Multiday cycles of heart rate modulate seizure likelihood at daily, weekly and monthly timescales: An observational cohort study

Tracking Seizure Cycles

Tracking Seizure Cycles is a non-interventional cohort study to measure cycles of seizure likelihood from seizure diaries, identify corresponding cycles from non-invasively measured physiological signals and quantify the relationship between physiological cycles and seizure occurrence. The study was approved by the St Vincent’s Hospital Human Research Ethics Committee (HREC 009.19) and first enrolment was in August 2019. Participants were over 18 with a confirmed epilepsy diagnosis and uncontrolled or partially controlled seizures. All participants provided written informed consent. Here, we report on one-year study outcomes.

Continuous data were collected via mobile and wearable devices for at least two months and up to two years. Participants wore a smartwatch (Fitbit) and manually reported seizure times in a freely available mobile diary app (Seer App). The smartwatch continuously measured participants’ heart rates (via photoplethysmography) at 5s resolution.

There were 46 participants (29 female) with a cumulative total of 143,142 hours (19,880 hours per person, mean M = 3,491, standard deviation SD = 2,880) of continuous heart rate recorded. Mean device adherence (at least one recording every hour) among participants was 66% (SD = 36%). The total diary duration across participants was 53 years (193,429 days per person, M = 464, SD = 543). Participant diaries included over 2,974 seizures (M = 71, SD = 80) with 1,818 seizures (M = 44, SD = 70) reported during the wearable monitoring period.

Multiday heart rate cycles

To evaluate the strength of heart rate cycles, participants were required to have at least two months of recordings with over 80% adherence (i.e., less than 20% missing data), with adherence defined as at least one heart rate recording every hour. Eligible participant demographics are given in Table 1 and Supplementary Table 1.

Cycles were measured at different periods using a wavelet transform approach that was previously proposed for the detection of multiday cycles of epileptic activity.⁵ The continuous heart rate signal was first down-sampled to one timestamp every minute. Linear interpolation was performed for up to 1 hour either side of a missing segment. Longer recording gaps of up to 3 days were interpolated with a straight line at the average value of all the data (see Appendix 1).

Candidate cycle periods were then tested using a continuous Morlet wavelet transform, where significant peaks in the global wavelet power were determined to be significant heart rate cycle periods (see Appendix 1). Heart rate cycle periods were considered to be circadian, weekly or monthly if they were within ±33.3% of 24 hours, 7 days or 30 days, respectively (consistent with the approach used by Baud et al. ⁵). However, all circadian cycles were detected at exactly 24 hours. Multiday cycles were any other periods longer than 32 hours and ultradian cycles were any other periods shorter than 16 hours.

Relationship between seizures and heart rate cycles

To evaluate the strength of co-modulation between seizure occurrence and heart rate cycles, participants were required to have at least five reported seizures (see Table 1). Co-modulation was measured from the degree of ‘phase-locking’ of seizure times with respect to underlying heart rate cycles, i.e. how tightly seizure times were linked a particular point (phase) of the cycle. Phase-locking was quantified by the synchronisation index (SI, also known as a phase-locking value or R-value). The amplitude of SI ranges from 0 to 1, where 0 indicates a uniform distribution (i.e., no relationship) and 1 indicates perfect alignment with respect to an underlying cycle. The angle or direction of SI indicates the preferred phase of seizure occurrence; for instance, seizures could be more likely near the peak or trough of average heart rate cycles.

To extract heart rate cycles, the standardised heart rate signal was bandpass filtered into distinct component frequencies matching the significant cycle frequencies (identified from wavelet decomposition). The bandpass filter applied at each significant cycle was a second-order zero-phase Butterworth bandpass filter with cut-off frequencies at ±33.3% of the cycle frequency. For instance, someone with significant cycles (wavelet spectrum peaks) at 24 hours, 9 days and 30 days would have three bandpass filters applied with cut-off frequencies of 16-32 hours, 6-12 days and 20-40 days, respectively. These cut-off frequencies were chosen to account for phase shifts in the cycle over the recording time.

The continuous phase of each bandpass filtered signal was estimated using the Hilbert transform. The times of seizure occurrence were mapped to the estimated phase of heart rate cycles. Seizure phases were then binned into 24 (circadian cycles) or 18 (all other periods) equal sized bins (ranging from 0 to 2π) to produce a phase distribution. The SI values (amplitude and angle) were computed from these phase distributions. The Omnibus (also called Hodges-Ajne) test was used to determine whether seizures were significantly (p<0.05) phase-locked to the heart rate cycle by testing the null hypothesis that the phase distribution was uniform.

Ambulatory video-EEG-ECG analysis

Retrospective analysis of ECG used a separate cohort of people undergoing at-home, ambulatory video-EEG-ECG diagnostic testing for epilepsy (Seer Medical). The study was approved by the St Vincent’s Hospital Human Research Ethics Committee (LRR 165/19). All participants provided written informed consent for their de-identified data to be used for subsequent research. This study included 29 participants’ records of at least 8 days (M = 9.7 days, SD = 1.3 days) duration. Note that as participants were undergoing diagnostic monitoring, they did not necessarily have an epilepsy diagnosis. Participant demographics and AEDs (if used) are given in Supplementary Table 2.

The current study analysed continuous ECG data, combined with epileptic event labels derived from video-EEG. To derive heart rate from ECG recordings, r-peak detection was performed and the number of beats per minute was computed with a 5 s sliding window. The peak-to-peak intervals (RR interval) were also analysed. Epileptic events were labelled using computer-assisted review, whereby event detection was first performed by a machine learning algorithm and then suspected events were reviewed and confirmed by expert neurophysiology and neurology review (see Appendix 3).

Given the limited duration of ECG studies it was not possible to determine whether weekly cycles were present. Instead, significant linear or quadratic trends were identified in the smoothed heart rate or RR interval signals. Heart rate and RR interval were smoothed using a 1-day moving average filter. Linear regression was then performed for a simple linear and quadratic model (see Appendix 3). The rationale for this approach was that the ECG might be recorded around the peak or trough of an about-weekly cycle (showing a curved or quadratic trend) or on the rising/falling slope of a cycle (showing a linear increase/decrease).

Tracking seizure cycles

There were 26 out of 39 participants with at least two months of data (range 2.1-13.9 months, M = 6.4, SD = 4.0) and at least 80% (M = 95%, SD = 5%) adherence. Among these 26 participants (Table 1), 17 had recorded at least 5 seizures during the wearable recording time and, therefore, were eligible for further seizure analysis.

Heart rate shows daily, weekly and monthly cycles

We first investigated whether multiday cycles of average heart rate existed. Figure 1 shows an example of heart rate cycles for two people. Multiday rhythms of heart rate were evident over about-weekly (Fig 1e) and about-fortnightly (Fig 1b) periods. Wavelet analysis confirmed significant cycles at daily (24 h), fortnightly (15 d) and monthly (27 d) periods for P39 (Fig 1c), and daily (24 h), weekly (8.5 d) and slower (43.5 d) periods for P11 (Fig 1f). These examples provide striking evidence that multiday cycles of cardiac activity exist over weekly and monthly timescales. Cycles can be seen from visual inspection of average heart rate and were robust over months (Fig 1a,d).

• Download figure
• Open in new tab

Figure 1. Examples of multiday heart rate cycles.

Data are shown for two different participants, P39 (a-c) and P11 (d-f). a,d: Heart rate (y-axis) smoothed with a 2-day moving average filter shows multiday cycles. Insets (blue) show circadian rhythms of heart rate. b,e: Bandpass filtered heart rate signals for different cycles (corresponding respectively to panels c,f). c,f: Wavelet power spectra for different scales (x-axis). Significant cycle periods are labelled.

Figure 2 shows the prevalence of multiday heart rate cycles across the cohort. It can be seen that cycles exist at weekly, fortnightly, monthly and longer timescales. Unsurprisingly, the average cycle strength across the cohort (Fig 2a) showed a clear peak at 24 hours, with significant circadian cycles detected for 100% of the cohort. More interesting were peaks in cycle strength around 7 days and 30 days, demonstrating that subsets of people have similar heart rate cycles at these periods (see Supplementary Figure 2 for individual analyses). Notably, multiday heart rate cycles were as prevalent as circadian cycles. We found that 23 people (88%) showed an about-weekly cycle and 24 people (92%) had about-monthly cycles. Apart from these periods, 26 people (100%) had a multiday cycle (> 32 hours) other than weekly or monthly and 19 people (73%) had a shorter ultradian cycle (< 16 hours).

• Download figure
• Open in new tab

Figure 2. Distribution of heart rate cycles.

a: Cycle strength (expressed as the normalised wavelet power, y-axis) for different periods (x-axis, logarithmic scale) averaged across the cohort. b: Raster plot showing cycle strength (colour bar) for each individual (y-axis) at different periods (logarithmic scale). c,d: Number of people (y-axis) with significant cycles at different periods up to 40 days (x-axis, logarithmic scale) for men and women, respectively. Note that the c,d x-axis (up to 40 days) is a subset of the a,b x-axis (up to 208 days) as indicated by the brace.

The distributions of cycles were similar for men and women (Fig 2c and d). Everyone showed daily cycles, and most showed about-weekly cycles. Of the people with about-monthly cycles, five were men (62% of all eligible male participants) and 16 were women (94% of females). Note that only seven (27%) of the about-weekly cycles were found at precisely 7 days, suggesting that, in general, cycles were not related to a particular day of the week. Three of these seven people (12% of total cohort) had seizures significantly locked to their 7-day heart rate cycle. Interestingly, this was consistent with a previous finding that 21% of people have a precise 7-day seizure cycle.⁴ None of the about-monthly cycles were found at 29.5 days (the lunar period).

Multiday heart rate cycles modulate seizure risk

We investigated whether multiday heart rate cycles and seizure likelihood were comodulated by analysing whether self-reported seizure times were significantly phase-locked to the underlying heart rate cycles (quantified by SI).

Individuals with more than five seizures were included in the analysis, resulting in a total of 17 (7 male). Of these, 13 of 17 (76%) had seizures significantly locked onto at least one heart rate cycle. Among participants with observed seizure modulation, 69% (9 of 13) had both fast (circadian or ultradian) and multiday cycles comodulated with seizure likelihood, 77% (10 of 13) had seizures significantly locked onto a multiday cycle and 92% (12 of 13) had seizures significantly locked onto a circadian or ultradian cycle. Of the total cohort of eligible participants, 65% had seizures significantly locked onto a circadian cycle, 35% had seizures significantly locked onto a weekly cycle and 29% had seizures significantly locked onto a monthly cycle.

Figure 3 shows individual examples of the occurrences of seizures with respect to multiday cycles of heart rate. Different participants’ seizures were strongly synchronised with their unique multiday heart rate cycle. Synchronisation may be stronger for multiday cycles than circadian cycles; for instance, P11 showed SI = 0.55 for their 8.5 day cycle and SI = 0.40 for their circadian cycle.

• Download figure
• Open in new tab

Figure 3. Examples of seizure occurrences locked to heart rate cycles.

Heart rate (y-axis) and self-reported seizures (dots) for two participants, a,c: P11 and e,g: P39. b,d,f,h: Corresponding circular histograms showing the phase distributions of seizures with respect to significant heart rate cycles. Note that multiday histogram bins have the same phase width (2π/18) although these correspond to different durations (labelled). The circadian histogram (b) bins have width of 1 hour (2π/24). Heart rate is shown after applying a moving average filter to highlight cycles (a uses a 1-hour MA, c,e use a 2-day MA and g uses a 7-day MA).

Figure 4 shows the SI vectors (magnitude and preferred phase) for every participant with significant seizure phase locking across different heart rate cycles. It can clearly be seen that seizures preferentially occurred on the rising phase of multiday cycles (weekly, monthly and other multiday periods, Figs 4b-d) for most people. In other words, increasing heart rate was linked to increased seizure likelihood.

• Download figure
• Open in new tab

Figure 4. Phase locking of seizures to heart rate cycles.

Each subplot shows individuals (arrows) with significant phase locking of seizure occurrence to their heart rate cycle. The lengths of the arrows indicate the strength of phase locking (radial axis, between 0 and 1), while the direction indicates the preferred phase (polar axis). a: Circadian cycle, all periods were 24 hours. b: About-weekly cycle, 6-8 day periods (colour bar). c: About-monthly cycle, 21-37 day periods (colour bar). d: Multiday cycle, 2-121 day periods (colour bar).

For circadian cycles, phase locking was also compared to clock time (Supplementary Figure 3), and a similar distribution was found. Across the cohort, 74% of seizures were reported during the day (between 8AM and 8PM), suggestive of a reporting bias that could affect phase locking of self-reported events to circadian cycles. However, it is important to note that a diurnal reporting bias would not affect phase-locking to multiday cycles.

Circadian cycles and weekly trends in continuous video-EEG-ECG

We investigated weekly modulation of heart rate and the cyclic co-modulation of heart rate and epileptic activity using continuous video-EEG-ECG records from a cohort of people undergoing diagnostic monitoring for epilepsy. Figure 5 shows example studies where significant weekly trends were observed in either heart rate or RR interval signals measured from continuous ECG. Trends differed between individuals, including linearly increasing (Fig 5a,d), decreasing (Fig 5b,e) or quadratic (Fig 5c,f) trends in average heart rate and RR interval, suggesting patient-specific modulation over the weeklong monitoring.

• Download figure
• Open in new tab

Figure 5. Example weekly trends identified from ambulatory ECG monitoring.

Black lines show average heart rate (a-c) and RR interval (d-f) and red lines show smoothed data (1-day moving average). Significant trends were identified from the smoothed data using logistic regression. a,d: Increasing linear trends of heart rate and RR interval. b,e: Decreasing linear trends of heart rate and RR interval. c,f: Curved (quadratic) trends of heart rate and RR interval. Insets show r-squared values and p-values for the linear (p₁) and quadratic (p₂) components of the models.

Across the EEG-ECG cohort, linear or quadratic trends were found for average heart rate from 15 out of 29 (52%) studies, with 9 studies showing a significant (p < 0.05) linear component and 11 showing a significant quadratic component. For RR interval, 14 out of 29 (48%) studies showed significant trends (10 linear and 5 quadratic). Exact p-values and R² statistics for linear and quadratic models of heart rate and RR intervals are given in Supplementary Tables 3 and 4. Although it was not possible to determine whether these trends were associated with about-weekly cycles, these findings support the observation that participants’ average heart rates were modulated over weekly timescales.

In addition to weekly trends, we investigated co-modulation of circadian cycles of heart rate and RR intervals with epileptic events (see Appendix 3). Both epileptic activity and self-reported seizures were phase locked to 24-hour cycles of heart rate and RR intervals; however, this relationship could not be differentiated from the relationship between epileptic events and time of day (Supplementary Figure 4 and Supplementary Table 5).

Mechanisms of multiday cycles in the heart and brain

Multiday cycles of epileptic activity are well documented,²⁴ and have been related to a progressive loss of resilience as the brain approaches a tipping point,⁶ corresponding to an increased susceptibility to seizures. Multiday heart rate cycles showed strikingly similar features to these cycles of epileptic brain activity, with a clear preference for seizures to occur on the rising phase of heart rate cycles (Fig 4), which is also a property of epileptic rhythms.^5,6 This relationship suggests that increasing heart rate is linked to increased seizure likelihood. In line with this, VNS treatment is generally found to decrease heart rate, albeit on a shorter time scale,²⁵ consistent with cycle periods where seizures were also less likely. More generally, increased heart rate is associated with increased sympathetic dominance, which has also been linked to epilepsy.²⁶ Nevertheless, the current study cannot elucidate whether co-modulated cycles of heart rate and seizures are causally linked or reflect two independent, correlated physiologies.

In contrast to epileptology, the existence of multiday cycles of heart rate has not been widely described in the cardiology literature, although some previous studies have documented weekly rhythms. One study of 86 premature infants found endogenous about-weekly (between 5-10 days) rhythms of either heart rate or blood pressure for 100% of the cohort (monitoring over 14-45 days).¹⁶ Other studies in adults have suggested 7-day modulation governs heart rate and blood pressure,^15,27 although these have been predominately limited to weeklong recordings, which cannot convincingly demonstrate multiday periodicity. These previous studies suggest that heart rate cycles also exist for people without epilepsy. The current work builds on earlier results and highlights the importance of longer recording periods, which reveal heart rate cycles at multiple timescales, including weekly, fortnightly and monthly cycles (Fig 2).

The causal mechanisms of multiday epileptic rhythms or heart rate cycles are not understood. In epilepsy, both circadian and multiday rhythms show an endogenous component, with seminal studies in animal models^10,28 and human subjects under continuous dim-light conditions²⁹ revealing free-running circadian^28,29 and multiday cycles.^5,9 In the current study, participants maintained their usual day-to-day routines, although the lack of obvious environmental cues suggests multiday heart rate cycles also had some intrinsic biological drivers. However, there remains no clear explanation for the prevalence of brain/heart cycles with weekly or monthly periodicities.

Many candidate hormonal, metabolic and other environmental factors have been hypothesised to play some role in generating endogenous multiday cycles in humans.^24,27 Immune proteins are known to fluctuate with approximate weekly periodicity in humans, although the causes of this cycle are unknown.³ Female sex hormones have been widely linked to monthly cycles of epileptic activity in women³⁰ and the menstrual cycle may also lead to changes in heart rate and HRV.³¹ However, hormonal changes cannot explain the similar prevalence of monthly cycles of seizures in children and men,^4,5,7 nor the present results demonstrating monthly cycles of heart rate in male subjects (Fig 2c). Chronic stress can promote autonomic imbalance affecting both heart rate³² and epileptic brain activity³³, and stress is also perceived as one of the leading factors triggering seizures³⁴. However, stress is not typically considered to follow regular, multiday cycles. In order to elucidate the drivers of co-modulated cardiac and epileptic activity, future studies should target a range of candidate mechanisms for long-term monitoring.

Clinical applications

Heart rate cycles can provide a biomarker for individual seizure likelihood, which may be used to guide epilepsy therapy or in seizure forecasting systems. Seizure forecasting is considered a key goal of epilepsy treatment, restoring a degree of control to people with refractory seizures; however, most forecasting algorithms have been deployed for intracranial EEG, which poses challenges for widespread availability of seizure warning systems.³⁵ Importantly, forecasting algorithms based on multiday cycles have emerged as the most accurate, recently surpassing all previous approaches on a benchmark human dataset⁶ as well as showing excellent performance for larger cohorts,³⁶ and using data from seizure diaries.²³ If heart rate can be shown to reliably track multiday cycles of seizure likelihood, then similarly powerful seizure forecasts may eventually be derived from wearable devices.

In addition to seizure forecasting, monitoring heart rate cycles may provide insight into the long-term role of heart rate and HRV in risk of SUDEP. There are stereotypical cardiorespiratory changes prior to the occurrence of SUDEP³⁷, with electrocardiographic abnormalities and impaired autonomic control associated with a higher risk of SUDEP.³⁸ Long-term ECG monitoring may be a predictive of SUDEP in people with epilepsy.^39,40 However, despite circadian modulation of SUDEP risk - with substantially higher incidence at night³⁷ - longer rhythms have not yet been investigated. The prevalence of co-modulated heart rate and seizure cycles makes heart rate cycles a feasible biomarker for the risk of both seizure occurrence and SUDEP.

Multiday heart rate cycles have implications for cardiology. The relevance of circadian cycles to cardiovascular disease has long been recognised. There is increased incidence of acute myocardial infarction and sudden cardiac death in the morning.⁴¹ Ischemia and stroke also show a clear morning peak, related to circadian changes in blood pressure and cardiac output.⁴² Circadian patterns are observed across most arrhythmic events regardless of whether an underlying heart condition is present.²² In addition to circadian rhythms of cardiovascular disease, seven-day and seasonal patterns have been documented. Multiple studies show a Monday peak in cardiac mortality and hospital admissions for cardiovascular disease.¹⁷ Although it is hypothesised that these weekly patterns may be related to endogenous variation in cardiac output,²⁷ they could also be linked to environmental factors.¹⁷ Seasonal peaks of cardiac mortality and hospitalisation are generally observed in winter, primarily attributed to environmental factors.² The existence of endogenous weekly and other multiday heart rate cycles suggest that incidence of cardiovascular disease should be explored in relation to underlying individual cycle periods to investigate potential high-risk phases for common dysfunctions. Just as circadian regulation leads to danger times for people with heart conditions, it is possible that risk factors for cardiac mortality also fluctuate over multiday timescales.

Future work

The current study was based on self-reported seizure times, so noise or reporting bias could influence results. We had previously demonstrated that multiday cycles established from seizure diaries align with cycles recorded from true electrographic seizures.²³ Nevertheless, it will be critical to validate the co-modulation of heart rate cycles and epileptic activity using chronic EEG recordings. Future work will extend these results to validate the existence of multiday heart rate cycles and their relationship to electrographic seizures using ultra long-term EEG and additional wearable physiological sensors. For now, it is our hope that the prospect of multiday cycles governing diverse physiological systems leads to new breakthroughs in understanding biological rhythms and treatment of disease.