Abstract
Objective
Sudden unexpected death in epilepsy (SUDEP) often follows generalized tonic–clonic seizures during sleep, likely resulting from impaired brainstem cardiorespiratory function. We used ictal electrocardiogram (ECG)‐based cross‐frequency phase–amplitude coupling (PAC) to detect cardiorespiratory disruptions, comparing SUDEP to non‐SUDEP cohorts. Leveraging respiratory modulation of ECG signals can provide a robust indirect proxy of respiratory monitoring despite high‐amplitude noise.
Methods
We analyzed ictal ECG and electroencephalographic recordings in 21 SUDEP cases and 21 non‐SUDEP epilepsy controls. Ictal ECG segments from 76 seizures (38 SUDEP, 38 non‐SUDEP) were processed using continuous wavelet transformation to compute PAC between respiratory (.1–.55 Hz, 6–33 breaths per minute) and cardiac (.7–3.7 Hz, 42–222 beats per minute) frequencies. Relative PAC coupling strength was evaluated for respiratory frequencies > .25 Hz (15 breaths per minute) and cardiac frequencies > 1.7 Hz (102 beats per minute). Furthermore, a 3 × 3 grid of PAC ranges was derived for each 20‐s window, yielding 18 features (mean and SD) as inputs to a logistic regression model.
Results
Elevated ictal PAC at higher respiratory (>.25 Hz, p < .0001) and cardiac (>1.7 Hz, p < .0142) frequencies in SUDEP patients suggests ictal respiration modulates ictal tachycardia, leading to cardiorespiratory dysfunction, probably brainstem‐mediated. The logistic model accurately distinguished 38 seizures in SUDEP cases from 38 seizures in non‐SUDEP cases (receiver operating characteristic area under the curve = 91%). Seizures in SUDEP patients had higher propensity scores (p < .001) both per seizure and per patient. All six test seizures (three SUDEP, three non‐SUDEP) were correctly classified using the optimal threshold.
Significance
Ictal ECG‐based PAC analysis is a potential noninvasive biomarker for SUDEP risk, capturing cardiorespiratory dysregulation during seizures. Its integration into wearable ECG devices could enable real‐time risk assessment, informing clinical interventions such as rescue medications, antiseizure medication adjustments, or surgical evaluations.
Keywords: brainstem, logistic model, neurodynamics, phase–amplitude coupling (PAC), risk assessment
Key points.
Risk assessment using ictal ECG‐based cross‐frequency PAC distinguishes SUDEP from non‐SUDEP seizures.
PAC analysis of ictal ECG reveals cardiorespiratory disruptions in SUDEP, showing higher coupling between respiratory and cardiac frequencies.
Elevated PAC in SUDEP reflects reduced brain network dynamism, linked to impaired cardiac and respiratory adaptability.
Noninvasive PAC analysis of ECG could enable real‐time SUDEP risk assessment via wearables, guiding clinical interventions.
1. INTRODUCTION
Sudden unexpected death in epilepsy (SUDEP) is hypothesized to result from seizure‐induced brainstem dysfunction that impairs respiratory and cardiac functions, potentially related to spreading depolarization 1 , 2 or spread of epileptiform activity 3 to medullary autonomic centers. SUDEP typically occurs following a generalized tonic–clonic seizure (GTCS) in sleep.
Cardiac SUDEP biomarkers offer accessibility and low cost but have proved elusive. Differences in heart rate variability and QT intervals could not distinguish SUDEP from non‐SUDEP cases, with a range of abnormalities linked to different seizure types. 4 ECG studies in SUDEP 5 , 6 , 7 have focused on interictal states where movement artifacts are minimal but provide no data on ictal cardiorespiratory functions. We reported a potential SUDEP biomarker using ictal electrocardiography (ECG), which identified a shift in cross‐frequency phase–phase coupling (PPC) to higher frequencies. 8 The lowest frequency PPC feature in SUDEP cardiac rhythms averaged >1 Hz higher than in non‐SUDEP patients, exhibiting a shift in the nonlinear interactions between two cardiac rhythms. Earlier studies of cross‐frequency coupling in epilepsy were conducted on electroencephalographic (EEG) recordings, which were less affected by artifact, and employed phase–amplitude coupling (PAC) analysis, identifying seizure location, 9 differences in postictal generalized EEG suppression, 10 , 11 and functional connectivity networks. 12
Two brainstem nuclei generate the respiratory rhythm: inspiratory pre‐Bötzinger complex and exhalatory retrotrapezoid nucleus. 13 Disrupted respiration and postictal apneas characterize SUDEP, 1 , 3 , 14 , 15 , 16 but quantitative measures of ictal respiratory functions are limited.
We hypothesized that a cardiorespiratory cross‐frequency coupling analysis of ictal ECG may reveal differences between SUDEP compared to non‐SUDEP patients; cardiac and respiratory rhythm and activities would be more disrupted in the SUDEP cohort and reflect alterations in brainstem functions.
2. MATERIALS AND METHODS
2.1. Ictal ECG recordings
Concurrent ECG and EEG recordings were obtained from 42 patients with epilepsy (21 SUDEP, 21 non‐SUDEP) from the Toronto Western Hospital (TWH; Toronto, Canada), Phramongkutklao Hospital (PH; Bangkok, Thailand), the North American SUDEP Registry (NASR; New York, NY, USA), and the University of Siena Scalp EEG Database 17 , 18 (SSED; Siena, Italy; Table 1). SUDEP mean age was 29.8 years (median = 26.0, SD = 10.1), and non‐SUDEP mean age was 36.7 years (median = 36, SD = 11.8), which were not significantly different (p = .06, two‐tailed Welch t‐test). The SUDEP population was 58% male/42% female, and non‐SUDEP 63% male/37% female. Ictal segments of ECG recordings were identified electrographically from concurrent EEG recordings by board‐certified neurologists/electroencephalographers. The patient populations included both focal and generalized epilepsy. Seizures categorized as GTCS included generalized as well as focal to bilateral secondary GTCS; non‐GTCS included seizures that did not generalize. Absence seizures were not included in the study. Both GTCS and non‐GTCS were analyzed (SUDEP: 20 GTCS, 21 non‐GTCS; non‐SUDEP: 19 GTCS, 24 non‐GTCS). Between one and five seizures per patient were analyzed, with a minimum ictal duration of 45 s. The mean duration of SUDEP seizures was 125 s (SD = 54), and mean duration of non‐SUDEP seizures was 100 s (SD = 83), with no significant difference between the populations (p = .13).
TABLE 1.
List of patients.
| Patient | Classification | Age, years | Sex | Non‐GTCSs, n (20‐s windows, n) | GTCSs, n (20‐s windows, n) | Seizure duration range, s | Recording source |
|---|---|---|---|---|---|---|---|
| 1 | Non‐SUDEP | 39 | F | 1 (3) | 120 | TWH | |
| 2 | Non‐SUDEP | 26 | M | 1 (3) | 100 | TWH | |
| 3 | Non‐SUDEP | 42 | M | 2 (2, 2) | 70–83 | SSED | |
| 4 | Non‐SUDEP | 34 | F | 3 (4, 2, 2) | 48–151 | SSED | |
| 5 | Non‐SUDEP | 36 | M | 3 (2, 2, 2) | 63–69 | SSED | |
| 6 | Non‐SUDEP | 25 | M | 4 (2, 2, 2, 2) | 51–69 | SSED | |
| 7 | Non‐SUDEP | 38 | M | 1 (2) | 68 | TWH | |
| 8 | Non‐SUDEP | 20 | M | 1 (3) | 100 | TWH | |
| 9 | Non‐SUDEP | 28 | M | 2 (2, 3) | 75–127 | TWH | |
| 10 | Non‐SUDEP | 56 | F | 1 (4) | 133 | TWH | |
| 11 | Non‐SUDEP | 49 | M | 1 (2) | 83 | SSED | |
| 12 | Non‐SUDEP | 42 | F | 1 (3) | 87 | TWH | |
| 13 | Non‐SUDEP | 35 | F | 1 (4) | 549 | TWH | |
| 14 | Non‐SUDEP | 19 | M | 2 (4, 2) | 62–133 | TWH | |
| 15 | Non‐SUDEP | 41 | M | 4 (4, 4, 4, 3) | 122–225 | TWH | |
| 16 | Non‐SUDEP | 58 | F | 1 (2) | 55 | SSED | |
| 17 | Non‐SUDEP | 55 | M | 4 (2, 2, 2, 2) | 54–74 | SSED | |
| 18 | Non‐SUDEP | 27 | F | 3 (2, 2, 2) | 59–80 | SSED | |
| 19 | Non‐SUDEP | 28 | M | 2 (3, 2) | 62–100 | TWH | |
| 20 | SUDEP | 30 | M | 1 (4) | 127 | PH | |
| 21 | SUDEP | 25 | F | 5 (3, 3, 4, 2, 4) | 82–136 | TWH | |
| 22 | SUDEP | 24 | F | 1 (4) | 300 | NASR | |
| 23 | SUDEP | 33 | F | 1 (4) | 3 (3, 4, 3) | 93–175 | NASR |
| 24 | SUDEP | 22 | M | 2 (2, 2) | 65–80 | NASR | |
| 25 | SUDEP | 25 | M | 3 (4, 4, 4) | 2 (4, 4) | 138–172 | NASR |
| 26 | SUDEP | 19 | M | 2 (4, 4) | 140–150 | NASR | |
| 27 | SUDEP | 49 | M | 1 (3) | 127 | PH | |
| 28 | SUDEP | 34 | F | 1 (3) | 100 | NASR | |
| 29 | SUDEP | 21 | F | 1 (4) | 241 | TWH | |
| 30 | SUDEP | 35 | M | 2 (4, 4) | 141–145 | NASR | |
| 31 | SUDEP | 26 | F | 1 (2) | 1 (2) | 60–86 | NASR |
| 32 | SUDEP | 34 | M | 3 (3, 3, 2) | 1 (3) | 48–116 | NASR |
| 33 | SUDEP | 30 | M | 1 (2) | 53 | TWH | |
| 34 | SUDEP | 26 | F | 1 (2) | 61 | TWH | |
| 35 | SUDEP | 54 | M | 1 (4) | 190 | NASR | |
| 36 | SUDEP | 20 | M | 1 (2) | 65 | NASR | |
| 37 | SUDEP | 16 | M | 1 (2) | 52 | NASR | |
| 38 | SUDEP | 43 | F | 1 (4) | 1 (4) | 129 | TWH |
| 39 | Non‐SUDEP | 41 | F | 2 (3, 3) | 107–123 | SSED | |
| 40 | Non‐SUDEP | 71 | M | 3 (2, 3, 2) | 63–96 | SSED | |
| 41 | SUDEP | 30 | F | 1 (3) | 99 | TWH | |
| 42 | SUDEP | 47 | M | 1 (2) | 1 (3) | 51–116 | TWH |
Note: Patients 39–42 were reserved as test subjects for the logistic model. Patients 41–42 died of SUDEP between 3 and 10 years from their last available recording, as distinct from our training population who died within 3 years.
Abbreviations: F, female; GTCS, generalized tonic–clonic seizure; M, male; NASR, North American SUDEP Registry; PH, Phramongkutklao Hospital; SSED, University of Siena Scalp EEG Database; SUDEP, sudden unexpected death in epilepsy; TWH, Toronto Western Hospital.
SUDEP patients were selected as those who had died within 3 years of their last available recording and classified as definite SUDEP 19 from TWH, PH, or NASR, having no comorbidities that could have contributed to sudden death. SUDEP Patients 41 and 42 from TWH who died between 3 and 10 years of their recording were reserved as test cases. Patients from TWH who had not died within 10 years of their last available recording were classified as non‐SUDEP and augmented by patients from SSED who were presumed non‐SUDEP. Non‐SUDEP patients were selected to create a distribution of age and sex that was not significantly different from available SUDEP recordings. All subjects were admitted to epilepsy monitoring units (EMUs) for presurgical evaluation at the time of their concurrent ECG/EEG recordings, with antiseizure medications suspended for this duration.
2.2. Ethics approval
The institutional review boards of the consortium formed by TWH, PH, and NASR approved the original collection of data used in this retrospective study. Consent for data use in research was obtained from patients undergoing monitoring in EMUs of respective consortium members. The ethical committee of the University of Siena approved collection and research use of SSED recordings in publicly available datasets.
2.3. Cross‐frequency PAC analysis
We apply PAC analysis to characterize ictal cardiorespiratory activity. PAC was used for EEG analysis in epilepsy 9 , 10 , 20 as well as in respiratory signal extraction from ECG. 21 PAC analysis of neurological signals is typically performed on 15‐s and 30‐s windows; longer durations reduce artifact but are less effective on shifting targets. 22 We calculated the continuous wavelet transform (CWT; Equation S3) for 20‐s windows of ECG signals (Figure 1A) using a complex Morlet wavelet for respiratory (.1–.55 Hz) and cardiac (.7–3.7 Hz) frequency ranges (Figure 1B,C). The Morlet wavelet had a central frequency of .8125 Hz and a bandwidth of 5 Hz.
FIGURE 1.
Phase–amplitude coupling analysis of seizure in sudden unexpected death in epilepsy (SUDEP) compared to non‐SUDEP patients. (A) Twenty‐second ictal electrocardiographic signal of a non‐SUDEP patient. (B) Continuous wavelet transform (CWT) of respiratory frequencies (.1–.55 Hz). (C) CWT of cardiac frequencies (.7–3.7 Hz). (D) Phase–amplitude cross‐frequency coupling comodulogram using respiratory frequencies in phase, and cardiac frequencies in amplitude. (E–H) Corresponding analysis applied to a SUDEP patient.
Using the Tort algorithm, 23 coupling between phases of the complex coefficients of the respiratory band CWT and magnitudes of the cardiac band CWT coefficients were calculated, resulting in PAC comodulograms (Figure 1D).
To characterize frequency‐specific differences in coupling strength between comodulograms, we subdivided the comodulogram into nine frequency regions. The mean and SD of each region were computed and are represented in 3 × 3 images (Figure S2). We studied 20‐s windows, using a minimum of one following onset and one preceding termination (Figure 2A), with a minimum seizure length of 45 s. Up to four windows were selected per seizure, including a third window at 40 s and another 40 s preceding termination, duration permitting. The 3 × 3 PAC grid was calculated for each window (Figure 2B), and the mean and SD of their values were calculated for each seizure (Figure 2C).
FIGURE 2.
Application of phase–amplitude coupling (PAC) range analysis to windowed recording. (A) Ictal electrocardiographic (ECG) recording of a non‐sudden unexpected death in epilepsy (SUDEP) patient, showing selected 20‐s windows in red, skipping 20‐s intervals in the same seizure. Maximum Lyapunov exponent (λ max) is indicated for each window. (B) Resulting PAC range analysis for corresponding window. (C) Mean and SD for each PAC range across all selected windows in seizure. (D–F) Applied to a SUDEP patient. PAC strength is elevated at higher frequencies in the SUDEP patient, and maximum Lyapunov exponent of ECG windows confirmed associated decreased complexity in the more highly coupled SUDEP case.
Figure 3 illustrates this approach for an example SUDEP patient compared to a non‐SUDEP patient. The SUDEP patient exhibited maximum coupling above .25 Hz in respiration (x‐axis), as well as above 1.70 Hz in cardiac rhythm (y‐axis; Figure 2F), compared to a non‐SUDEP seizure (Figure 2C), where maximum coupling remained below these frequencies. Elevated PAC strength has been associated with a decrease in brain network dynamism. 12 In cardiac systems, loss of dynamism is linked to a reduction in complexity measures, such as the maximum Lyapunov exponent (λ max), 24 compared to normal ECGs. 25 Demonstrating the link between increased cross‐frequency coupling and decreased complexity measures, the SUDEP seizure had a mean of λ max = .098 (SD = .003) over three successive 20‐s windows (Figure 2D), which was significantly lower (p = .007, two‐tailed Welch t‐test) than that for the non‐SUDEP seizure (Figure 2A): λ max = .122 (SD = .001).
FIGURE 3.
Elevated cardiorespiratory phase–amplitude coupling (PAC) strength. Relative coupling was significantly elevated at higher cardiac and respiratory frequencies compared to typical physiological ranges for n sz (number of seizures) = 38 sudden unexpected death in epilepsy (SUDEP; 18 generalized tonic–clonic seizure [GTCS], 20 non‐GTCS; orange) compared to n sz = 38 non‐SUDEP seizures (19 GTCS, 19 non‐GTCS; blue), including both generalized and nongeneralized seizures. Vertical axis represents the ratio of PAC coupling strengths. Boxplots extend from first to third quartiles, horizontal line indicates mean, and whiskers extend 1.5 × interquartile range from the box. Estimation statistics are presented to the right of boxplots. (A) Relative respiratory coupling above .25 Hz was exceptionally significant (p < .0001, Welch t‐test). (B) Relative cardiac coupling above 1.7 Hz was also significant but less prominent than respiration (p < .0142, Welch t‐test). (C–F) When separated by seizure type, GTCS only (C, D), and non‐GTCS only (E, F), these measures were significantly different in all cases except the cardiac ratio in GTCS (D).
To assess relative PAC strength, we adapt the SumMI method proposed by Caiola et al. 22 for evaluating the strength of PAC features at particular frequencies, taking the ratio of mean coupling within the region of interest against mean coupling outside that region (Equations S7 and S8).
2.4. Risk assessment model
For direct clinical translation of the risks represented by elevated PAC strength at higher frequencies, we trained a logistic model to serve as a risk assessment tool.
The nine features corresponding to mean PAC of windows and nine features corresponding to SD of PACs of windows (Figure 2C) served as the input vector to a logistic model trained on 76 seizures (38 non‐SUDEP, 38 SUDEP) from 38 patients (19 non‐SUDEP, 19 SUDEP).
Each seizure was assigned a propensity score 26 between 0 and 1 by the resulting model, with higher scores indicating SUDEP‐typical seizures. Estimation statistics were applied to the propensity scores to assess the significance of score difference between the populations. The mean of each patient's seizure propensity scores was taken as their overall propensity score.
A receiver operating characteristic (ROC) curve was used to assess the model and identify an optimal threshold for classifying a seizure as at‐risk for SUDEP. The optimal threshold was calculated as the point that minimized the ROC distance to perfect accuracy.
Six reserved seizures (three non‐SUDEP, three SUDEP) from four patients, two non‐SUDEP (Patients 39 and 40), two SUDEP (Patients 41 and 42), not included in the training set, were then scored by the model. The two reserved SUDEP patients were selected as they succumbed to SUDEP later than those in the training set; their recordings were within 10 years of death, whereas the training set recordings were within 3 years.
2.5. Statistical analysis
Results are compared using estimation statistics 27 in addition to null hypothesis significance testing. Reported p‐values are calculated by a two‐tailed Welch t‐test, 28 with an alpha level of p < .05 considered significant.
3. RESULTS
Analysis of cardiorespiratory PAC revealed significant differences between SUDEP and non‐SUDEP seizures. PAC strength was elevated in SUDEP when considering means of seizure windows, as well as elevated SD between windows in SUDEP in most ranges (Figure S3). Elevation of mean coupling strength appeared mostly above .25 Hz in respiratory frequencies (15 breaths per minute), and above 1.7 Hz (~100 bpm) in cardiac frequencies. We therefore evaluated coupling strength above these frequencies to quantify the degree to which cardiorespiratory coupling had moved away from resting‐state dynamics (Equations S7 and S8).
Applying relative coupling strength to these ranges revealed markedly elevated coupling at higher respiratory (Figure 3A) and cardiac frequencies (Figure 3B) in 38 SUDEP seizures compared to 38 non‐SUDEP seizures, including both generalized and nongeneralized seizure types. Respiratory coupling above .25 Hz was highly significant (p < .0001; Figure 3A), whereas cardiac coupling above 1.7 Hz was modestly significant (p < .0142; Figure 3B).
3.1. Risk assessment model
The individual seizure values underlying Figure S3 served as features for training a logistic model, producing propensity scores for each seizure, which further distinguished SUDEP from non‐SUDEP cases.
Figure 4A illustrates the distribution of propensity scores for individual seizures. A comparison of all non‐SUDEP versus SUDEP seizure scores revealed highly significant differences (p < .0001; Figure 4B). When averaged by patient, these scores remained significantly distinct (p < .0001; Figure 4C), highlighting robust differentiation at seizure and patient levels.
FIGURE 4.
Seizure propensity scores from training set. (A) Result of applying methodology to logistic model for each seizure (blue dots indicate non‐sudden unexpected death in epilepsy (SUDEP), orange dots indicate SUDEP). Each seizure for a patient is plotted in a single column. Horizontal lines in the boxplot represent the mean of seizure scores for a given patient. Boxplots extend from first to third quartile seizures scores for a patient; whiskers extend 1.5 × interquartile range from the box. (B) Comparing all non‐SUDEP seizure scores and all SUDEP seizure scores. The scores of the populations were significantly different (p < .0001). n sz denotes the number of seizures. (C) Comparison of average non‐SUDEP seizure scores by patient against average SUDEP seizure scores by patient (p < .0001). n pt denotes the number of patients. (D–F) Trained and tested on generalized tonic–clonic seizures (GTCSs) only. (G–I) Trained and tested on non‐GTCSs only.
The performance of the SUDEP risk assessment model was evaluated using ROC curves and seizure score analyses. For training set seizures, the model achieved an area under the curve (AUC) of 91% (Figure 5A) and remained robust under fivefold cross validation (Figure S5). Each of the six training seizures reserved for testing was correctly classified by the model (Figure 5C) using the optimal point from the ROC as the threshold for classification.
FIGURE 5.
Performance of sudden unexpected death in epilepsy (SUDEP) risk assessment. (A) Receiver operating characteristic (ROC) curve for training set seizures (n sz = 38 SUDEP, 38 non‐SUDEP seizures; area under the curve [AUC] = 91%). Optimal threshold is marked as a red dot. (B) Boxplots of SUDEP versus non‐SUDEP training set seizure scores (38 SUDEP, 38 non‐SUDEP seizures), extending from 1st to 3rd quartile, and whiskers extending 1.5 × interquartile range from the box, with the mean indicated as a horizontal line. (C) Seizure scores of patients not included in training (blue indicates non‐SUDEP; orange indicates SUDEP). Threshold for SUDEP classification is marked as a dashed red line. All seizures correctly were classified. (D) ROC curve for training set patients, using the mean of their respective seizure scores (AUC = 93%). (E) Boxplots of SUDEP compared to non‐SUDEP training set patient scores (19 SUDEP, 19 non‐SUDEP patients), extending from 1st to 3rd quartile, and whiskers extending 1.5 × interquartile range from the box, with the mean indicated as a horizontal line. In panels B and E, n pt denotes the number of patients, and n sz denotes the number of seizures. (F) Scores of patients (#39–42) not included in training (blue indicates non‐SUDEP; orange indicates SUDEP). All patients were correctly classified.
Results improved when adding classification scores per patient, taking the mean of their seizure scores, achieving 93% AUC in the training set (Figure 5D) and classifying the test patients correctly.
4. DISCUSSION
ECG PAC analysis of both GTCSs and non‐GTCSs successfully distinguished SUDEP from non‐SUDEP patients. Relative PAC at heightened respiratory frequencies as well as heightened cardiac frequencies significantly differentiated between these populations. A PAC‐based logistic model successfully predicted SUDEP risk from seizures of SUDEP and non‐SUDEP patients. When tested against six seizures from two patients not included in the training set, but who succumbed to SUDEP between 3 and 10 years from their recordings, all seizures were correctly classified. Ictal ECG‐based PAC analysis may be an inexpensive and accessible noninvasive biomarker of autonomic dysfunction and SUDEP risk.
4.1. Role of cardiorespiratory coupling
SUDEP often follows a GTCS in sleep and results from seizure‐induced impairments of arousal and respiratory and cardiac functions. Brainstem centers mediate arousal and autonomic functions and are linked to SUDEP in animal models showing spreading depolarization 1 , 2 or spread of epileptiform activity 3 to medullary autonomic centers, and in human EEG 29 and neuroimaging 30 studies.
SUDEP cases reveal a progression of severe postictal impairment of arousal, followed by intermittent states of apnea, and ultimately cardiac failure. 1 , 3 , 14 , 15 , 16 Ictal autonomic activity leading to SUDEP remains poorly characterized. Ictal autonomic markers of SUDEP risk would provide critical biomarkers to identify patients at increased SUDEP risk and potentially lead to life‐saving nocturnal monitoring and alarm of a caregiver or on‐demand stimulation of arousal during the postictal state.
Postictal impaired arousal and respiratory distress in SUDEP 31 , 32 , 33 is consistent with disruption of coupled inhalatory and exhalatory oscillators in the brainstem. 13 We reported a PAC‐based technique for the extraction of respiratory signals from ECG, 21 demonstrating that PAC‐derived respiration predicts the same respiratory frequencies as ECG envelope‐based extraction. 34 , 35 We applied a similar approach to directly compare cardiorespiratory neuromodulation and monitor disruptions of brainstem‐driven coupled oscillations, bypassing the task of isolating each rhythm.
We found significantly increased ictal cardiorespiratory coupling strength in SUDEP cases. This atypical pattern of cardiorespiratory coupling is consistent with seizure activity interfering with normal brainstem rhythms, 36 leading to a shift in cardiorespiratory dynamics to higher frequencies. Elevated PAC strength is suggestive of a decrease in brain network dynamism 12 that may lead to impairment of the adaptability of the cardiac system to its environment. Clinically, sustained coupling in respiratory frequencies above .25 Hz is suggestive of elevated ictal respiratory rates.
Other proposed biomarkers based on ictal ECG 8 and EEG 37 recordings were sensitive to seizure type and necessitated seizure type determination and isolated modeling for each. Although our results showed differences between PAC features for GTCS and non‐GTCS (Figure S2), both coupling strength ratios and the logistic classifier successfully distinguished SUDEP cases without explicit seizure type labeling.
4.2. Clinical translation
A continuous monitoring and SUDEP risk assessment of seizures by a lightweight single‐lead ECG device could enable effective SUDEP prevention interventions. Equipping patients with real‐time, noninvasive detection of SUDEP‐typical cardiorespiratory patterns would inform clinicians of the need for immediate medical intervention. Current EEG approaches rely on larger, more cumbersome devices with arrays of electrodes that are impractical for continuous wear, require frequent calibration, and are resisted by patients. 38 , 39 Seizure prediction studies using commercial wearable devices such as smartwatches emphasize analysis of step counts and heart rate but fail to fully exploit increasingly available high‐resolution ECG signals. 40
Our approach relies on access to continuous nontemplated ECG signals, as contrasted with common wearable devices that monitor cardiovascular metrics by proxies such as variation in oxidation or electrodermal activity. 41 Although smartwatches have gained this capability, most are constrained by a short recording window and require the subject to maintain contact between their finger and an electrode, making them unsuitable for ictal surveillance. A ring device (e.g., BIOPAC's research ring 42 ) may allow for continuous ECG without sustained attention by the subject.
Our findings offer a potentially low‐cost and accessible SUDEP risk assessment solution, evaluating each ECG‐monitored seizure against SUDEP‐typical PAC metrics, and identifying an optimal threshold for classifying the patient as having increased risk for SUDEP. In the context of other clinical information, including seizure frequency, rate of GTCSs, and occurrence of status seizures, clinicians would have sufficient support to recommend more aggressive intervention. These may include urgent changes to increase nocturnal monitoring as well as reducing seizure frequency and intensity with antiseizure medications and consideration of surgical options (e.g., resection 43 vagus nerve stimulation, 44 deep brain stimulation, responsive neurostimulation 45 , 46 ).
4.3. Limitations and future directions
Although this study achieved patient age and sex group distributions that did not significantly differ between SUDEP and non‐SUDEP, and because the number of patients included is very limited, future work would benefit from an age/sex‐matched approach to eliminate the possibility of differences due to these variables. Validation of the methodology introduced in this retrospective study requires the capture of ictal events from a cohort of epilepsy patients equipped with wearable ECG systems to compare those who succumb to SUDEP against non‐SUDEP patients. This will benefit from using improved commercial wearables to include a continuous‐contact second electrode and integrating risk assessment analysis with existing seizure onset and termination detection algorithms. Integration of our proof‐of‐principle approach with developments in wearable technology promises to transform SUDEP risk management by enabling personalized, real‐time interventions while minimizing patient burden.
5. CONCLUSIONS
Ictal ECG‐based PAC analysis is a promising noninvasive biomarker to identify SUDEP risk. Elevated PAC at higher respiratory and cardiac frequencies in SUDEP patients suggests seizure‐induced brainstem dysfunction, leading to impaired cardiorespiratory adaptability. Our logistic risk assessment model distinguished SUDEP from non‐SUDEP cases. Integrating this approach into wearable ECG devices could enable real‐time SUDEP risk assessment, facilitating timely clinical interventions such as enhanced nocturnal monitoring, medication adjustments, or surgical evaluations. Prospective studies using advanced wearable technology can potentially validate this biomarker and support its advancement into clinical practice, potentially preventing some SUDEPs through personalized risk management.
AUTHOR CONTRIBUTIONS
Adam C. Gravitis conducted the analysis and prepared the draft manuscript. Richard Wennberg, Peter L. Carlen, and Orrin Devinsky contributed patient data and revised the manuscript. Yotin Chinvarun, Victor Lira, and Juliana Laze contributed to or assisted in processing patient data. Berj L. Bardakjian supervised the study and revised the manuscript.
CONFLICT OF INTEREST STATEMENT
The authors report no competing interests. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.
Supporting information
Figure S1. Validation of respiratory range. Phase–amplitude couplings (PACs) were calculated for a standard dataset containing concurrent electrocardiogram and impedance pneumography (IP). The frequency of the PAC feature with strongest coupling was compared to the instantaneous frequency of the IP signal. Each marker represents one test subject (n = 28). This validated that respiratory rate was correlated with the phase frequency of maximum PAC coupling for subjects at rest.
Figure S2. Subdivision of phase–amplitude coupling (PAC) comodulogram into nine features. (A) PAC comodulogram calculated using respiratory frequencies for phase (x‐axis), and cardiac frequencies for amplitude (y‐axis). (B) The means of each of the nine PAC subdivisions form a 3 × 3 grid for analysis of the window.
Figure S3. Summary of cardiorespiratory coupling results. Average phase–amplitude coupling (PAC) mean and SD by seizure were calculated for all generalized seizures (19 non‐sudden unexpected death in epilepsy [SUDEP], 18 SUDEP; left) and nongeneralized seizures (19 non‐SUDEP, 20 SUDEP; right). SUDEP results reveal a pattern of elevated PAC means at higher frequencies, indicative of reduced dynamism and decreased system complexity.
Figure S4. Relative phase–amplitude coupling strength. (A) Elevated coupling above .25 Hz was calculated by taking the ratio of sum of coupling strengths above this frequency (shown in pink) against the region below .25 Hz (in blue). (B) Elevated cardiac coupling above 1.7 Hz was calculated similarly.
Figure S5. Cross‐validation of logistic regression. Fivefold cross‐validation of training seizures demonstrates the model remaining robust, with an average 84% area under the curve.
ACKNOWLEDGMENTS
B.L.B. acknowledges support from the Natural Sciences and Engineering Research Council of Canada and the SciNet HPC Consortium funded by the Canada Foundation for Innovation; the Government of Ontario; the Ontario Research Fund (Research Excellence); and the University of Toronto. NASR is supported by funding from FACES (Finding a Cure for Epilepsy and Seizures).
DATA AVAILABILITY STATEMENT
There are legal restrictions on sharing a deidentified dataset for a subset of patients, imposed by NASR, due to agreements between them and their consortium review boards. NASR accepts data access requests at the following address: North American SUDEP Registry, NYU Langone Health Comprehensive Epilepsy Center, 223 East 34th St., New York, NY 10016, USA; Email: info@sudepregistry.org. ECG recordings of additional non‐SUDEP patients with epilepsy were obtained from SSED.
REFERENCES
- 1. Aiba I, Noebels JL. Spreading depolarization in the brainstem mediates sudden cardiorespiratory arrest in mouse SUDEP models. Sci Transl Med. 2015;7(282):282ra46. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Aiba I, Noebels JL. Spreading depolarization and seizures: end of the beginning, or beginning of the end? Epilepsy Curr. 2025;25:335–340. 10.1177/15357597251328827 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Lertwittayanon W, Devinsky O, Carlen PL. Cardiorespiratory depression from brainstem seizure activity in freely moving rats. Neurobiol Dis. 2020;134:104628. [DOI] [PubMed] [Google Scholar]
- 4. Surges R, Adjei P, Kallis C, Erhuero J, Scott CA, Bell GS, et al. Pathologic cardiac repolarization in pharmacoresistant epilepsy and its potential role in sudden unexpected death in epilepsy: A case‐control study. Epilepsia. 2010;51(2):233–242. 10.1111/j.1528-1167.2009.02330.x [DOI] [PubMed] [Google Scholar]
- 5. Szurhaj W, Leclancher A, Nica A, Périn B, Derambure P, Convers P, et al. Cardiac autonomic dysfunction and risk of sudden unexpected death in epilepsy. Neurology. 2021;96(21). 10.1212/WNL.0000000000011998 [DOI] [PubMed] [Google Scholar]
- 6. Chen ZS, Hsieh A, Sun G, Bergey GK, Berkovic SF, Perucca P, et al. Interictal EEG and ECG for SUDEP risk assessment: a retrospective multicenter cohort study. Front Neurol. 2022;13:13. 10.3389/fneur.2022.858333 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7. Ramirez AD, Tanwir A, Szabo C. Cardiac biomarkers for SUDEP: retrospective evaluation of nine patients with EMU admissions (S24.003). Neurology. 2024;102(7_supplement_1). 10.1212/WNL.0000000000205126 [DOI] [Google Scholar]
- 8. Gravitis AC, Tufa U, Zukotynski K, Streiner DL, Friedman D, Laze J, et al. Ictal ECG‐based assessment of sudden unexpected death in epilepsy. Front Neurol. 2023;14:14. 10.3389/fneur.2023.1147576 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9. Guirgis M, Chinvarun Y, del Campo M, Carlen PL, Bardakjian BL. Defining regions of interest using cross‐frequency coupling in extratemporal lobe epilepsy patients. J Neural Eng. 2015;12(2):026011. 10.1088/1741-2560/12/2/026011 [DOI] [PubMed] [Google Scholar]
- 10. Grigorovsky V, Jacobs D, Breton VL, Tufa U, Lucasius C, del Campo JM, et al. Delta‐gamma phase‐amplitude coupling as a biomarker of postictal generalized EEG suppression. Brain Commun. 2020;2(2). 10.1093/braincomms/fcaa182 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Grigorovsky V, Breton VL, Bardakjian BL. Glial modulation of electrical rhythms in a neuroglial network model of epilepsy. IEEE Trans Biomed Eng. 2021;68(7):2076–2087. 10.1109/TBME.2020.3022332 [DOI] [PubMed] [Google Scholar]
- 12. Tufa U, Gravitis A, Zukotynski K, Chinvarun Y, Devinsky O, Wennberg R, et al. A peri‐ictal eeg‐based biomarker for sudden unexpected death in epilepsy (SUDEP) derived from brain network analysis. Front Netw Physiol. 2022;2. 10.3389/fnetp.2022.866540 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Feldman JL, Del Negro CA, Gray PA. Understanding the rhythm of breathing: So near, yet so far. Annu Rev Physiol. 2013;75. 10.1146/annurev-physiol-040510-130049 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Devinsky O, Hesdorffer DC, Thurman DJ, Lhatoo S, Richerson G. Sudden unexpected death in epilepsy: epidemiology, mechanisms, and prevention. Lancet Neurol. 2016;15(10):1075–1088. [DOI] [PubMed] [Google Scholar]
- 15. Bozorgi A, Lhatoo SD. Seizures, cerebral shutdown, and SUDEP: SUDEP—a perfect storm. Epilepsy Curr. 2013;13(5):236–240. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Salam MT, Montandon G, Genov R, Devinsky O, Del Campo M, Carlen PL. Mortality with brainstem seizures from focal 4‐aminopyridine–induced recurrent hippocampal seizures. Epilepsia. 2017;58(9):1637–1644. [DOI] [PubMed] [Google Scholar]
- 17. Detti P. Siena Scalp EEG Database. 2020. Available from: https://physionet.org/content/siena‐scalp‐eeg/1.0.0/
- 18. Detti P, Vatti G, de Zabalo Manrique Lara G. EEG synchronization analysis for seizure prediction: a study on data of noninvasive recordings. Processes. 2020;8(7):846. 10.3390/pr8070846 [DOI] [Google Scholar]
- 19. Devinsky O, Bundock E, Hesdorffer D, Donner E, Moseley B, Cihan E, et al. Resolving ambiguities in SUDEP classification. Epilepsia. 2018;59(6):1220–1233. 10.1111/epi.14195 [DOI] [PubMed] [Google Scholar]
- 20. Colic S, Wither RG, Lang M, Zhang L, Eubanks JH, Bardakjian BL. Prediction of antiepileptic drug treatment outcomes using machine learning. J Neural Eng. 2017;14(1):016002. 10.1088/1741-2560/14/1/016002 [DOI] [PubMed] [Google Scholar]
- 21. Garcia CAS, Gravitis AC, Bardakjian BL. Respiratory signal extraction from ECG using a phase‐amplitude cross‐frequency coupling index. In: 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC). Orlando, FL: IEEE. 2024;2024:p. 1–4. 10.1109/EMBC53108.2024.10782382 [DOI] [PubMed] [Google Scholar]
- 22. Caiola M, Devergnas A, Holmes MH, Wichmann T. Empirical analysis of phase‐amplitude coupling approaches. PLoS One. 2019;14(7):e0219264. 10.1371/journal.pone.0219264 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23. Tort ABL, Komorowski R, Eichenbaum H, Kopell N. Measuring phase‐amplitude coupling between neuronal oscillations of different frequencies. J Neurophysiol. 2010;104(2):1195–1210. 10.1152/jn.00106.2010 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24. Rosenstein MT, Collins JJ, De Luca CJ. A practical method for calculating largest Lyapunov exponents from small data sets. Physica D. 1993;65(1–2):117–134. 10.1016/0167-2789(93)90009-P [DOI] [Google Scholar]
- 25. Goldberger A. Is the normal heartbeat chaotic or homeostatic? Physiology. 1991;6(2):87–91. 10.1152/physiologyonline.1991.6.2.87 [DOI] [PubMed] [Google Scholar]
- 26. Streiner DL. Statistics commentary series: commentary no. 30: propensity scores. J Clin Psychopharmacol. 2019;39(1):9–12. 10.1097/JCP.0000000000000991 [DOI] [PubMed] [Google Scholar]
- 27. Ho J, Tumkaya T, Aryal S, Choi H, Claridge‐Chang A. Moving beyond P values: data analysis with estimation statistics. Nat Methods. 2019;16:565–566. [DOI] [PubMed] [Google Scholar]
- 28. Welch BL. The generalization of ‘student's' problem when several different population variances are involved. Biometrika. 1947;34(1–2):28–35. 10.1093/biomet/34.1-2.28 [DOI] [PubMed] [Google Scholar]
- 29. Vilella L, Lacuey N, Hampson JP, Zhu L, Omidi S, Ochoa‐Urrea M, et al. Association of peri‐ictal brainstem posturing with seizure severity and breathing compromise in patients with generalized convulsive seizures. Neurology. 2021;96(3):e352–e365. 10.1212/WNL.0000000000011274 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Mueller SG, Nei M, Bateman LM, Knowlton R, Laxer KD, Friedman D, et al. Brainstem network disruption: a pathway to sudden unexplained death in epilepsy? Hum Brain Mapp. 2018;39(12):4820–4830. 10.1002/hbm.24325 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Bauer PR, Thijs RD, Lamberts RJ, Velis DN, Visser GH, Tolner EA, et al. Dynamics of convulsive seizure termination and postictal generalized EEG suppression. Brain. 2017;140(3):655–668. 10.1093/brain/aww322 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32. Lhatoo SD, Faulkner HJ, Dembny K, Trippick K, Johnson C, Bird J. An electroclinical case‐control study of sudden unexpected death in epilepsy. Ann Neurol. 2010;68(6):787–796. [DOI] [PubMed] [Google Scholar]
- 33. Ryvlin P, Rheims S, Lhatoo S. Risks and predictive biomarkers of SUDEP. Curr Opin Neurol. 2019;32(2):205–212. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 34. Charlton PH, Bonnici T, Tarassenko L, Clifton DA, Beale R, Watkinson PJ. An assessment of algorithms to estimate respiratory rate from the electrocardiogram and photoplethysmogram. Physiol Meas. 2016;37(4):610–626. 10.1088/0967-3334/37/4/610 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35. Charlton PH, Bonnici T, Tarassenko L, Alastruey J, Clifton DA, Beale R, et al. Extraction of respiratory signals from the electrocardiogram and photoplethysmogram: technical and physiological determinants. Physiol Meas. 2017;38(5):669–690. 10.1088/1361-6579/aa670e [DOI] [PubMed] [Google Scholar]
- 36. Friedman D, Chyou J, Devinsky O. Sudden death in epilepsy: of mice and men. J Clin Invest. 2013;123(4):1415–1416. 10.1172/JCI67759 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37. Gravitis AC, Sivendiran K, Tufa U, Zukotynski K, Chinvarun Y, Devinsky O, et al. Wavelet phase coherence of ictal scalp EEG‐extracted muscle activity (SMA) as a biomarker for sudden unexpected death in epilepsy (SUDEP). PLoS One. 2024;19(8):e0298943. 10.1371/journal.pone.0298943 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38. Bruno E, Simblett S, Lang A, Biondi A, Odoi C, Schulze‐Bonhage A, et al. Wearable technology in epilepsy: the views of patients, caregivers, and healthcare professionals. Epilepsy Behav. 2018;85:141–149. 10.1016/j.yebeh.2018.05.044 [DOI] [PubMed] [Google Scholar]
- 39. Hoppe C, Feldmann M, Blachut B, Surges R, Elger CE, Helmstaedter C. Novel techniques for automated seizure registration: patients' wants and needs. Epilepsy Behav. 2015;52:1–7. 10.1016/j.yebeh.2015.08.006 [DOI] [PubMed] [Google Scholar]
- 40. Nasseri M, Stirling RE, Viana PF, Cui J, Nurse E, Karoly PJ, et al. Forecasting epileptic seizures with wearable devices: a hybrid short‐ and long‐horizon pseudo‐prospective approach. Epilepsia. 2025;66. 10.1111/epi.18466 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 41. Poh MZ, Loddenkemper T, Swenson N, Goyal S, Madsen J, Picard RW. Continuous Monitoring of Electrodermal Activity During Epileptic Seizures Using a Wearable Sensor. In: Proceedings of the 32rd Annual International Conference of the IEEE Engineering in Medicine and Biology Society. Buenos Aires: IEEE Engineering in Medicine and Biology Conference; 2010;p. 4415–4418. [DOI] [PubMed] [Google Scholar]
- 42. BIOPAC Systems . BIOPAC Research Ring Product Sheet. 2025. Accessed July 7, 2025. Available from: https://www.biopac.com/wp‐content/uploads/RESRING‐PACK.pdf
- 43. Rheims S, Sperling MR, Ryvlin P. Drug‐resistant epilepsy and mortality—why and when do neuromodulation and epilepsy surgery reduce overall mortality. Epilepsia. 2022;63(12):3020–3036. 10.1111/epi.17413 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44. Simpson HD, Schulze‐Bonhage A, Cascino GD, Fisher RS, Jobst BC, Sperling MR, et al. Practical considerations in epilepsy neurostimulation. Epilepsia. 2022;63(10):2445–2460. 10.1111/epi.17329 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 45. Nair DR, Laxer KD, Weber PB, Murro AM, Park YD, Barkley GL, et al. Nine‐year prospective efficacy and safety of brain‐responsive neurostimulation for focal epilepsy. Neurology. 2020;95(9):e1244–e1256. 10.1212/WNL.0000000000010154 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 46. Devinsky O, Friedman D, Duckrow RB, Fountain NB, Gwinn RP, Leiphart JW, et al. Sudden unexpected death in epilepsy in patients treated with brain‐responsive neurostimulation. Epilepsia. 2018;59(3):555–561. 10.1111/epi.13998 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Figure S1. Validation of respiratory range. Phase–amplitude couplings (PACs) were calculated for a standard dataset containing concurrent electrocardiogram and impedance pneumography (IP). The frequency of the PAC feature with strongest coupling was compared to the instantaneous frequency of the IP signal. Each marker represents one test subject (n = 28). This validated that respiratory rate was correlated with the phase frequency of maximum PAC coupling for subjects at rest.
Figure S2. Subdivision of phase–amplitude coupling (PAC) comodulogram into nine features. (A) PAC comodulogram calculated using respiratory frequencies for phase (x‐axis), and cardiac frequencies for amplitude (y‐axis). (B) The means of each of the nine PAC subdivisions form a 3 × 3 grid for analysis of the window.
Figure S3. Summary of cardiorespiratory coupling results. Average phase–amplitude coupling (PAC) mean and SD by seizure were calculated for all generalized seizures (19 non‐sudden unexpected death in epilepsy [SUDEP], 18 SUDEP; left) and nongeneralized seizures (19 non‐SUDEP, 20 SUDEP; right). SUDEP results reveal a pattern of elevated PAC means at higher frequencies, indicative of reduced dynamism and decreased system complexity.
Figure S4. Relative phase–amplitude coupling strength. (A) Elevated coupling above .25 Hz was calculated by taking the ratio of sum of coupling strengths above this frequency (shown in pink) against the region below .25 Hz (in blue). (B) Elevated cardiac coupling above 1.7 Hz was calculated similarly.
Figure S5. Cross‐validation of logistic regression. Fivefold cross‐validation of training seizures demonstrates the model remaining robust, with an average 84% area under the curve.
Data Availability Statement
There are legal restrictions on sharing a deidentified dataset for a subset of patients, imposed by NASR, due to agreements between them and their consortium review boards. NASR accepts data access requests at the following address: North American SUDEP Registry, NYU Langone Health Comprehensive Epilepsy Center, 223 East 34th St., New York, NY 10016, USA; Email: info@sudepregistry.org. ECG recordings of additional non‐SUDEP patients with epilepsy were obtained from SSED.