIBIB P41-RR014075. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 13An increasing number of functional magnetic resonance imaging (fMRI) studies are examining not only the task-related increases in the fMRI signal within specific brain areas, but also the correlation of time series fluctuations between different brain regions. These studies are driven by the observation that low-frequency (less than 0.1 Hz) fluctuations in the fMRI signal intensity time series, which can take place either as a result of task-induced signal modulations or in the absence of an external stimulus or explicit task (resting state), are frequently correlated between functionally related areas. The general hypothesis is that these correlated fluctuations reveal synchronized variations in the neuronal activity of a network of regions. fMRI could then provide a window into the interaction, or connection, among brain areas. The exploration of these networks by analysing coherent signal fluctuations has therefore become known as functional or effective connectivity analysis [1,2], whereby the latter term implies that causal interactions are sought [2]. In particular, the recent discovery that spontaneous brain activity is not random noise, but is specifically organized in resting-state networks (RSNs) has generated considerable interest, as RSNs provide a way to probe brain function without needing explicit tasks to drive brain activity [3]. Specifically, consistent correlations among low-frequency fluctuations between the fMRI time series corresponding to different areas at rest were initially revealed by Biswal et al. [4]. Subsequent studies identified numerous consistent and distinct RSNs, including motor, auditory, visual and attention networks [5,6], as well as the presence of RSNs using additional modalities (electroencephalography—EEG and magnetoencephalography—MEG) [7,8]. Correlated resting-state fluctuations have been also identified between regions of the default mode network (DMN)—a set of brain areas that regularly deactivate through a wide range of cognitive tasks [9]. This supports the view that the brain at rest comprises a number of periodically active and synchronized networks. However, the subject test–retest reliability of RSN connectivity is lower than what would be expected if patterns were solely anatomically determined [10]. This is partly owing to the reconfiguration of functional networks, but could also reflect time-variability and the effect of non-neuronal sources. In this context, RSN non-stationarities are being increasingly considered, mainly using sliding-window techniques combined with seed voxel/region analysis and independent component analysis (ICA) [8,11–14]. Recent studies have focused on dynamic fMRI-based RSN functional connectivity [11,13,15]. Furthermore, the blood oxygen level-dependent (BOLD) signal measured by fMRI is only indirectly related to the underlying neuronal activity, as determined by neurovascular coupling mechanisms and quantified by the hemodynamic response function (HRF) [16]. The effect of non-neuronal physiological variability (e.g. heart rate, arterial CO2, respiration) on fMRI measurements and connectivity has been well established [17,18]. The cerebral vasculature is exquisitely sensitive to arterial CO2 fluctuations; it has been shown that spontaneous arterial CO2 fluctuations affect both cerebral blood flow in the middle cerebral artery [19] and the fMRI BOLD signal regionally [17,20]. Additionally, slow changes in respiration depth and rate were found to significantly affect resting-state functional connectivity of the DMN [21] and HRV, as well as its low- and high-frequency power components, have been shown to be correlated to regional BOLD signal variations [18]. RSN studies are particularly susceptible to physiological noise, as they are based on a single experimental trial [22,23]. No previous studies have examined the effect of physiological factors on dynamic (time-varying) resting-state functional connectivity to our knowledge. Therefore, the aim of this study is to determine whether dynamic resting functional connectivity is modulated by the time-varying properties of simultaneously recorded physiological signals by using end-tidal CO2 (PETCO2) and HR measurements obtained during scanning. In order to do so, we assess dynamic functional connectivity and its corresponding network degree for the DMN, visual and somatosensory RSNs and quantify its relation to time-varying physiological signal power. The results reveal a modulatory effect of the latter on the obtained dynamic functional connectivity patterns for both CO2 and HR. By using the discrete wavelet decomposition, we also show that this effect is more pronounced in specific frequency bands. Finally, we show that the observed modulation effects were considerably clearer when processing was done in the individual functional space, compared to when it was done in standard (Montreal Neurological Institute, MNI) space. Twelve healthy volunteers (seven male, aged 29.2±4.6 years) underwent resting-state fMRI at the Cardiff University Brain Imaging Centre. Imaging was performed using a 3 T scanner (General Electric). Head movement was minimized with a bite bar. For each subject, a T1-weighted FSGPR structural scan (256×256×172 voxels of 1×1×1 mm) was acquired and used to assist in placing individual subjects' data into a common stereotactic space. During the resting-state scan, the subjects were instructed to keep their eyes closed and remain awake. The duration of the experiment was 630 s, corresponding to 210 time points (gradient echo EPI sequence, TR=3 s, TE=35 ms, FOV/slice=20.5 cm/3.2 mm, flip=90°, 53 slices with 91×109×91 voxels of 3.2×3.2×3.2 mm). PETCO2 was recorded using a nasal cannula attached to a gas analyser (AEI Technologies). The HR and HRV signals were extracted from photoplethysmography waveforms obtained during the scan by the built-in scanner pulse oximeter. Specifically, before proceeding to compute these signals, we checked the raw signal quality using a signal quality index (SQI) which labels every 10 s of data as ‘reliable’ or ‘unreliable’, based on a set of physiologically relevant checks, followed by QRS template matching [24]. For the subsequent computation and analysis, we used only time series which were labelled as ‘reliable’ for more than 80% of the time. Subsequently, QRS detection was performed using the Hamilton & Tompkins [25] algorithm. The HRV signal at each beat was defined as the periods between consecutive R-peaks, whereas the HR signal was defined as the inverse of these periods multiplied by 60—in units of beats per minute. The final HR and HRV signals were obtained by applying spline interpolation at a frequency of 4 Hz and subsequently filtering with a median filter in order to remove spurious spike artefacts. Note that the results obtained by the HR and HRV signals were found to be almost identical; therefore, we present only results obtained using HR. We implemented a processing pipeline that was developed using FSL [26] and in-house-developed Matlab scripts. Data pre-processing steps included rigid body motion correction and skull removal for functional images using BET. For each of the subjects, motion parameters were calculated to reflect head motion in six directions with respect to the mean image. The resting-state functional images were normalized into MNI standard space using linear registration implemented using FSL-FLIRT [26]. We implemented two different versions of seed-based resting-state analysis. According to the first, we used anatomical masks defined in the MNI space (FSL atlas) in order to define regions of interest (ROIs). The following grey matter ROIs, which define the DMN, were defined: medial prefrontal cortex (mPFC), posterior cingulate cortex (PCC), anterior cingulate cortex (ACC), thalamus and precuneus. The average BOLD fMRI time series in each region were extracted and correlated with each other in order to obtain a functional connectivity matrix in sliding windows of 150 s duration (50 time lags), overlapped by 30 s (10 time lags). In order to select the window length, we obtained correlation matrices for different window lengths. Specifically, the cross-correlation between two time-series is estimated from finite data samples using the following sum The power spectral density of the PETCO2 and HR signals was calculated in the same sliding windows by using the Welch method [27]. To quantify correlation between different ROIs, we calculated both the absolute maximum and the absolute average cross-correlation values between 0 and ±5 time lags. In addition to performing these calculations in standard MNI space, we repeated the procedure in the functional space of each subject in order to examine the effects of registration (of the BOLD time series to the MNI space) on the results. To do so, we registered the ROI masks originally defined in MNI space to the functional space of each subject (FLIRT). After obtaining correlation values between all areas in the examined RSNs, the overall time-varying connectivity of the network was quantified by computing the graded network degree (i.e. we did not convert brain networks to binary graphs) [28] for all 150 s windows. Specifically, the graded degree is defined as We performed standard seed-based analysis by first selecting a seed voxel, correlating all voxels of the brain to this voxel and applying a threshold in order to define the spatial pattern of the corresponding RSN. After examining different threshold values, we present results for a threshold value of 0.65, as they yielded spatial RSN patterns relatively similar to the mask-based analysis, whereby similarity is defined as spatial overlap between the regions defined by mask-based and seed-based analysis (i.e. number of overlapping voxels over number of voxels of the smaller mask which in this case was yielded by the seed-based analysis). Note that the precise threshold value was not found to affect the overall results, as discussed below. For the DMN, the seed voxel was selected within the PCC, whereas for the visual and somatosensory networks, the seed voxels were selected in the occipital cortex and postcentral gyrus, respectively. The time-varying RSN-graded degree for all networks was computed as described above. Wavelet transforms are multi-resolution decompositions that can be used to analyse signals and images offering good time- and scale-resolution [29,30]. In the case of one-dimensional signals, such as the ones we are dealing with, the discrete wavelet transform is essentially a decomposition into a sum of basis functions. These basis functions are shifted and dilated versions of a (bandpass) wavelet function and shifted versions of a (low-pass) scaling function. This decomposition, which is achieved via a fast filter bank algorithm and includes a digital filtering and dyadic subsampling step at each level, results in a set of wavelet and scaling coefficients. The signal can then be reconstructed via the inverse scheme [29]. By decomposing the fMRI and physiological signals via the discrete wavelet decomposition and reconstructing them using the coefficients of a single decomposition level at each time, we are basically ‘breaking down’ the signal into its different frequency sub-bands which then makes it possible to determine whether the effect of the physiological signals on the fMRI signal is more pronounced within a specific frequency sub-band. An advantage of the approach compared with other frequency-based analysis approaches is that the temporal information of the signal is preserved. In addition, application of the wavelet filter bank allows us to search for modulatory effects at different frequency sub-bands, whereas still taking the entire signal into consideration. Using the techniques described and using the wavelet toolbox in Matlab, we applied a five-level wavelet decomposition to the PETCO2 and HR signals, obtained the corresponding frequency decompositions and calculated correlations with the corresponding time-varying RSN degree at all the resulting frequency sub-bands. Because the frequency sub-bands resulting from a wavelet decomposition are the result of digital filtering and dyadic subsampling, they depend on the sampling rate of the signal analysed. They were, therefore, different for PETCO2 and HR because they were sampled at 1/3 and 4 Hz, respectively, as shown in figure 3. Lastly, it is worth mentioning that the choice of basis function did not make a significant difference in our analysis. The results presented were obtained by using the Coiflets wavelet basis [30] (figure 1). Figure 1. The first three levels of the employed one-dimensional discrete wavelet transform for (a) PETCO2 and (b) HR/HRV.
In each case, we quantified the modulating effect of time-varying physiological signal power for both PETCO2 and HR on the time-varying RSN degree by calculating the Spearman rank correlation coefficient, which is suitable in our case due to the fact that our results are paired with respect to time (network degree versus time-varying physiological signal power). We performed the analysis on all 12 subjects (11 for HR analysis, as one subject did not satisfy the SQI mentioned above) and then calculated the mean and standard deviation of the resulting correlation coefficients over all subjects. This was done both for the mask-based analysis (functional and MNI space), seed-based analysis as well as for different wavelet sub-bands. Finally, statistical significance for the computed correlation coefficients was assessed by calculating p-values for all subjects, accounting for multiple comparisons by using Hommel's correction method [31]. First, we note that according to the methodology outlined above for assessing signal quality (SQI), we discarded the HR data from one subject, owing to poor quality; therefore, we present results for 11 of 12 subjects for the HR signal. Representative global BOLD time series, along with the corresponding PETCO2 and HR time series are shown in figure 2. Note that in functional space the correlations between spontaneous PETCO2 fluctuations and the BOLD signal (figure 2a, top panel) are clearer compared with the MNI space (figure 2a, bottom panel), owing to the blurring effect of registering the BOLD time series to the MNI space. Figure 2. (a) Representative global BOLD fMRI (functional and MNI space) and PETCO2time series. The PETCO2 signal is shown after linear detrending to illustrate that the correlations between the PETCO2 and global BOLD signals are more evident in the individual functional space compared with MNI space and (b) representative instantaneous heart rate (HR) signal. (Online version in colour.)
The time-varying DMN resting functional connectivity matrices for one representative subject are shown in figure 3a for successive windows of 50 s length, where considerable variability can be observed. The results shown in the figure were obtained using the mask-based analysis in the individual functional space. As we can see from the corresponding colour maps, the correlation values between DMN areas are overall high. In figure 3b, we show the power spectral density of the PETCO2 and HR signals in the same successive 50 s windows, where it can be seen that there exist time variations in the spectral content of the signals. Figure 3. (a) Dynamic resting DMN functional connectivity matrices obtained using the maximum absolute correlation value in the individual functional space (mask-based analysis) of one representative subject for 12 successivetime windows. (b) Time-varying power spectral density of the HR and PETCO2 signals in the same windows. Each successive window has a length of 50 s and the overlap between successive windows is 30 s. (Online version in colour.)
The DMN is shown in standard MNI space in figure 4a and in a representative functional space in figure 4b. Figure 4. (a) The default mode network (DMN) as defined in the MNI space using mask analysis, containing the following anatomical areas: posterior cingulate cortex, anterior cingulate cortex, precuneus, thalamus and medial prefrontal cortex. (b)The DMN as defined in the functional space of a representative subject using mask analysis. (c) The DMN as defined in the individual functional space using seed-based analysis. (Online version in colour.)
Figure 5 illustrates the time-varying network degree as well as the time-varying PETCO2 and HR signal power for a representative subject in the case of mask-based analysis in the individual functional and MNI spaces. In the case shown here, the degree was computed for the case in which correlations between areas were quantified using the average absolute cross-correlation value between 0 and ±5 time lags (similar results were obtained when the absolute maximum cross-correlation value was used). The results were similar for smaller/longer window sizes but a window length of 50 s, as mentioned in Materials and methods, achieved a good balance between time resolution and reliable cross-correlation function estimation. We can observe considerable differences between the results in the functional and MNI spaces. Specifically, in figure 5a,c, it can be seen that the time-varying DMN degree follows the time-varying HR and PETCO2 power, respectively. On the other hand, in figure 5b,d, the same relation is not as clear, which is reflected on the obtained correlation coefficients (table 1). Figure 5. DMN degree and power of HR (top four panels) and PETCO2 (bottom four panels) signals as a function of time for two representative subjectsobtained from mask-based analysis in (a) functional space, (b) functional space, wavelet level A4 (0–0.25 Hz) (c) MNI space (d) MNI space, wavelet level A4 (0–0.25 Hz) (e) functional space, (f) functional space, wavelet level A2 (0–0.08 Hz) (g) MNI space and (h) MNI space, wavelet level A4 (0–0.02 Hz). (Online version in colour.)
In figure 5, we also present results obtained when using the wavelet transform, for the frequency sub-bands within which the correlation between time-varying network degree and band-limited signal power was found to be highest. In both functional and MNI spaces, the A4 level (0–0.25 Hz) yielded the highest correlations for the HR signal, whereas for the PETCO2 signal, the A2 level (0–0.08 Hz) yielded the highest correlations in the functional space, owing to that resting PETCO2 fluctuations exhibit most of their power below 0.05 Hz [32]. In the MNI space, level A4 (0–0.02 Hz) yielded the stronger correlations for the PETCO2 signal, possibly due to the fact that registration of the BOLD time series to the MNI space essentially results in low pass filtering. Overall, the correlations between time-varying degree and signal power were found to be stronger for the time-varying band-limited power compared with total signal power (see also tables 1 and 2).
The correlation coefficient values between time-varying DMN degree and the total/band-limited power of the PETCO2 and HR signals are given in table 1 for all subjects (mean±s.d.). These suggest the presence of temporal correlations between network degree and signal power. Correlations were found to be much higher in the functional space compared with the MNI space. The correlation coefficients for the band-limited power are given for the wavelet sub-bands that yielded the highest correlation values (A4 (0–0.25 Hz) for HR in both spaces, A2 (0–0.08 Hz) and A4 (0–0.02 Hz) for PETCO2 in the functional and MNI spaces, respectively). In both cases, using the band-limited time-varying signal power yielded higher mean correlations. For instance, the correlation coefficients for PETCO2 increased from around 0.6 (total signal power) to almost 0.8 (A2 level) in functional space. The values obtained in functional space are also considerably higher, suggesting that registration to the MNI space blurs the modulatory effect of PETCO2 on time-varying network degree. Table 2 illustrates the obtained p-values for the Spearman correlation coefficients for all subjects in each space/sub-band. In agreement to the above observations, statistical significance (p<0.05) was observed for most subjects in the individual functional space when total signal power was used and for almost all subjects (10 out of 12 for PETCO2, 11 out of 11 for HR) when the corresponding optimal sub-band was used. On the other hand, statistical significance was not achieved for most subjects in the MNI space, even when the optimal frequency sub-band was used. In the present and next sections, we present results obtained from the individual functional space of each subject, because it was found that registration to the MNI space and the resulting blurring of the corresponding BOLD time series considerably affects the results, as shown above. The spatial overlap between mask-based and seed-based analysis for the DMN areas was equal to 94.82% for the ACC, 44.65% for the PCC, 76.33% for the thalamus, 83.15% for the precuneus and 92.85% for the medial prefrontal cortex ROIs. The Spearman correlation coefficients between time-varying DMN degree and band-limited PETCO2/HR power are given in table 3 for all subjects (mean±standard deviation). These suggest the presence of temporal correlations between network degree and the PETCO2/HR power, as in the case of mask-based analysis (somewhat lower values—not shown—were obtained for total power, as above). As before, correlations were found to be much higher in the wavelet sub-band A2 for PETCO2 (0–0.08 Hz) and in the wavelet sub-band A4 for HR (0–0.25 Hz). Table 4 illustrates the p-values for the obtained Spearman correlation coefficients for all subjects in each space/sub-band. Overall, the results from the seed-based analysis are similar to the mask-based analysis, with statistical significance attained in most subjects (nine out of 12 for PETCO2, nine out of 11 for HR; figure 6). Figure 6. DMN degree and power of HR (top two panels) and PETCO2 (bottom two panels) signals as a function of time for two representative subjects obtained from seed-based analysis in(a) functional space, power of CO2 versus network degree, (b) functional space, power of CO2 at wavelet level A2 (0–0.08 Hz) versus network degree, (c) functional space, power of HR versus network degree and (d) functional space, power of HR at wavelet level A4 (0–0.25 Hz) versus network degree. All correlations were quantified using the average cross-correlation absolute value between 0 to ±5 time lags. (Online version in colour.)
We applied the procedure described above (seed-based analysis only) on two additional RSNs, namely the visual and somatosensory networks (figure 7). The results were overall similar to those obtained for the DMN and revealed a modulatory effect for both PETCO2 and HR on the time-varying degree for both networks. The wavelet-based analysis yielded higher correlations than those achieved for the total signal power for the same frequency sub-bands as above (tables 5 and 6). Statistically significant correlations between the time-varying visual and somatosensory RSN degree and band-limited HR and PETCO2 power were obtained for all subjects in the case of the visual cortex, whereas, interestingly, for the somatosensory cortex significance was obtained for all subjects in the case of HR but for only two out of 12 subjects for PETCO2—even though results were marginally significant for all the remaining subjects (table 6). Figure 7. (a) Resting-state visualnetwork (high visual: orange and primary visual: cyan) and (b) resting-state somatosensory network as obtained from seed-based analysis. (Online versionin colour.)
This paper examined the modulation of dynamic, resting-state connectivity by physiological signal fluctuations. The results reveal a clear effect of time-varying signal power for both HR and PETCO2 on the time-varying network degree for three well-described RSNs: the default-mode, visual and somatosensory RSNs, revealing brain–heart interactions in the context of fMRI-based RSN connectivity studies. This effect was found to be more pronounced for the fluctuations in the physiological spectral content in specific frequency sub-bands (time-varying band-limited power), as revealed by wavelet analysis. In addition, it was found that the observed modulations were not apparent when the analysis was performed in the MNI space, using anatomical masks to define the RSNs of interest. Despite the well-established effect of physiological signals on fMRI connectivity and particular RSN connectivity [17,18,21–23], to the best of our knowledge, this is the first study that demonstrates that the time-varying properties of physiological signals may affect dynamic resting-state functional connectivity. This has important implications, as it suggests that even moderate modulations in the power of these signals can considerably influence RSN analyses and that a significant source of dynamic variations in resting-state connectivity is physiological in nature. Given that resting-state (spontaneous) fluctuations of physiological signals such as HR and PETCO2 are of small magnitude, they are not expected to significantly affect neuronal activity per se; therefore, the observed modulatory effects are likely physiological in origin. Fluctuations in fMRI-based functional brain networks have been observed in time scales from tens of seconds to several minutes [11,12,15]. For instance, recent fMRI studies with high temporal resolution [15] have demonstrated the presence of time-varying patterns that are related to large-scale topological properties of the brain. It has been speculated that these fluctuations may achieve more efficient information transfer and energy expenditure. We selected the sliding window length (50 time lags) as a good trade-off between time resolution and obtaining a good cross-correlation estimate. Interestingly, several studies related to dynamic functional connectivity have used sliding windows of around 50 s (some with lower TR values than 3 s) [11,33], whereas in [15], 60 s windows were used, albeit with a subsecond TR (60 s=83 time lags). In addition, in [34], it was suggested that average correlation values within and between RSNs stabilize at approximately 240 s. Therefore, the selected window length is reasonable given these observations and is similar to the window lengths examined in other studies [35]. In order to obtain an overall measure of dynamic functional RSN connectivity, we computed the time-varying average degree of the brain network defined by the corresponding RSN, as this graph-theoretic measure is a straightforward way to quantify overall connectivity in a brain network and to monitor its variations over time. Alternative graph measures such as clustering coefficient or efficiency are not relevant for our purpose, as we are interested in how ‘connected’ a specific RSN is within a given time window. Note that we computed the graded degree [28], i.e. we did not convert the networks to binary networks, as often done in practice. This is due to that we were interested in the precise degree of modulation of dynamic connectivity by physiological fluctuations, without being concerned if and which network connections are deemed ‘significant’ (non-zero), which would further complicate matters and importantly would also depend on the method of network binarization (i.e. simple thresholding, surrogate data [36], etc.). Furthermore, we selected to quantify functional and not effective connectivity, as the use of the latter has been questioned in the case of fMRI-based connectivity (particularly RSN connectivity), owing to the low-pass filtering introduced by the HRF, which limits the time resolution and the ability to infer causal effects [37]. However, it is worth noting that methods for estimating the HRF from resting-state data have been recently proposed [38], which could in turn provide more reliable effective connectivity estimates from fMRI RSN data [37,39]. This is an important point that deserves further attention. We examined two different methods to define RSNs (mask-based analysis and seed-based analysis) as the network node definition procedure has been shown to influence both connectivity analyses [37] and statistical comparisons between different conditions [40] in fMRI studies. Here, it was not found to influence the results to a large degree. On the other hand, registration of the BOLD time series to the MNI atlas was found to blur the modulatory effects of physiological fluctuations to a large extent. This suggests that sensitivity in tracking dynamic functional connectivity changes is lost when working in the MNI space and that possibly the overall connectivity patterns may not be as accurate as in the individual functional space. Overall, our results were found to be similar among all RSNs examined, with HR yielding overall more pronounced modulations of time-varying degree than PETCO2 (tables 2, 4 and 6). Interestingly, it was found that the somatosensory network yielded marginally significant results for PETCO2 for most subjects in comparison with other networks (table 6), which could be due to vascular anatomy differences between these networks. With regard to the effect of threshold value in the seed-based analysis, we repeated our analysis for different threshold values. For values between around the reported value (between 0.55 and 0.75), the resulting areas from seed-based analysis were relatively cohesive, and the results were very similar to those reported above. For larger threshold values, the seed-based areas become more disjointed; however, they can be constrained anatomically to yield the same number of DMN areas (ACC, PCC, etc.), and the physiological modulations were found to be overall similar. For smaller threshold values, the seed-based areas become more extensive (and possibly overlapping), but the effect of physiological modulations is still present. Finally, extreme threshold values yield either a few voxels or very extensive areas and are not of much practical interest. In order to examine the effects of standard physiological correction methods on the results, we repeated the analysis after implementing two different schemes of physiological correction (RETROICOR [41]) to regress out the effects of (i) HR and respiration and (ii) HR, respiration and PETCO2. The results are summarized in table 7, where we show results (Spearman correlation coefficients) for the broadband signals (i.e. without using wavelets).
When HR and respiration were regressed out, the modulatory effects of both signals were found to be reduced but not completely removed. When all signals are regressed out, the drop in modulation is more substantial for CO2. Therefore, modulating physiological effects do not disappear even after physiological correction—particularly for HR, for which the correlation coefficient remains fairly high (0.53). Moreover, it is worth noting that, whereas in most fMRI studies, HR and respiration are regressed out using RETROICOR or other similar techniques, PETCO2 data are typically not recorded and consequently, the effects of CO2 are not regressed out. The results suggest that collecting PETCO2 data in resting-state studies is important and also that better physiological denoising algorithms for such data are needed. Furthermore, we selected not to perform white matter (WM) or cerebrospinal fluid (CSF) regression, which is a relatively common practice in similar studies, as it is rather difficult to obtain accurate individualized masks for WM and/or CSF and consequently, some fraction of the grey matter signal may be regressed from the data, masking out the pure effects of physiological noise, in which we are interested [22]. The frequency sub-bands that yielded the stronger correlations between time-varying degree and band-limited power are related to each signal's characteristic and suggest that fluctuations in these bands have a clearer modulatory effect. For PETCO2 modulations in very low-frequency power (0–0.08 Hz in functional space and 0–0.02 in MNI space) were found to have the clearer effect, which is consistent with the previously described spectral characteristics of PETCO2 fluctuations [19,32]. For HR, a wider frequency band was identified (0–0.25 Hz) reflecting the fact that HR has a richer spectral content than PETCO2. Additionally, the identified frequency band contains the low-frequency (LF—around 0.1 Hz) spectral peak and possibly also the high-frequency (HF—around 0.2 Hz) spectral peak of the HR signal. While, often, these two spectral peaks are assumed to correspond to cardiac sympathetic and cardiac parasympathetic neural activity, respectively, the relative contributions of each mechanism still remain the subject of controversy and investigation. What is generally accepted is that both LF and HF components of the HR signal (or HRV—note that fluctuations in these two signals have the same spectral characteristics) are affected by the complex interactions between both parasympathetic and sympathetic nerve fibres as well as mechanical, and other factors on the pacemaker cells usually located in the sinoatrial node [42]. The effects of these two components could be disentangled by obtaining estimates of instantaneous HF and LF HR power by using, e.g. point-process models [43]. F.N. performed data analysis, prepared all figures and wrote the paper. C.O. performed data analysis and wrote the paper. P.P. performed data analysis. K.M. collected the data and provided comments on the paper. R.G.W. collected the data and provided comments on the paper. G.D.M. conceived and designed the study, wrote the paper and approved the final version. We declare we have no competing interests. We received no funding for this study. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 14In recent years, a growing number of studies have investigated the role of different cortical and subcortical brain regions involved in autonomic control during a variety of different tasks and sensory stimuli, and our recent neuroimaging meta-analysis has summarized the diversity of brain regions supporting differential control of the central autonomic network (CAN) [1]. For example, while the amygdala, insula and mid-cingulate cortices were found to form the core of the CAN, regional specificity was found when comparing sympathetic and parasympathetic control networks, as well as when comparing central autonomic response with different tasks and stimuli, such as pain. While the telencephalic and diencephalic circuitry supporting autonomic outflow has been imaged in previous studies, the brainstem circuitry—which includes several important premotor and modulatory autonomic nuclei—has proved more difficult to investigate. The brainstem acts as a relay and processing station between the spinal cord, cerebellum and neocortex. It contains vital nodes of many functional systems in the central nervous system, including the visual, auditory, gustatory, vestibular, somatic and visceral senses, as well as the autonomic nervous system (ANS). Despite this indisputable importance, the brainstem has been largely neglected in attempts to measure or model brain functions, especially in human neuroscience. One reason for this neglect is that the anatomical characteristics of the brainstem—specifically, its close vicinity to large arteries and ventricles, high levels of physiological noise driven by cardiac pulsatility, the proximity to a steep magnetic susceptibility gradient produced by the air–tissue boundary posterior to the oral cavity and the small size of its nuclei—present inherent challenges to neuroimaging analysis, including functional magnetic resonance imaging (fMRI). Spatial resolution and the signal- or contrast-to-noise ratio, especially using 1.5 T, are typically inadequate for precise brainstem evaluation. Nevertheless, the field of brainstem functional imaging has significantly advanced in recent years, largely due to the development of new investigation and analysis tools that facilitate studying this critical brain structure [2]. Ultrahigh-field fMRI, which allows for improved signal-to-noise ratio and spatial resolution, promises to further our understanding of brainstem and ANS physiology. Pain is a strong modulator of the ANS, and autonomic dysfunction has been linked with clinically relevant parameters in chronic pain patients, such as fibromyalgia and migraine populations [3–7]. A number of studies have demonstrated that evoked pain stimuli induce increased sympathetic and/or decreased parasympathetic outflows using indirect measures such as heart rate (HR) [8–10], heart rate variability (HRV) [11,12] and skin conductance [13,14]. It is possible that pain-processing nuclei in the brainstem may also be linked with autonomic premotor nuclei, contributing to central autonomic control and to autonomic dysfunction in chronic pain patients. Exploring the interactions between pain and autonomic modulation in the brainstem is thus highly relevant to clinical applications. In this study, we use ultrahigh-magnetic-field (7 T) fMRI, a technology with a high signal-to-noise ratio, allowing for improved spatial resolution for an ANS/fMRI analysis focused on the brainstem. Our recent studies have successfully integrated peripheral ANS recordings and fMRI data [15,16] to identify the purported central control regions for autonomic outflow. Our goal here is to investigate specific brainstem nuclei involved in the processing of a strong modulator of autonomic activity, namely evoked pain. Eleven healthy right-handed subjects (8 males, 3 females, 33±4 years old, mean±s.d.) were enrolled in the study. All experiments took place at the Athinoula A. Martinos Center for Biomedical Imaging in Charlestown, MA, USA. Subjects experienced fMRI scan runs under a resting or sustained pain condition. For sustained pain, a deep pain sensation was induced by inflating a pressure cuff placed on the subjects' left lower leg (gastrocnemius muscle). Inflation commenced just before each sustained pain run. Prior to the imaging session, the pressure was percept-matched between subjects by asking each subject to signal to the experimenter inflating the cuff when pain was first perceived and then when the pain sensation would be rated as 40 on a scale ranging from 0 (meaning ‘no pain’) to 100 (meaning ‘the most intense pain tolerable’). The calibration procedure was repeated at least three times, and the resulting pressure values were then averaged, in order to ensure a robust estimation. As scans were 6 min in duration, for safety reasons, subjects were instructed to slightly move the toes on their right foot should the pain sensation become too strong during the scan run. This would signal to the investigator to reduce the pressure by 10 mmHg. During the session, participants underwent five fMRI scan runs—three resting-state (hereinafter, REST) runs and two sustained pain stimulation runs (PAIN)—randomized in order and lasting 6 min each. At the end of each PAIN run, subjects were asked to rate the experienced pain intensity on the same 0–100 scale used in the calibration procedure. Ratings were collected retrospectively (i.e. at the end of each PAIN run) regarding both the average pain intensity during the entire run as well as the pain intensities related to three 2-min consecutive windows (initial, middle and final, i.e. each one-third of the entire scan run). Finally, subjects were also asked to rate their average level of anxiety during the PAIN run, on a scale ranging from 0, meaning ‘not anxious’, to 100, meaning ‘very anxious’. Whole-brain blood oxygen level-dependent (BOLD) fMRI data were collected on a Siemens 7 T whole-body scanner (Siemens Healthcare, Erlangen, Germany) using a custom-built 32-channel receive array and birdcage transmit coil [17]. Functional data were acquired with gradient-echo single-shot echo-planar imaging (EPI) using a blipped–controlled aliasing in parallel imaging (blipped-CAIPI) simultaneous multi-slice acquisition [18] with multi-band factor 2 and the following parameters: 1.2×1.2 mm in-plane resolution (field of view=192×192 mm2), 126 oblique sagittal slices, 1.2 mm slice thickness, repetition time (TR)=3.5 s, echo time (TE)=23 ms, flip angle=80°, no partial Fourier, band width = 1562 Hz pix−1, effective echo-spacing=0.19 ms, using R=4 in-plane (generalized autocalibrating partially parallel acquisitions (GRAPPA)) acceleration calibrated with 128 fast low-angle excitation echo-planar technique (FLEET) reference lines [19]. In each run, 100 time-series measurements were acquired. For anatomical reference datasets, distortion-matched T1-weighted EPI data (using TR=8 s, TE=23 ms, flip angle=90°, 7/8 partial Fourier) were acquired with a slab-selective adiabatic frequency-offset corrected inversion (FOCI) and a permutation of the temporal ordering of slice acquisition to achieve 20 inversion times per slice [20]. T1 values were fitted to these inversion recovery curves to produce a T1 map for the image volume, from which a synthetic T1-weighted volume closely resembling the contrast of the Montreal Neurological Institute (MNI) template was generated [21]. Concurrent with BOLD data, peripheral autonomic physiological signals were collected at 400 Hz using a 16-channel Powerlab DAQ System (ADInstruments, Colorado Springs, CO, USA) and the Chart Data Acquisition Software (ADInstruments) running on a conventional Windows OS laptop. In order to estimate the timings of cardiac contraction, a piezo-electric pulse transducer was placed on the index finger of the right hand to record the cardiac pulse signal via blood pressure fluctuation. Respiration was monitored through a custom-built pneumatic belt placed around the subjects' chest/abdomen. Low-compliance tubing connected this belt to an air pressure transducer (PX138-0.3D5 V; Omegadyne, Inc., Sunbury, OH, USA), thereby producing voltage data that corresponded to changes in respiratory volume [22]. Of the original 55 scan runs (11 subjects, five runs per subject), four were discarded due to low-quality pulse pressure signals. Pulse pressure signals were annotated through an automated method followed by manual adjustment, in order to obtain correct peak detection. Beat-to-beat intervals were then fed into a point-process algorithm used to develop local likelihood HR estimation, in order to compute instantaneous estimates of HR and HRV. The previously validated approach models the stochastic structure assumed to generate the pulse pressure peaks as a history-dependent inverse Gaussian process, as its explicit probability density is derived directly from an elementary, physiologically based integrate-and-fire model [23]. The mean of the beat-to-beat interval lengths is modelled as a linear function of the last k beat-to-beat intervals, and this allows for the estimation of the dependence of such intervals on the recent history of parasympathetic and sympathetic inputs to the sinoatrial (SA) node of the heart. Finally, from this set of k regressive coefficients, the total spectral power was computed, and the low-frequency (LF-HRV, 0.04–0.15 Hz) and high-frequency (HF-HRV, 0.15–0.5 Hz) spectral components were extracted. This approach offers the advantage of estimating the dynamics of the model parameters, and, consequently, the time-varying behaviour of the spectral indices, at any time resolution. Thus, the temporal resolution of the HRV index was set to match the temporal resolution of the fMRI signal time series, as detailed below. As in our previous work [16], the instantaneous HF-HRV index was chosen as a metric for parasympathetic activity, in order to evaluate the modulation induced by sustained pain stimulation on cardiovagal outflow. Furthermore, the adopted point-process approach also provided instantaneous HR estimation. Both HR and HF-HRV series were estimated using a fixed model order k=8, every Δ=2 ms, low-pass filtered at 0.14 Hz, and resampled at the fMRI TR time points. Each HF-HRV power series went through an additional thresholding below the 98th percentile, in order to remove outlier time-series values, and to normalize the time series to a common peak, thus enhancing sensitivity to the full-range dynamics. The resulting signal was then used as a regressor in the general linear model (GLM) fMRI analysis (see below). For respiratory data, the maxima and minima for each breathing cycle were identified through an automated method followed by manual adjustment of the respiratory signal, and the changes in the respiration volume per time (RVT) were estimated as in [24], in order to include a regressor of no interest related to respiration activity in the GLM design matrix (see below). The fMRI data were preprocessed using the Oxford Centre for Functional MRI of the Brain (FMRIB) Software Library (FSL; v. 5.0.6) and Analysis of Functional NeuroImages (AFNI) [25]. A first preprocessing stage included image-based retrospective correction (RETROICOR), slice timing correction (using a script developed in-house in order to take into account the simultaneous multi-slice acquisition of our dataset), motion correction and brain extraction. Then, a spatial preprocessing step was performed, consisting of affine coregistration and nonlinear warping of the T1-weighted EPI to the nonlinear version of the ICBM152 MNI template. The estimated transform matrices were then inverted, in order to be applied to a brainstem mask originally defined in the ICBM152 MNI space, and to transform this mask to the individual functional space. This step was performed in order to mask and retain only the brainstem region from the whole-brain BOLD data, thus reducing the volume of interest in the fMRI analysis and masking out vascular and other non-parenchymal structures surrounding the brainstem, known to be heavily affected by physiological noise [26]. The brainstem mask was taken from Beissner et al. [26]. It comprised the entire medulla, pons and mesencephalon and was based on grey and white matter tissue maps of the ICBM152 template thresholded at a tissue probability of 0.9. Finally, minimal spatial smoothing (full width at half maximum (FWHM)=2 mm) was applied to the BOLD data. We first evaluated autonomic physiological data to assess whether sustained deep pain altered autonomic activity. Mean HR, HR variance (HRvar), LF-HRV and HF-HRV power, LF/HF ratio and respiration rates were averaged over sessions and across subjects for both REST and PAIN conditions, and statistically tested in order to evaluate the presence of significant differences between the two conditions. As a second step, the indices showing significant changes (i.e. a significant difference between REST and PAIN conditions) were evaluated in the initial, middle and final 2-min windows of the entire 6-min run, in order to estimate temporal variability across the entire 6-min run and identify potential time frames during which a stronger effect was induced. A Kolmogorov–Smirnov test evaluated Gaussianity of the data, and subsequent statistical analyses were performed accordingly. All post hoc comparisons were Bonferroni-corrected for multiple comparisons, and the statistical significance threshold was set at p=0.05. The purported brainstem control nuclei supporting cardiovagal modulation in response to deep pain were obtained using the HF-HRV power series as a regressor of interest in a GLM analysis. The preprocessed parasympathetic measure was convolved with a canonical gamma haemodynamic response function (HRF; SD=3 s, mean lag=6 s). In addition to HF-HRV power, two regressors of no interest (cardiac and respiratory activity) were also included in the design matrix, as follows. The HR time series, resampled at the TR of the fMRI acquisition, were convolved with a specific cardiac response function [27] known to reflect the generalized pulsatile artefact resulting from the cardiac contraction pressure wave. Similarly, respiration-related artefacts were reduced by convolving the RVT series with a specific respiration response function [28], and including the result in the design matrix as a second confound regressor. Statistical parametric mapping was carried out using the fMRI Expert Analysis Tool (FEAT v. 6.00; FSL). In the first-level analysis, separate subject-level GLMs were evaluated for each run, for a total of 51 estimations. Parameter estimates derived from each REST and PAIN run were then normalized to MNI space and passed up, with their variances, to second-level fixed-effects analyses for each subject, in order to obtain individual statistical maps for the two examined conditions (leading to 22 second-level estimates). Finally, two group-level mixed-effects analyses (FMRIB's Local Analysis of Mixed Effects (FLAME), FEAT, FSL) [29] produced separate statistical maps for both the REST and PAIN conditions. Furthermore, a REST versus PAIN difference map was evaluated using a paired t-test. Following the results of the 2-min windows analysis of physiological data, BOLD series were similarly split and the same three-level analysis was implemented for specific windows showing a significant effect for HF-HRV modulation. All resultant statistical brainstem maps noted above were corrected for multiple comparisons and family-wise error with a cluster-forming threshold of Z=2.3 and cluster-corrected at p<0.05. In the 2-min windows analyses, for both autonomic physiological data and ANS/fMRI, we chose to compare PAIN periods with the correspondent REST periods (e.g. initial 2 min of PAIN versus initial 2 min of REST). Thus, in order to rule out potential carry-over effects of pressure from the preceding PAIN condition on the initial 2-min of REST, both physiological and fMRI data were first tested for the presence of significant variations due to time. All subjects tolerated the sustained pain stimulation procedure. Pressure values used to evoke a deep pain sensation rated as 40 on a scale from 0 to 100 in the PAIN runs ranged between 100 and 240 mmHg (171.82±50.75 mmHg, mean±s.d.). The pressure estimated through the calibration procedure was used in both PAIN runs or reduced by 10 mmHg in the second run when requested by the subject. Two subjects signalled to reduce the pressure by 10 mmHg during the run. One subject asked for a substantial reduction of pressure (−90 mmHg) for the second PAIN run. This same subject also reported the highest levels of anxiety (60/100 in the first PAIN run, 55/100 in the second run) relative to any other subject and with respect to the average anxiety rating (19.69±22.25), which demonstrated that on average subjects experienced only mild levels of anxiety during scanning. The average pain ratings (0–100 scale) were significantly higher in the first PAIN run than in the second run (first PAIN run: 49.44±9.45; second PAIN run: 44.62±7.05, paired t-test, p<0.05), but no significant differences were found when comparing ratings relative to the consecutive 2-min time windows (first PAIN run: initial 2 min, 45.94±11.72; middle 2 min, 45.62±9.29; final 2 min, 49.69±11.61; second PAIN run: initial 2 min, 43.12±11.93; middle 2 min, 42.5±8.45; final 2 min, 44.37±8.63). Interestingly, a significant reduction in anxiety level was found between the two PAIN runs (first PAIN run: 22.50±22.52; second PAIN run: 16.87±23.13, paired t-test, p<0.05), thus suggesting that the reduction in average perceived pain intensity in the second run could be due to a reduction in the anxiety level, or vice versa. Finally, no significant effect for time window was found in the consecutive 2-min ratings, thus allowing us to exclude any temporal summation or adaptation mechanisms in the examined group. Kolmogorov–Smirnov testing demonstrated lack of Gaussianity for all the autonomic indices considered (HR, HRvar, LF-HRV power, HF-HRV power, LF/HF ratio and respiration rate), thus subsequent statistical analyses were carried out using non-parametric tests. For the initial 6-min run analysis, autonomic indices were compared between REST and PAIN through a Wilcoxon signed-rank test (table 1). A significant decrease during PAIN with respect to REST was found only for HF-HRV power (REST: 1330.89 (336.85;2134.77) ms2; PAIN: 1012.75 (306.51;1383.3) ms2; median(interquartile range; IQR), p=0.042) (figure 1a). An Ansari–Bradley test also showed a significant reduction in the dispersion of HF-HRV values in PAIN runs with respect to REST (p≪0.001), demonstrating that a decrease in fluctuation amplitude accompanies the reduction in median values. No significant changes were found for any other indices.
Figure 1. (a) Cardiovagal (HF-HRV) response to pain (compared with rest) for the entire 6-min run demonstrated reduced HF-HRV power during PAIN. (b) HF-HRV response to pain (compared with rest), evaluated in three consecutive 2-min windows across the entire run, demonstrated reduced HF-HRV power during PAIN for the initial 2-min period only. *p<0.05; statistical significance testing was Bonferroni-corrected for multiple comparisons.
Given the significant effect induced by pain on our index for cardiovagal modulation, we also inspected HF-HRV power for the 2-min time windows, thus dividing the 6-min runs into three consecutive timeframes (‘initial’, ‘middle’ and ‘final’). A significant effect of time window was confirmed through a Skillings–Mack test [30], used as a substitute for the Friedman test in the presence of an unbalanced design (p<0.001). A first set of post hoc comparisons found a lack of significant differences among the three 2-min segments of HF-HRV power during REST, supporting a lack of any potential carry-over effect from preceding PAIN stimulations. Subsequent post hoc comparisons showed a significant decrease in HF-HRV power during PAIN with respect to REST in the first 2-min interval (initial-REST: 1486.42(348.66;2409.64) ms2; initial-PAIN: 890.88(282.6;1581.12) ms2, p<0.05). No significant differences were found for the second (middle-REST: 1389.44(338.72;2042.82) ms2; middle-PAIN: 1119.44(291.03;1370.09) ms2) or the third time window (final-REST: 1125.82(325.03;2128.66) ms2; final-PAIN: 917.13(365.57;1246.99) ms2) (figure 1b). As HF-HRV showed significant pain-evoked modulation (figure 1a), this index was used as the regressor of interest in the fMRI data analysis. The differential map resulting from the paired t-test between REST and PAIN conditions is reported in figure 2a, and the brainstem regions identified are shown in table 2. The significant clusters (Z>2.3, P<0.05) are overlaid on an average of individual functional BOLD images, previously normalized to MNI space. We also report the mean signal intensity response for both REST and PAIN, with an average per cent-change score extracted from a 2 mm radius sphere centred on the peak voxel of each significant cluster. These bar plots show which of the two conditions was driving the significant difference. Figure 2. (a) Differential map (PAIN–REST) for the HF-HRV/fMRI analysis of the entire (6 min) runs. Group maps are compared with graphical representations of brainstem nuclei as in the Duvernoy atlas [31] (white squares indicate the correspondence on the functional maps). A significant difference between REST and PAIN conditions was found in the pontine nuclei (Pn), the rostral ventromedial medulla (RVM) and the nucleus reticularis medullae oblongatae centralis, including the nucleus ambiguus (Rt/NAmb). (b) Normalized values of HF-HRV power (black line) and BOLD signal from Rt/NAmb (grey line) from a representative subject during a REST and a PAIN run. (c) Per cent signal change from the significant clusters showing a reduction in the negative correlation (Gi/RVM cluster) or a shift to positive correlation (Pn and Rt/NAmb clusters) during PAIN. Bar plot error bars represent s.e.m.
For the entire 6-min runs, compared with REST, we found that PAIN produced a reduction in the anti-correlation between HF-HRV and the fMRI signal in a cluster located in the upper medulla, presumably containing the rostral ventromedial medulla (RVM) including the gigantocellular nucleus (Gi) (based on reference atlases [31,32]). At the same time, a switch from anti-correlation to positive correlation was found in a lower medullary region, consistent with the dorsal, intermediate and ventral nucleus reticularis medullae oblongatae centralis (Rt), including the nucleus ambiguus (NAmb), and in two pontine regions consistent with pontine nuclei (Pn) (figure 2c). An example of the switch from anti-correlation to positive correlation is shown in figure 2b, where the fMRI signal extracted from the Rt/NAmb cluster is plotted together with the HF-HRV power regressor for a representative subject (both time series are scaled between 0 and 1 for visualization). The same HF-HRV/fMRI differential map was then evaluated in the initial 2-min time window of the runs, when the HF-HRV power showed a significant reduction (figure 2b). Also in this case, a residual influence on brain activity during the initial 2 min of REST due to previous painful stimulation was first investigated through an ANOVA for repeated measures including the three consecutive HF-HRV/fMRI maps for each subject. No regions were identified by the analysis; therefore, we evaluated the HF-HRV/fMRI differential map using the initial 2 min of both REST and PAIN. The resulting map (figure 3, table 2) shows a partial overlap with the one obtained by analysing the entire run: significant clusters were found in Pn, in an upper-medullary region including Rt and NAmb, and in the lower medulla, encompassing Rt, NAmb, the dorsal motor nucleus of the vagus (DMNX), nucleus gracilis (Gr) and the nucleus of the solitary tract (NTS). Interestingly, a new cluster, not highlighted by the previous analysis, was also found in the upper portion of the pons, consistent with the locus coeruleus (LC). The anti-correlation between the HF-HRV power and the fMRI signal during REST is confirmed also within the initial 2-min window, as well as the reduction of this anti-correlation or shift to a positive correlative association during PAIN for specific brainstem nuclei (figure 3). No regions were identified by the REST>PAIN contrast, for both the 6-min and the initial 2-min analyses. Figure 3. Differential map (PAIN–REST) for the HF-HRV/fMRI analysis of the initial 2-min time window. Group maps are compared with graphical representations of brainstem nuclei from the Duvernoy atlas [31] (white squares indicate the atlas image correspondence on the functional maps). A significant difference between REST and PAIN conditions was found in pontine and medullary regions, including the locus coeruleus (LC) and nucleus of the solitary tract (NTS). Per cent signal change from the significant clusters shows a reduction in the negative correlation or a shift to positive correlation during PAIN. Bar plot error bars represent s.e.m.
Central autonomic control nuclei in the human brainstem have been difficult to evaluate. Our ultrahigh-field (7 T) fMRI study used a sustained pain stimulus to modulate ANS outflow and applied a combined fMRI/HF-HRV general linear model framework to evaluate putative brainstem nuclei that control and/or sense the cardiovagal modulation induced by deep pain. Physiological data analysis showed a significant reduction of parasympathetic activity (HF-HRV) during PAIN compared with REST runs. Moreover, the pain-induced reduction in HF-HRV was most prominent in the initial 2-min time window during this scan run, and this time frame was further evaluated to determine which brainstem nuclei were associated with reduced cardiovagal outflow. Analysis of the full 6-min scan run found a significant PAIN>REST contrast in several clusters consistent with the RVM, Rt/NAmb and Pn. Analysis of the initial 2-min period of the scan, when HF-HRV was most prominently reduced by PAIN, found a significant PAIN>REST contrast in clusters consistent with the previous 6-min result (Rt/NAmb, Pn), but also identified clusters consistent with DMNX/NTS as well as LC. The brainstem regions implicated in supporting the cardiovagal response to pain in our study include different structures and nuclei known to be involved in autonomic and nociceptive functions from animal experiments. Both the 6-min and 2-min analyses identified a cluster in the lower pons, which was anatomically consistent with Pn. These nuclei project to the cerebellum and are believed to act as a relay station between the cerebrum and the cerebellum [33]. Furthermore, since Pn also receive projections from hypothalamic and limbic structures, it has been proposed that these connections play a role in the cerebellar contributions to specific cognitive tasks, as well as to the integration of emotional information in movement execution [34]. Recently, an fMRI study reported ventral pons activation ipsilateral to the location of pain in a patient with cluster headache, during typical pain attacks. Given the known projections of these nuclei, the authors hypothesized that Pn are also involved in pain avoidance as they are also associated with the activation of structures involved in motor function and reactive behaviour [35]. A second partially overlapping cluster extended longitudinally in the medulla, and encompassed different nuclei: its upper portion was consistent with RVM including the gigantocellular nucleus (Gi), while its lower portion was localized to the nucleus reticularis (Rt), including the NAmb. For the initial 2-min time window, a similar cluster also included the DMNX, nucleus gracilis (Gr) and NTS, which extends into the dorsolateral medulla. Interestingly, both NTS and Gr are known to receive nociceptive inputs, the former via the vagal nerve and spinal afferents, the latter from the dorsal column pathway [36]. Rt, and particularly its dorsal portion, has also been proposed as a primary pro-nociceptive centre in the brain's endogenous pain modulatory system, integrating multiple excitatory and inhibitory functions for nociceptive processing [37,38]. On the other hand, NAmb and DMNX are both premotor nuclei implicated in the generation of autonomic response patterns evoked by physiological and various sensory stimuli [39]. Both of these nuclei are involved with efferent parasympathetic outflow and play a critical role in parasympathetic reflexes, accepting input from the NTS, which is the principal nucleus for incoming signals, particularly from the viscera via the afferent vagus nerve [40]. Finally, the RVM has been identified as one of the key endogenous pain modulatory areas of the brain, conveying descending pain modulatory influences from the periaqueductal grey to neurons located in the dorsal horn of the spinal cord [41,42]. Thus, it is conceivable that sustained pain stimulation linked pain processing in nociception-associated brainstem nuclei with autonomic response via known autonomic premotor nuclei in the medulla. Interestingly, the initial 2-min time-window analysis identified another cluster, located in the right upper pons and consistent with the LC. LC is the main source of noradrenaline in the forebrain, and is implicated in autonomic regulation, with an excitatory influence on sympathetic outflow, and an inhibitory influence on parasympathetic outflow [43]. Importantly, the LC also plays a major role in the processing of pain, channelling nociceptive information to the somatosensory cortex and exerting an inhibitory influence on pain sensation [44,45]. Therefore, its involvement in the first minutes of pain stimulation could be interpreted as a coping mechanism in our healthy adult subjects, in an attempt to decrease the pressure-evoked pain sensation. In both the 6-min and the initial 2-min analyses, we found that an anti-correlation between HF-HRV and the fMRI signal seen during REST was reduced, or even shifted to a positive correlation, during the sustained pain stimulus. In our previous ANS/fMRI study, where we used similar methods to evaluate the CAN response to visual nauseagenic stimulation, this aversive stimulus increased anti-correlation in cortical areas such as the insula [16]. Such common anti-correlations, typically in neocortical regions, may be due to inhibitory links between the identified regions and premotor cardiovagal outflow nuclei (i.e. NAmb). The present brainstem-focused results, which demonstrated that evoked pain induced a positive correlation between the HF-HRV signal and brain activity in important ponto-medullary nuclei such as Pn, LC and most importantly NAmb, suggest that, under sustained nociceptive drive, NAmb activity became in-phase with HF-HRV power, possibly contributing to the decreased cardiovagal modulation induced by this stimulus. However, as NAmb is an elongated nucleus with a very small cross-sectional area in the axial plane (<1 mm2), future studies may need to push for an even better spatial resolution and improved temporal resolution to definitely assess the association between NAmb activity and HF-HRV power. Indeed, the modelling of the association between HF-HRV and fMRI signal could be suboptimal, especially in terms of the specific haemodynamic response linking the two signals. Further research should thus explore more precise modelling of the neurovascular coupling/haemodynamic response, leading to a deeper understanding of pain-induced changes in ANS/fMRI analyses. Interestingly, the two differential (PAIN—REST) analyses, one using the entire run (6 min, figure 2) and the other using the initial 2 min when HF-HRV reduction was most pronounced (figure 3), were partially overlapping, thus suggesting that (i) a 2-min period of data collection contains a high enough signal-to-noise ratio to perform such analyses and (ii) an early, stronger modulation of autonomic outflow may be driving the significant differences found in HF-HRV power and in ANS/fMRI analysis results for the entire 6-min data collection period. This finding also supports the reliability of the time-varying assessment of HF-HRV achievable by the point-process framework, which is able to provide instantaneous estimates (at a temporal resolution matched to the BOLD acquisition time points) within windows as short as 2 min. This may not be true with other approaches, which are based on the assumption of quasi-stationarity across the entire time window, typically at least 1 or 2 min in duration. Thus, analysis of the initial 2-min time window, when autonomic outflow was most differentiated between REST and PAIN, supported the specificity of the brainstem nuclei involved as relevant to the control and sensory feedback of cardiovagal modulation by sustained pain stimulation. Although the results yielded by the two analyses were quite similar, it should be noted that, when the analysis was focused on the specific time window characterized by greater cardiovagal reduction, the differential map demonstrated the involvement of other brainstem nuclei known to be involved in autonomic and nociceptive modulation, not identified when considering the entire length of the run. One possible reason for this result could be neurovascular response habituation to the nociceptive stimulus, leading to attenuation of the BOLD haemodynamic response for this longer time period. Alternatively, specific nuclei such as LC could show a more phasic response, which was only detectable in the initial stage of the stimulation. While our study successfully identified cardiovagal-associated nuclei in the brainstem response to pain, several limitations should also be noted. The number of subjects was somewhat restricted, and even greater power with a larger cohort may have better identified other brainstem nuclei known from animal models to be related to both pain and cardiovagal outflow/autonomic control, such as the periaqueductal grey. Furthermore, brainstem nuclei typically have a very small cross-sectional area in the axial plane. As different individuals probably have slight differences in the exact location of these nuclei within the brainstem, even if the gross brainstem is perfectly aligned between individuals, and spatial resolution is optimized at ultrahigh-field MRI, spatial smoothing will still be necessary for any group analysis. Future studies may also explore individual subject analyses to further assess CAN physiology, non-invasively, in the human. Second, the anxiety ratings were assessed only at the end of each PAIN run, and not on consecutive 2-min time windows as with the pain intensity ratings. Therefore, it was not possible to control for a potentially increased anxiety level in the first 2 min. On the other hand, on average, subjects experienced only mild levels of anxiety (19.69±22.25, mean±s.d., on a scale ranging from 0 to 100), with a significant reduction in the second PAIN run. Thus, the contribution of anxiety level to the examined brainstem activity is likely to be limited. Finally, the pulse signal measured at the finger is about 300 ms delayed with respect to the contraction of the heart. Theoretically, this delay should be incorporated into the HRF convolution used for the HF-HRV GLM regressor to model the BOLD response in the brain. In practice though, this delay should not significantly affect the canonical BOLD HRF, which was used for this study and peaks approximately 5 s after the assumed neuronal activity. A delay (about 150 ms) also exists between vagal nerve activity and cardiac contraction via the sinoatrial node [46,47]. Thus, future modelling should (i) better estimate the HRF in the brainstem and (ii) explicitly account for all peripheral delays between neuronal activity and pulse pressure signal recording at the fingertip. As mentioned in the Introduction, the motivation for conducting this study at 7 T was the increased sensitivity and resolution afforded by ultrahigh-field MRI compared with lower field strengths. Both image signal-to-noise ratio and T2* contrast increase with field strength, therefore the functional contrast-to-noise ratio increases dramatically with field strength [48]. This increased sensitivity to activation can be used to reduce voxel sizes to better sample the small nuclei of the brainstem. However, there are multiple experimental challenges at ultrahigh field strengths. One is increased image distortion in EPI—especially pronounced around air–tissue interfaces such as the oral cavity just anterior to the brainstem. This challenge was mitigated in this study through the use of accelerated parallel imaging [49,50], at the expense of some sensitivity. Another challenge that is particularly relevant to this study is the increased intensity of physiological noise fluctuations with increasing field strength [51]. Physiological noise generated by local magnetic field changes driven by the respiratory cycle are stronger close to the chest [52,53] and therefore are more pronounced in the brainstem than in the cerebral cortex. Fortunately, physiological noise contributions are suppressed in small voxels, where thermal noise dominates [51,54], and previous studies using larger voxels than those employed here demonstrated that approximately 10% of the total noise in the BOLD fMRI time series sampled from the nearby primary visual cortex could be explained by respiratory noise [53]; therefore, this noise source is not expected to contribute much to the total signal variability. A larger concern is the strong cardiac cycle-driven physiological noise, which is prevalent within voxels sampling the brainstem due to partial volume effects between the tissue and surrounding cerebrospinal fluid (CSF), where physiological noise is highest in the brain [55], and this can be a major confound in brainstem fMRI [56]. Smaller voxels also provide reduced partial volume effects that allow tissue signals to be separated from signals emanating from adjacent CSF, as has been demonstrated previously [57]. In this work, the physiological noise sources from within the CSF were explicitly avoided by masking the brainstem, and therefore were largely absent from the fMRI signals. This accurate masking was enabled in part by our use of an anatomical reference dataset with identical distortion to the BOLD fMRI data, namely the T1-weighted EPI. Thus our use of low-distortion, high-resolution fMRI acquisition—enabled by accelerated parallel imaging and the higher sensitivity provided at ultrahigh fields—provided fine sampling of the human brainstem while avoiding noise sources from the surrounding CSF regions. In summary, successfully exploiting high spatial resolution fMRI and high temporal resolution HF-HRV estimation, this work presents the association between specific areas in the brainstem and pain-induced autonomic modulation. Results demonstrate the existence of a change in the relationship between parasympathetic outflow and numerous brainstem nuclei known to be involved in autonomic regulation and pain processing, mainly located in the medullary and pontine portions of the brainstem. Written informed consent was obtained from all participants, and the protocol was approved by the Human Research Committee of the Massachusetts General Hospital. The article's supporting data may be available upon request by emailing the corresponding author. R.S.: acquisition, analysis and interpretation of data, article drafting and final approval; F.B.: conception and design, acquisition of data, manuscript revision and final approval; G.D.: acquisition of data, manuscript revision and final approval; J.R.P.: acquisition of data, manuscript revision and final approval; L.L.W.: conception and design, manuscript revision and final approval; N.W.K.: interpretation of data, manuscript revision and final approval; J.K.: acquisition of data, manuscript revision and final approval; R.G.G.: acquisition of data, manuscript revision and final approval; V.R.: acquisition and analysis of data, manuscript revision and final approval; A.M.B.: interpretation of data, manuscript revision and final approval; S.C.: interpretation of data, manuscript revision and final approval; V.N.: conception and design, acquisition, analysis and interpretation of data, manuscript revision and final approval; R.B.: analysis and interpretation of data, manuscript revision and final approval. We have no competing interests. We thank the following organizations for funding support: Regione Lombardia and Fondazione Cariplo, Project ‘THINK&GO’ (R.S.); National Institute of Biomedical Imaging and Bioengineering (NIBIB), NIH: K01-EB011498, Center for Functional Neuroimaging Technologies (J.R.P.), P41-EB015896 (J.R.P., L.L.W.); German Research Foundation grant no. BE4677/1-1 (F.B.); National Center for Complementary and Integrative Health (NCCIH), NIH: K01-AT008225 (G.D.), P01-AT006663 (V.N.), R01-AT007550 (V.N.); National Institute for Arthritis and Musculoskeletal and Skin Diseases (NIAMS), NIH: R01-AR064367 (V.N.); National Institute of Mental Health (NIMH), NIH: R21-MH103468 (R.G.G., V.N.); National Institute of Neurological Disorders and Stroke (NINDS), NIH: R21-MH103468 (V.N.); Instrumentarium Science Foundation, Swedish Cultural Foundation in Finland, Academy of Finland, grant no. 265917 (V.R.). This work also involved the use of instrumentation supported by the NIH Shared Instrumentation Grant Program and/or High-End Instrumentation Grant Program; specifically, grant nos. S10RR019371, S10RR023034, S10RR023401, and S10RR020948. We thank Drs Thomas Witzel and Boris Keil for supporting hardware and software developments for 7 T fMRI. FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 15Spontaneous fluctuations in the blood oxygen level-dependent (BOLD) signal are correlated with local field potential activity [1]. Analyses of the functional connectivity (FC) between spontaneous low-frequency BOLD signal fluctuations have identified a number of large-scale intrinsic connectivity networks (ICNs). These ICNs are related to sensory, motor, language, social–emotional and cognitive functions, suggesting that spontaneous BOLD fluctuations play a fundamental role in encoding brain function. The above-mentioned patterns are consistent across different levels of consciousness ranging from wakefulness down to sleep and anaesthesia [2], and find correlates with other imaging modalities including electroencephalography (EEG) and magnetoencephalography [3,4]. Recent studies have revealed that these ICNs can be identified by transiently synchronized spontaneous BOLD ‘events’ in the distributed brain regions, which can for example be identified as peaks above the threshold. A growing amount of evidence points to these spontaneous BOLD events governing the brain dynamic at rest [5–7]; these spontaneous BOLD events can be revealed by point process analysis (PPA) [6], the core idea of it being in this context to isolate events in the BOLD time series (for example, peaks in the standardized time series) and to look at their spatial and temporal distribution. Compared with static FC maps constructed from correlations between the whole time series, the FC maps derived by PPA appear to be similar but carry more information on the individual states characterizing brain dynamics [6,8]. Both dynamic FC and PPA are dependent on the variance of BOLD signal, thus they are sensitive to the contribution from non-neuronal fluctuation. A wide range of artefacts may induce changes in BOLD signal, such as thermal noise, hardware limitations [9] and participant's movements inside the scanner [10]. In addition, as the BOLD signal is a measurement of changes in blood flow, oxygenation and volume [11], these changes may be caused by neuronal activity through neurovascular coupling, or alternatively arise from any other physiological processes that affect blood oxygenation or volume [12]. Accordingly, noise and physiological fluctuations may contribute to the rich spatio-temporal information content of dynamic brain activity and connectivity [13]. To reduce confounds deriving from non-neural activity-related processes, plenty of advanced noise cleanup methods have been proposed [14]. Owing to the lack of ground truth in noise removal, most of them focus on improvements in temporal signal-to-noise ratio (tSNR) or reproducibility of FC maps. Yet, no study has explored how and to what extent these confounds affect sparse spontaneous events. As proposed in previous studies, the spontaneous BOLD point process events in resting-state fMRI are assumed to be induced by these spontaneous neural events. Then, it would be possible to retrieve the corresponding haemodynamic response function (HRF) of the spontaneous neural event at rest [15,16]. Apart from the variation of amplitude in BOLD signal, additional temporal characteristics of the haemodynamic response, not available from tSNR and FC maps such as time to peak, could be revealed by statistical analysis of spontaneous point process HRF. Unlike thermal noise, physiological fluctuations can introduce fluctuations in the fMRI signal that are uncoupled from neural activity, and are among the most important confounds in BOLD signal change [17]. In fact, cardiac mechanisms include changes in cerebral blood flow/volume and arterial pulsatility [18]. Respiration effects include changes in B0 and arterial CO2 partial pressure [19]. Although cardiac and respiratory cycles have relatively high frequencies in contrast to the typical low-frequency (less than 0.1 Hz) BOLD fluctuations, aliasing of physiological components to lower frequency range will inevitably occur owing to lower sampling rate in BOLD fMRI than cardiac and respiratory cycles [20]. Recent studies have shown that these nuisance confounds can significantly alter FC maps of the intrinsic brain networks, such as the default mode network [21–,23]. Nonetheless, ample evidence has been collected to support that resting-state FC does have a neuronal underpinning and cannot purely be the result of physiological noise. To date, it is still not clear to what extent the physiological confounds affect the haemodynamic response retrieved by PPA, and more information on this would be helpful for understanding the physiological foundation of functional coupling among brain regions [24]. A number of methods have been developed to reduce physiological confounds in the BOLD signal [21,22,25–27]. Retrospective image space correction of physiological noise (RETROICOR) is one of the most employed methods to correct the cardiac and respiratory quasi-periodic fluctuations [25]. However, it only filters cyclic effects aliased in the fMRI signal, whereas the physiology-related low-frequency fluctuations remain in the data. The time-shifted respiratory volumes per unit time (RVT) and heart rate (HR) time series were introduced to account for more variance in BOLD signal than that induced by non-periodic fluctuations arising from cardiac and respiratory processes [21,22]. These physiological noise correction models on BOLD have been well validated. In the light of what said so far, we feel it is important to examine their influence on haemodynamic response retrieved from spontaneous point processes. Unfortunately, the advanced technique to remove the physiological confounds may also remove meaningful variance components reflecting activity in the autonomic nervous system (ANS). ANS activity should be considered as theoretically meaningful information, especially when studying brain areas involved in decision-making, conflict resolution and the experience of emotion [28]. A seed-based static FC analysis of posterior cingulate cortex has shown that general ANS activity is significantly related to spontaneous BOLD activity in default mode network (DMN) and task positive network [29]. The sliding window-based dynamic FC analysis further reveals that heart rate variability (HRV) covaries with temporal changes in dorsal anterior cingulate cortex (dACC) and amygdala connectivity maps [30]. These studies indicate that resting-state BOLD activity also contains both physiology-related spontaneous neuronal activity and non-neural fluctuations [12,21,31,32]. Therefore, it is critical to differentiate BOLD point process from ANS modulation and physiological noise confounding. HRV is a popular non-invasive indicant for assessing the activity in ANS. An analysis reporting how spontaneous point process HRF covaries with HRV could therefore explore the nature of autonomic regulation on brain activity at rest. In this study, we investigate to what extent the estimation of the spontaneous point process haemodynamic response is affected by changes in physiological noise, rather than solely by central processes such as neural or astrocytic control. The combination of RETROICOR, RVT and HR is employed to deconvolve the physiological fluctuation influence. As HRV is estimated from cardiac activity, we explore only physiological noise correction effect of the quasi-periodic and non-periodic cardiac fluctuations. Then, spontaneous point process HRF maps are retrieved from the residual BOLD signal. Quantitative analysis on HRF map affected by cardiac fluctuation is performed. Finally, correlation analysis between HRV and spontaneous point process HRF is further explored. Two different resting-state (rs) fMRI datasets are included in this study. The first dataset is the enhanced Nathan Kline Institute-Rockland Sample (NKI-RS), acquired from 3 T Siemens scanners [33]. Here we focus on two different TRs (TR=0.645 s, TE=30 ms, FA=60°, 3 mm isotropic voxels, 900 volumes; and TR=2.5 s, TE=30 ms, FA=80°, 3 mm isotropic voxels, 120 volumes) sequentially collected by multiband (acceleration factor=4) and conventional echo-planar imaging (EPI) sequence. Anatomical images were obtained using an MPRAGE sequence with a resolution of 1 mm3 isotropic. Right-handed subjects in release 4 with complete demographic information were employed in our analysis (n=67, 17 females, age: 12–85 with mean 50.6 and standard deviation (s.d.) 20.8 years). They were instructed to keep their eyes open and fixate a crosshair. The 7 T rs fMRI test–retest dataset used in this study has been publicly released by the Consortium for Reliability and Reproducibility (CoRR) project [34]. Twenty-two participants (10 females) were scanned during two sessions spaced one week apart. The subjects were instructed to stay awake, keep eyes open and focus on a cross. Their age ranged from 21 to 30 years with mean 25.1 and s.d. 2.2, one left-handed subject was excluded, resulting in all right-handed subjects. Each session included two 1.5 mm isotropic whole-brain resting-state scans (TR=3.0 s, TE=17 ms, FA=70°, 1.5 mm isotropic voxels, 300 volumes, GRAPPA acceleration with iPAT factor of 3) and gradient echo field map. Structural images were acquired by three-dimensional MP2RAGE sequence with a resolution of 0.7 mm isotropic. Physiological data (respiratory and cardiac traces) were simultaneously recoded for each rs-fMRI scan. The original data in 7 T dataset (5000 Hz) were down-sampled to 100 Hz. The data in NKI-RS dataset are recorded at a sample rate of 62.5 Hz. Two cardiac fluctuation correction models were constructed to account for components related to (i) cardiac phases (CP) and (ii) heart rate (HR). The respiration fluctuations are also included to account for the physiological noise influences: (i) respiratory phases (RP) and the interaction effects between CP and RP (InterCRP) and (ii) respiratory volume per unit time (RVT). Models for cardiac and respiratory phases and their interaction effects were based on RETROICOR [25] and its extension [35]. Cardiac and respiratory response functions were employed to model HR and RVT onto physiological process of the fMRI time series [21,22,26,27]. For each subject, a set of 20 physiological regressors (i.e. fourth-order Fourier expansion for RP, third-order Fourier expansion for CP, second-order Fourier expansion for InterCRP, RVT and HR) was created using the Matlab PhysIO toolbox (http://www.translationalneuromodeling.org/tnu-checkphysretroicor-toolbox/) for each slice in each fMRI run. Cardiac fluctuation correction based on different combinations of these regressors was studied to investigate the effect of cardiac pulse, performing by a generalized linear model (GLM). The combinations are
HRV analysis was performed on the interbeat interval (IBI) time series in each resting-state session scan, using the HRV analysis software (HRVAS, https://github.com/jramshur/HRVAS). The IBI time series were calculated as the peak-to-peak interval of photoplethysmography signal. IBI outliers in each session were removed. The outliers were defined as intervals deviating 20% from the previous interval. To alleviate any non-stationarities within IBI time series, wavelet packet detrending was used before HRV analysis. Finally, time domain and frequency-domain measures were derived from IBI series, including: mean IBI; the standard deviation of the normal-to-normal (NN) interval series, SDNN; the root mean square of successive differences of the IBI series, RMSSD; spectral power of low-frequency (LF: 0.04–0.15 Hz) and high-frequency (HF: 0.15–0.4 Hz) band power and LF/HF ratio, which represents a measure of sympathovagal balance. All structural images in both datasets were manually reoriented to the anterior commissure and segmented into grey matter, white matter and cerebrospinal fluid (CSF), using the standard segmentation option in SPM 12 [36]. Resting-state fMRI data pre-processing was subsequently carried out using both AFNI and SPM12 package with default parameters [36,37], including slice timing correction (T), registration (R), physiological noise model correction (C), despiking (D) and normalization (N). To examine the pre-processing procedure effect on point process acquisition, three commonly used orders of pre-processing steps were applied to the dataset: (i) DCTRN, (ii) DRCTN, and (iii) DTRCN. The raw volumes were despiked using AFNI's 3dDespike algorithm to mitigate the impact of outliers. In slice timing step, the EPI volumes of each run were corrected for the temporal difference in acquisition among different slices to match the middle time slice or half TR (for TR=0.645 s); in the registration step, the images were realigned to the first volume of the first run, the gradient echo field map was processed to create a voxel displacement map and used to correct the realigned images for geometrical distortion (only for 7 T dataset), then the generated mean image across all realigned volumes was coregistered with the structural image, and the resulting warps applied to all the realigned volumes; in physiological noise correction step, physiological noise models were regressed as the covariates, the physiological model regressors from middle time slice were used for DTRCN, whereas each slice data were regressed by physiological model regressors constructed from different time acquisition for DCTRN and DRCTN. Finally, all the processed BOLD images were spatially normalized into MNI space. A conjunction mask was then created to sufficiently cover for all participants in each dataset. Six head motion parameters obtained in the realigning step, Legendre polynomials up to second order were included in a linear regression to remove possible spurious variances from the data. Then, the residual time series were temporally band-pass filtered (0.008–0.1 Hz) and submitted for further HRF retrieval and statistical analysis, including the s.d. and coefficient of variation (CV, i.e. s.d./mean). We employed a blind deconvolution technique to retrieve spontaneous point process HRF from resting-state BOLD fMRI signal [15]. A linear time-invariant model for the observed resting-state BOLD response is assumed. We hypothesize that a common HRF is shared across the various spontaneous point process events at a given voxel, resulting in a more robust estimation. After cardiac fluctuation correction, the BOLD signal y(t) at a particular voxel is given by 2.1 where x(t) is a sum of time-shifted delta functions centred at the onset of each spontaneous point process event and h(t) is the haemodynamic response to these events, c is a constant term indicating the baseline magnitude of the BOLD response, ε(t) represents additive noise and ⊗ denotes convolution. The noise errors are not independent in time owing to aliased biorhythms and unmodelled neural activity, and are accounted for using an AR(p) model during the parameter estimation (we set p=1 in this study). Although no explicit external inputs exist in rs fMRI acquisitions, we still could retrieve the timing of these spontaneous events by means of the blind deconvolution technique [15]. The lag between the peak of neural activation and the peak of BOLD response is presumed to be k×(TR/N) seconds (where2.2 In order to avoid pseudo-point process events induced by motion artefacts, a temporal mask with framewise displacement (FD)<0.3 was added to exclude these bad pseudo-event onsets from timing set S by means of data scrubbing [38]. A smoothed finite impulse response (sFIR) model is employed to retrieve the spontaneous point process HRF shape [39].To characterize the shape of the haemodynamic response, three parameters, namely response height (including non-normalized and normalized by baseline magnitude c, i.e. percent signal change, hereafter referred to as response height, and response height-PSC), time to peak and full width at half maximum (FWHM), were estimated, which could be interpretable in terms of potential measures for response magnitude, latency and duration of neuronal activity [40]. After we retrieved the resting-state HRF for each cardiac fluctuation correction model, response height (-PSC) outlier was rejected by the Grubbs test, then the corresponding HRF parameters for each subject were spatially smoothed (8 mm FWHM), finally individually entered into a random-effects analysis (one-way ANOVA within subjects, with three covariates (age, gender and mean FD) to identify regions which showed significant haemodynamic differences after cardiac fluctuation correction; subjects with mean FD>0.3 were excluded in the statistical analysis). Correlation between each HRV indicator and HRF parameter was analysed by multiple regressions with three covariates (age, gender and mean FD). Type I error owing to multiple comparisons across voxels was controlled by familywise error rate (FWE, voxel-wise correction, p<0.05, cluster size 20). Four regressors of physiological noise correction models are entered into mass univariate GLM analysis, and the adjusted R-square of cardiac fluctuations are estimated by nested model. To include head motion parameters (obtained in the realigning step), here we report only the results after DTRCN pre-processing. Figure 1 shows the averaged fraction of variance explained by quasi-periodic and non-periodic cardiac fluctuation regressors at voxel level over subjects. Most higher adjusted R-square values for quasi-periodic cardiac fluctuation are distributed on the brainstem (the anatomical locations were obtained using the maximum probability tissue atlas from the OASIS project (http://www.oasis-brains.org/) included in SPM12 and provided by Neuromorphometrics, Inc., under academic subscription (http://neuromorphometrics.com/)). For HR, the adjusted R-square value distribution is much lower and more homogeneous, and higher explained variance can also be found in cortical networks, such as DMN. Figure 1. Spatial distribution of voxelwise adjusted R-squared values for different cardiac fluctuations in different TRs. First column:
HRF parameters of each voxel are estimated and mapped on a brain template (figure 2). The median maps of each HRF parameter exhibit spatial heterogeneity across different physiological noise correction models (figure 3). They present similar spatial distributions: higher response height/FWHM/time to peak is present in the occipital/frontal lobe and precuneus, higher response height-PSC is distributed in the brainstem and surrounding areas. The baseline amplitudes in different MRI scanners exhibit different spatial distributions. The interested reader can find spatial maps of these parameters in other datasets [16]. Figure 2. Median maps of HRF parameters (first, fourth, sixth, seventh rows) and BOLD s.d./mean/CV (second, third, fifth rows) across subjects (pre-processed by DRCTN). (Online version in colour.)
Figure 3. Main effect on HRF parameters of four cardiac fluctuation correction models with different pre-processing procedures (repeated-measures ANOVA F-value, p<0.05, FWE correction). CTR, DCTRN; RCT, DRCTN; TRC, DTRCN. (Online version in colour.)
Repeated-measures ANOVA reveals that HRF response height is significantly different across models. The main effect of the cardiac fluctuation correction model on response height is mainly located in the brainstem and the surrounding pulsatile CSF regions and cortex (figure 3; p<0.05, FWE-corrected). The results are strongly affected by different pre-processing procedures and magnetic field strengths. The post hoc analyses suggest that HR gives a limited contribution to the variance of the point process response height (and PSC), whereas the significant magnitude increase is caused by cardiac cycle. The main effect maps of response height show highly similar spatial distribution with CV, s.d. and the response height-PSC (s.d. and response height are not shown in figure 3). The other HRF parameters are less sensitive to cardiac fluctuation correction. The cardiac fluctuation correction on FWHM exhibits high sensitivity to different pre-processing procedures. While the post hoc analyses further indicate that cardiac cycle extends the response duration. The only significant differences found in time to peak are in the 7 T dataset (TR=3 s); the post hoc tests show quasi-periodic cardiac fluctuation extends the time latency in precuneus. Only the correlation maps that passed the conjunction analysis obtained from two or three pre-processing procedures are reported. The correlation map reveals that the results appear to be dependent on the TR (figure 4; p<0.05, FWE-corrected). For TR=0.645 s, FWHM appears to be the only HRF parameter significantly linearly correlated with two HRV indicators. One is the mean IBI in midbrain, pons and surrounding areas: culmen, parahippocampal gyrus, thalamus, insula, superior temporal gyrus and dorsal anterior cingulate; the other is LF power in midbrain and cerebellum anterior lobe (figure 4, top). These positive correlations are also significant without cardiac fluctuation correction in all pre-processing procedures. For TR=2.5 s, only the response height and PSC are significantly correlated with some HRV indicators: LF power and SDNN (figure 4, middle and bottom). The positive linear relationship with LF power/SDNN in response magnitude map is mainly distributed in midcingulate cortex (MCC). More regions are found to be also correlated between LF power and response magnitude-PSC, namely cuneus, precuneus, inferior parietal lobule, angular, precentral gyrus, anterior cingulate cortex (ACC), medial/superior frontal gyrus and superior parietal lobule. The positive correlation between response magnitudes PSC and SDNN shows a spatial pattern similar to the one of LF power, apart from above-reported regions, but extends to include hippocampus, parahippocampal gyrus, caudate, middle/inferior/superior temporal gyrus, supramarginal gyrus, postcentral gyrus and inferior/middle frontal. Figure 4. Correlation maps between HRV and HRF parameters (p<0.05, FWE correction). Top: TR=0.645 s. Middle and bottom: TR=2.5 s (SDNN, LF). The warm (cool) colour denotes positive (negative) T-value. (Online version in colour.)
We investigated how cardiac fluctuations affect the resting-state point process haemodynamic response estimation. These quasi-periodic fluctuations appear to influence the point process HRF magnitude and duration mainly in the brainstem and surrounding cortical. In addition, our results suggest that HRF parameters are sensitive to pre-processing procedures. However, we found a robust correlation between spontaneous point process response duration and mean interbeat interval (higher IBI correspond to slower heart rate)/low-frequency power in brainstem at short TR dataset (TR=0.645). Such positive correlations are persistent and not affected by cardiac fluctuation correction. Previous neuroimaging studies aiming to elucidate brain–heart interactions have looked at the regional cerebral blood flow, derived from positron emission tomography (PET) or arterial spin labelling (ASL) [31,41], at brain activity in PET/task fMRI or brain connectivity in rs fMRI [12,22,31]. Our results show for the first time, to the best of our knowledge, that the reconstruction of haemodynamic response at rest is affected by cardiac activity confounds, especially affecting the temporal information on the duration and latency of the BOLD signature of cortical events. Our findings are consistent with the PET and above-mentioned fMRI studies, involving the brainstem and surrounding areas, insula and dorsal anterior cingulate. The brainstem, including the medulla oblongata, pons and midbrain, is the most important integrative control centre for ANS function and plays an important role in the regulation of cardiac and respiratory function [42,43]. There are more potential sources of signal variance in the brainstem than in any other part of the brain, owing to its anatomical structure: it is highly vascularized with arteries and veins in midbrain, is surrounded by the pulsatile flow of the CSF and it is more connected to the lungs. These factors have been reported to induce stronger changes in the magnetic field B0 [44,45]. The well-established physiological noise correction methods sharply regress out a very large proportion of spurious variation in the brainstem signal. Nonetheless, the linear correlation analysis across subjects still shows that brainstem activity is associated with HRV. The simple mean value of heart rate reveals such relationship with spontaneous point process response durations. In addition, LF power, which is generally thought to be modulated by both sympathetic and parasympathetic activity, is robustly correlated with the response duration in the midbrain. These phenomena are not affected by different processing pipelines, but cannot be evidenced with longer TRs. This may be explained by the fact that a more precise estimation of haemodynamic response duration requires a higher sample rate. To further confirm the effect from different magnitude field strength, a short TR acquisition with a 7 T MRI scanner would be a great resource. With longer TRs, the associations between HRF parameters and HRV are more sensitive to the processing steps. No significant correlation could be found after performing the physiological noise correction first. Moreover, the spontaneous point process response magnitude and its normalization (PSC) are the only indexes that are correlated with HRV parameters. Apart from LF power, SDNN is also significantly correlated with brain areas involved in autonomic activity. SDNN also reflects both sympathetic and parasympathetic activity, providing an index of total HRV [46,47]. Our results reveal that LF and SDNN share regions in MCC that are correlated with response magnitude and its normalization. They are canonical brain areas associated with sympathetic regulation [43]. Other regions showing significant associations in the current analysis have been reported to be related to autonomic activity in previous studies [43]. In fact, the insular cortex is posited to act as an integrator on the brain–heart axis [48]: it has a prominent role in limbic–autonomic integration and is involved in the perception of emotional significance [49]; it also participates in visceral motor regulation, including blood pressure control, in cooperation with subcortical autonomic centres [50–52]. The dACC is also involved in autonomic control [30,53]; the network consisting of insula, dACC and amygdala has been described as crucial in the regulation of central ANS [54]. A human neuroimaging meta-analysis on electrodermal activity and high-frequency HRV revealed that midbrain, insula and supramarginal gyrus are associated with sympathetic and parasympathetic regulation; ACC, thalamus and primary somatosensory cortex are associated with sympathetic regulation, whereas precuneus, superior temporal gyri and angular are associated with parasympathetic regulations [43]. The precuneus and angular gyrus are also among the key nodes of DMN. Several studies have shown that FC maps of the DMN are modulated by heart rate and RVT [12,22]. Intriguingly, the spatial extent of the correlation map with normalized response magnitude is larger than when raw response magnitude is considered. It is worth mentioning that there is no significant correlation between HRF estimation and HRV after conjunction analysis in the 7 T dataset. Apart from stringent thresholds for significance, magnetic field strength and age distributions are different in the two datasets: a study has shown that SDNN index exhibits a linearly correlated pattern of decline with ageing for both genders [47]. A recent study reported a significantly decreased test–retest reliability in FC by physiological noise correction techniques [23]. These results were explained by assuming that these physiological fluctuations are similar and reproducible within a subject across sessions, but to a lesser extent than between subjects. Another explanation given in the same study posited that physiological noise correction could also remove the signal of interest. Physiological fluctuations have been shown to be proportional to magnetic field strength [55]: the physiological processes may therefore contribute much more to variance in BOLD signal when data are acquired with a strong field. Apart from cardiac fluctuations, respiration is another physiological fluctuation that has also been found to strongly modulate the rs fMRI BOLD signal [26,27]. Respiration fluctuations will induce variations in arterial level of CO2, then cause either validations or vasoconstriction, resulting in blood flow and oxygenation changes [56]. HRF magnitude variation is intrinsically related to the CO2 centration owing to vascular reactivity [19,57,58]. In this study, a vascular modulator such as a breath-holding task was not present in all datasets; RVT is used as a surrogate for arterial CO2 concentration to capture breathing rate and depth from respiratory belt measurements suggested by [26]. In addition, as respiration and cardiac pulsations are tightly correlated [59,60], we first partial out respiratory fluctuations before investigating the impact of cardiac fluctuations on the estimation of resting-state point process HRF. The spontaneous events retrieved by point process may include actual neural events, autonomic activities and their interactions. However, in the estimation of the HRF, we hypothesize that they result in a common shape. This may intrinsically limit the disambiguation of the two. To reduce the computational cost and the bias in the linear estimation framework, we employ canonical functions for HR and RVT haemodynamic response [22,27]. Moreover, the lagged RVT and HR regressors are not included in our analysis. These may reduce the contribution of HR in the regression model. However, increasing the number of regressors may induce more bias in GLM especially for short time series (120 volumes in TR=2.5). The more flexible sFIR model could then minimize the risk of assumptions about the spontaneous point process HRF shape [39]. In addition, the sFIR model may also include the components related to cardiac fluctuation in the haemodynamic response, when the latter is not eliminated from the BOLD signal. It has been shown that different processing steps could dramatically change the tSNR in fMRI BOLD signal. In particular, volume registration before physiological noise correction and not performing slice timing correction before physiological noise correction will result in the greatest reduction of temporal noise [61]. Such effect on estimation of spontaneous point process HRF has not been investigated. Resting-state point process is dependent on the variance of BOLD signal; repeated-measures ANOVA results show that HRF magnitudes are similar to BOLD CV and SD in most cases. Our results confirm that different processing steps affect the HRF estimation, not only of its magnitude but also of its temporal parameters (latency and duration). In addition, we find that the effects of HR on HRF estimation are more evident when the DTRCN processing procedure is used together with 3 T dataset, and when DRCTN is used with short TR (0.645 s). Apart from the differences owing to processing, the sFIR model is essentially more sensitive to temporal noise [40]. This study has some limitations that should be noted. First, the proposed method to retrieve the HRF at rest only uses BOLD data. In [16], we have started to explore some validation strategies involving models, ASL, PET and simultaneous EEG-fMRI data. The most convincing validation would nonetheless involve data where the neural activity and the BOLD signal are both extracted, for instance an experiment with simultaneous BOLD signal and intracortical recordings of neural signals [62]. It is worth noting that the processing order ‘DTRCN’ may not capture the aliased physiological perturbation owing to the placement of RETROICOR after slice-timing correction, especially for long TR datasets. Performing despiking before physiological noise correction could improve the regression model fitting by removing large spikes. The despiking procedure was skipped in [38,61]; nonetheless, it appears to improve the results of volume registration over time as illustrated in [63]. However, each procedure has potential disadvantages: temporal correction (slice-timing correction and despking) before realignment may interpolate signals from different brain regions if there is significant head movement; temporal correction after realignment, on the other hand, may shift voxels to adjacent slices and hence disturb temporal order: this problem is especially relevant for interleaved and multiband acquisitions such as those used in this study. To cope with the latter problem, motion-modified RETROICOR has been proposed to include slice contribution to every voxel [61]. This procedure, however, might induce higher computational costs and a bias in the regression model. Therefore, the influences of different pre-processing procedures on HRF estimation should be further explored. In addition, we did not find any significant association between amygdala with HRV parameters. Moreover, no significant correlation with HF power, RMSSD or LF/HF was found after conjunction analysis. It is well known that head motion is an unavoidable source of noise in the BOLD signal [10]. To avoid motion-related artefact contribution to point process detection, in addition to adding motion parameters as a nuisance regressor in the GLM, data scrubbing was performed [38], and mean FD of each subject was included as a covariate for further statistical analysis [64]. This procedure ensures that our findings are unlikely to be affected by motion artefact. This study has demonstrated the impact of physiological noise correction on resting-state HRF estimation, validated at different TRs and magnetic field strength. Several processing pipelines are employed to explore the sensitivity in estimation of resting-state HRF. Intersubject correlation analyses between HRF and HRV parameters suggest that ANS fluctuations modulate the estimation of spontaneous point process response in brainstem. NKI Rockland data are available and described at http://dx.doi.org/10.3389/fnins.2012.00152. CoRR 7 T TRT data are available and described at http://dx.doi.org/10.1038/sdata.2014.54. G.R.W. and D.M. designed the research; G.R.W. analysed the data; G.R.W. and D.M. wrote the paper. We declare we have no competing interests. G.R.W. was supported by the Natural Science Foundation of China (grant no. 61403312), and the Fundamental Research Funds for the Central Universities (grant no. 2362014xk04). FootnotesOne contribution of 16 to a theme issue ‘Uncovering brain–heart information through advanced signal and image processing’. References
Page 16
You have to enable JavaScript in your browser's settings in order to use the eReader. Or try downloading the content offline DOWNLOAD |
What neurotransmitter increases cardiac output?
Postagens relacionadas
direito autoral © 2024 estudarpara Inc.
