OBJECTIVE:Novel subcutaneous electroencephalography (sqEEG) systems enable prolonged, near-continuous cerebral monitoring in real-world conditions. Nevertheless, the feasibility, acceptability and overall clinical utility of these systems remain unclear. We report on the longest observational study using ultra-long-term sqEEG to date. METHODS:We conducted a 15-month prospective, observational study including 10 adult people with treatment-resistant epilepsy. After device implantation, patients were asked to record sqEEG, to use an electronic seizure diary, and to complete acceptability and usability questionnaires. sqEEG seizures were annotated visually, aided by automated detection. Individualized temporal patterns of seizure occurrence were assessed via circadian circular statistics and via Fano factor analysis. RESULTS:Over a median duration of 438 days, 10 patients recorded a median 18.8 h/day, totaling 71 984 h of real-world sqEEG data. Adherence and acceptability remained high throughout the study. Although 754 sqEEG seizures were recorded across patients, more than half (52%) of these were not reported in the patient diary. Of the 140 (27%) diary reports not associated with an identifiable sqEEG seizure, the majority (68%) were reported as seizures with preserved awareness. The sqEEG to diary F1 agreement score was highly variable, ranging from .06 to .97. Patient-specific patterns of circadian seizure occurrence and seizure clustering were found, including several relevant discrepancies between sqEEG and diary. SIGNIFICANCE:We demonstrate feasibility and high acceptability of ultra-long-term (months-years) sqEEG monitoring. These systems help provide real-world, more objective seizure counting compared to patient diaries. It is possible to objectively monitor individual temporal fluctuations of seizure occurrence.

OBJECTIVE:Seizure unpredictability can be debilitating and dangerous for people with epilepsy. Accurate seizure forecasters could improve quality of life for those with epilepsy but must be practical for long-term use. This study presents the first validation of a seizure-forecasting system using ultra-long-term, non-invasive wearable data. METHODS:Eleven participants with epilepsy were recruited for continuous monitoring, capturing heart rate and step count via wrist-worn devices and seizures via electroencephalography (average recording duration of 337 days). Two hybrid models-combining machine learning and cycle-based methods-were proposed to forecast seizures at both short (minutes) and long (up to 44 days) horizons. RESULTS:The Seizure Warning System (SWS), designed for forecasting near-term seizures, and the Seizure Risk System (SRS), designed for forecasting long-term risk, both outperformed traditional models. In addition, the SRS reduced high-risk time by 29% while increasing sensitivity by 11%. SIGNIFICANCE:These improvements mark a significant advancement in making seizure forecasting more practical and effective.

BACKGROUND:Multiday cyclic patterns underlying the timing of seizures are well-established in adults with epilepsy. However, longer-term patterns underpinning these models are yet to be explored extensively in pediatric cohorts. This study aims to identify and compare multiday seizure cycles between pediatric and adult cohorts, followed by a preliminary validation of cycle-based methods for estimating seizure likelihood in a pediatric cohort. METHODS:Multiday seizure cycles were extracted retrospectively from 325 (71 pediatric) electronic seizure diary users with confirmed epilepsy. Cycles were grouped (k-means clustering) and seizure cycles quantified (synchronisation index) with significant cycles identified (Rayleigh test (p < 0.05)). Wilcoxon rank-sum test assessed differences in prevalence and strength of cycle groups between pediatric and adult cohorts. The accuracy of cycle-based models to track pediatric seizure occurrence was calculated from the receiver operating characteristic (area under the curve; AUC) comparing estimated cycles to shuffled surrogate data and further validated with a moving average model. FINDINGS:30,019 seizures (pediatric: Median = 51, IQR (Q1 = 30, Q3 = 115), Range (21-661), adult: Median = 46, IQR (Q1 = 31, Q3 = 93), Range (20-1112) were analysed and seizure cycles grouped across circadian (0.5-1.5 days), about-weekly (2-12 days), about-fortnightly (13-22 days) and about-monthly (23-32 days) periodicities. Significant cycles were identified in each cycle group, with no differences in prevalence or cycle strength between pediatric and adult cohorts. Estimated cycles showed a reliable assessment of observed seizure occurrence (significantly (p < 0.05) better performance compared to random models for 88% (44 of 50) and moving average models for 50% (25 of 50) of observed daily seizure occurrence). SIGNIFICANCE:Multiday seizure cycles estimated from seizure diaries present a viable model for identifying longer-term seizure patterns in a pediatric cohort. Knowledge of these individual seizure cycles has potential to reduce the unpredictability of seizure timing and inform clinical decision-making.

OBJECTIVE:Seizures, or seizure-like events, can indicate over 50 neurological disorders, and interpretation of patient-reported symptoms is subjective, leading to misdiagnosis. This study aimed to determine diagnostic utility of multi-choice questions to assess seizures in a clinical setting. PATIENTS & METHODS:This retrospective study of Mayo Clinic neurology patients (2016 - 2021) analysed de-identified Electronic Health Records from adults with a paroxysmal disorder who completed Mayo Clinic's 'Epilepsy Pre-screening Questionnaire'. Diagnostic groups were epilepsy, functional neurological disorder (FND), syncope, and other neurological disorders. The main outcome was group-level separability of survey responses between diagnostic groups (assessed from response proportions using a chi-squared test with Holm-Sidak correction for multiple comparisons). RESULTS:4,130 records were included for epilepsy (n = 3,194), FND (n = 214), syncope (n = 101), other (n = 621). Seven multichoice questions provided significant separability between diagnostic groups, with event duration, specific warnings and symptoms providing greatest diagnostic utility. Those with syncope were significantly more likely to report short events, of a few seconds, compared to every other group (syncope: 27 %, other: 11 %, epilepsy: 9 %, FND: 1 %). People with FND were also significantly more likely to report events > 7 min (27 %), compared to those with epilepsy (11 %). Conversely, events lasting < 1 min were still reported by 23 % of people with FND and long seizures (>7 mins) were reported by almost 10 % of people with epilepsy. Pre-seizure warnings were similar for focal and generalised epilepsy, contradicting the assumption that auras are mainly associated with focal epilepsy. CONCLUSION:The results validate the diagnostic utility of general questions about event duration, related symptoms and comorbidities in a meaningfully sampled patient cohort that was representative of people presenting with seizures or seizure-like events presumed to be epilepsy (i.e. not people presenting with an obvious diagnosis of non-epileptic seizures). Findings could inform development of standardised, patient-reported surveys to aid differential diagnosis across common neurological conditions.

PURPOSE:This study aims to assess the diagnostic yield of routine EEG (rEEG) followed by long-term ambulatory EEG (aEEG) in a retrospective cohort, focusing on the rates of abnormal EEG findings, and overall event capture. METHODS:Data were retrospectively collected from deidentified clinical reports of patients who underwent both rEEG and subsequent aEEG, with both modalities including video recordings. The study included 95 patients, with demographic, clinical information, and EEG findings extracted for analysis. Statistical analyses included chi-squared proportion tests and Wilcoxon rank-sum tests to assess the influence of variables such as age, sex, referral source, and aEEG duration on outcomes. Bayes factors were calculated to evaluate the power of the statistical tests. RESULTS:Among the 95 patients, 33 % were 16 years old or younger. The median duration of aEEG was 3.9 days. Abnormal EEG findings increased from 18 % with rEEG to 33 % with aEEG. Epileptic seizures were captured in 3 % of rEEG and 8 % of aEEG, while non-epileptic events were captured in 35 % of aEEG compared to none in rEEG. Younger age was associated with higher rates of abnormal findings, but this was not adequately powered. Females had a higher likelihood of event capture on aEEG, though this finding was also underpowered. The majority of adult and paediatric patients with a normal rEEG went on to have a normal aEEG. CONCLUSION:Ambulatory EEG significantly improves the diagnostic yield for both epileptic and non-epileptic events compared to routine EEG, particularly in adults. This study supports the broader use of aEEG for comprehensive epilepsy evaluation and suggests further research to optimise its clinical utility.

ObjectiveOver recent years, there has been a growing interest in exploring the utility of seizure risk forecasting, particularly how it could improve quality of life for people living with epilepsy. This study reports on user experiences and perspectives of a seizure risk forecaster app, as well as the potential impact on mood and adjustment to epilepsy.MethodsActive app users were asked to complete a survey (baseline and 3-month follow-up) to assess perspectives on the forecast feature as well as mood and adjustment. Post-hoc, nine neutral forecast users (neither agreed nor disagreed it was useful) completed semi-structured interviews, to gain further insight into their perspectives of epilepsy management and seizure forecasting. Non-parametric statistical tests and inductive thematic analyses were used to analyse the quantitative and qualitative data, respectively.ResultsSurveys were completed by 111 users. Responders consisted of “app users” (n = 58), and “app and forecast users” (n = 53). Of the “app and forecast users”, 40 % believed the forecast was accurate enough to be useful in monitoring for seizure risk, and 60 % adopted it for purposes like scheduling activities and helping mental state. Feeling more in control was the most common response to both high and low risk forecasted states. In-depth interviews revealed five broad themes, of which ‘frustrations with lack of direction’ (regarding their current epilepsy management approach), ‘benefits of increased self-knowledge’ and ‘current and anticipated usefulness of forecasting’ were the most common.SignificancePreliminary results suggest that seizure risk forecasting can be a useful tool for people with epilepsy to make lifestyle changes, such as scheduling daily events, and experience greater feelings of control. These improvements may be attributed, at least partly, to the improvements in self-knowledge experienced through forecast use.

Forecasting events in multichannel electroencephalographic (EEG) brain recordings remains a formidable task given the noise and complexity in neural systems. Here we compare two dynamical systems motivated approaches to forecasting brain events. The first follows previous state-of-the-art (SOTA) research of time-series features of critical slowing down (autocorrelation, variance) as biomarkers of impending events. The second involves a novel long-term-short-term (LSTM) neural network-based filter to estimate the neurophysiological feature variables of mathematical neural population models of the EEG. Previous critical slowing research presented forecasting results for the best EEG channel, however, in practice the best channel cannot be known a priori. Therefore, here we also consider forecasting by combining the different features across the different EEG channels using logistic regression. One application area where forecasting brain events is important is epileptic seizure prediction. Epileptic seizures are debilitating events and up to 50 million people worldwide with drug-resistant epilepsy could benefit by receiving warnings of impending seizures. Here we apply the above methods to a long-term epileptic seizure prediction dataset from 15 patients. It was found that seizure forecasting with (1) logistic regression and critical slowing features, (2) logistic regression and neurophysiological features, and (3) the best channel using critical slowing features, respectively, achieved median sensitivities of 70, 54 and 67% and median time in low seizure risk of 84, 84, and 81%. This indicates that a multichannel model approach can perform as well as the best channel approach, removing the need to find the best channel. It also suggests neurophysiological features could be used to increase time in low risk. Future work exploring other features, machine learning models and their various combinations could yield further improvements.

A significant challenge of video-electroencephalography (vEEG) in epilepsy diagnosis is timing monitoring sessions to capture epileptiform activity. Given the significant consequences of misdiagnosis or delayed diagnosis, new techniques to improve diagnostic yield of vEEG are needed. In this study, we introduce and validate “pro-ictal EEG scheduling”, a method to schedule vEEG monitoring to coincide with periods of heightened seizure probability as a low-risk approach to enhance the diagnostic yield. A database of long-term ambulatory vEEG monitoring sessions ( n =5038) of adults and children was examined. Data from linked electronic seizure diaries were extracted (minimum 10 self-reported events over 12-months) to generate cycle-based estimates of seizure risk. VEEG monitoring sessions coinciding with periods of estimated high-risk were allocated to the high-risk group (adults n =305, children n =82) and compared to remaining studies (baseline: adults n =3586, children n =1065). Test of Proportions and Risk-Ratios (RR) were used to index differences in proportions and likelihood of capturing outcome measures (abnormal report, confirmed seizure and diary event) during monitoring. The impact of clinical and demographic factors (sex, epilepsy-type, medication) was also explored. During vEEG monitoring, the high-risk group was 25% more likely to have an abnormal vEEG report (190/305:62.3% vs 1790/3586:49.9%, RR=1.25, 95% CI[1.137:1.370], p <0.001), 63% more likely to present with a confirmed seizure (56/305:18.4% vs 424/3586:11.3%, RR=1.63, 95% CI[1.265:2.101], p <0.001) and 42% more likely to report an event (153/305:50.2% vs 1267/3586:35.3%, RR=1.420, 95% CI[1.259:1.602], p <0.001). In children, the high-risk group was 93% more likely to have a confirmed seizure (21/82:25.6% vs 141/1065:13.2%, RR=1.93, 95% CI[1.297:2.885], p= 0.002). Similar effects were observed across clinical and demographic features. This study provides the first large-scale validation of pro-ictal EEG scheduling in improving the yield of vEEG. This innovative approach offers a pragmatic and low-risk strategy to enhance the diagnostic capabilities of vEEG monitoring, significantly impacting epilepsy management.

Objective:Novel subcutaneous electroencephalography (sqEEG) systems enable prolonged, near-continuous cerebral monitoring in real-world conditions. Nevertheless, the feasibility, acceptability and overall clinical utility of these systems remains unclear. We report on the longest observational study using ultra long-term sqEEG to date. Methods:We conducted a 15-month prospective, observational study including ten adult people with treatment-resistant epilepsy. After device implantation, patients were asked to record sqEEG, to use an electronic seizure diary and to complete acceptability and usability questionnaires. sqEEG seizures were annotated visually, aided by automated detection. Seizure clustering was assessed via Fano Factor analysis and seizure periodicity at multiple timescales was investigated through circular statistics. Results:Over a median duration of 438 days, ten patients recorded a median 18.8 hours/day, totalling 71,984 hours of real-world sqEEG data. Adherence and acceptability remained high throughout the study. While 754 sqEEG seizures were recorded across patients, over half (52%) of these were not reported in the patient diary. Of the 140 (27%) diary reports not associated with an identifiable sqEEG seizure, the majority (68%) were reported as seizures with preserved awareness. The sqEEG to diary F1 agreement score was highly variable, ranging from 0.06 to 0.97. Patient-specific patterns of seizure clustering and seizure periodicity were observed at multiple (circadian and multidien) timescales. Interpretation:We demonstrate feasibility and high acceptability of ultra long-term (months-years) sqEEG monitoring. These systems help provide real-world, more objective seizure counting compared to patient diaries. It is possible to monitor individual temporal fluctuations of seizure occurrence, including seizure cycles.

The risk of seizures in epilepsy fluctuates in cycles with multiday periodicity. The strength of these patient-specific seizure risk cycles can be modulated by disease processes. There is a lack of computational models of epilepsy that describe the progression and modulation of multiday seizure risk cycles. We developed a state space model (SSM) for epilepsy progression that learns individualized multiday seizure risk cycles from intracranial EEG (iEEG) data. To capture the cyclical nature of seizure risk, our model incorporated cyclical dynamics by using a special rotation matrix structure for the state transition matrix. The model learned patient-specific multiday cycles using a novel expectation-maximization algorithm. We evaluated the model on real-world data from one of the longest continuous iEEG recordings in people with epilepsy. The model forecast iEEG and inferred periods of heightened risk of seizures better than or comparable to baseline models, and provided novel insight into biological factors that modulate seizure risk cycles. To demonstrate the value of the model in developing brain stimulation treatment, the proposed SSM was integrated with reinforcement learning to reduce seizure risk in silico. Our model holds significant potential for addressing clinically important problems.

Objective This work aims to determine the AVEM duration and number of captured seizures required to resolve different clinical questions, using a retrospective review of ictal recordings. Methods Patients who underwent home-based AVEM had event data analyzed retrospectively. Studies were grouped by clinical indication: seizure differential diagnosis, classification, or treatment assessment. The proportion of studies where the conclusion was changed after the first seizure was determined, as was the AVEM duration needed for at least 99% of studies to reach a diagnostic conclusion. Results The referring clinical question was not answered entirely by the first event in 29.56% (n=227) of studies. Diagnostic and classification indications required a minimum of 7 days for at least 99% of studies to be answered, whilst treatment-assessment required at least 6 days. Conclusions At least 7 days of monitoring, and potentially multiple events, are required to adequately answer these clinical questions in at least 99% of patients. The widely applied 72 hours or single event recording cut-offs may be insufficient to correctly answer these three indications in a substantial proportion of patients. Significance Extended duration of monitoring and capturing multiple events should be considered when attempting to capture seizures on AVEM.

Sleep duration, sleep deprivation and the sleep-wake cycle are thought to play an important role in the generation of epileptic activity and may also influence seizure risk. Hence, people diagnosed with epilepsy are commonly asked to maintain consistent sleep routines. However, emerging evidence paints a more nuanced picture of the relationship between seizures and sleep, with bidirectional effects between changes in sleep and seizure risk in addition to modulation by sleep stages and transitions between stages. We conducted a longitudinal study investigating sleep parameters and self-reported seizure occurrence in an ambulatory at-home setting using mobile and wearable monitoring. Forty-four subjects wore a Fitbit smartwatch for at least 28 days while reporting their seizure activity in a mobile app. Multiple sleep features were investigated, including duration, oversleep and undersleep, and sleep onset and offset times. Sleep features in participants with epilepsy were compared to a large (n=37921) representative population of Fitbit users, each with 28 days of data. For participants with at least 10 seizure days (n=29), sleep features were analysed for significant changes prior to seizure days. A total of 3894 reported seizures (M = 88, SD = 130) and 17078 recorded sleep nights (M = 388, SD = 351) were included in the study. Participants with epilepsy slept an average of 2 hours longer than the average sleep duration within the general population. Just 1 of 29 participants showed a significant difference in sleep duration the night before seizure days compared to seizure-free days. However, 11 of 29 subjects showed significant differences between either their sleep onset (bed) or offset (wake) times prior to seizure occurrence. Conversely to existing studies, the current study found oversleeping was associated with a 20% increased seizure risk in the following 48h (p < 0.01), possibly due to nocturnal seizures occurring overnight. Nocturnal seizures were associated with both significantly longer sleep durations and increased risk of a seizure occurring in the following 48h. We also observed that oversleeping only significantly contributed to seizure risk when the participant was in a high-risk state, according to a cycles-based forecasting algorithm. Overall, the presented results demonstrated that day-to-day changes in sleep-duration had a minimal effect on reported seizures, which bed- and wake-times were more important for identifying seizure risk the following day. Oversleeping was also linked to seizure occurrence, most likely due to nocturnal seizures driving oversleep. Wearables can be utilised to identify these sleep-seizure relationships and guide clinical recommendations or improve seizure forecasting algorithms.

The factors that influence seizure timing are poorly understood, and seizure unpredictability remains a major cause of disability. Work in chronobiology has shown that cyclical physiological phenomena are ubiquitous, with daily and multiday cycles evident in immune, endocrine, metabolic, neurological, and cardiovascular function. Additionally, work with chronic brain recordings has identified that seizure risk is linked to daily and multiday cycles in brain activity. Here, we provide the first characterization of the relationships between the cyclical modulation of a diverse set of physiological signals, brain activity, and seizure timing.

BACKGROUND:Seizure risk forecasting could reduce injuries and even deaths in people with epilepsy. There is great interest in using non-invasive wearable devices to generate forecasts of seizure risk. Forecasts based on cycles of epileptic activity, seizure times or heart rate have provided promising forecasting results. This study validates a forecasting method using multimodal cycles recorded from wearable devices. METHOD:Seizure and heart rate cycles were extracted from 13 participants. The mean period of heart rate data from a smartwatch was 562 days, with a mean of 125 self-reported seizures from a smartphone app. The relationship between seizure onset time and phases of seizure and heart rate cycles was investigated. An additive regression model was used to project heart rate cycles. The results of forecasts using seizure cycles, heart rate cycles, and a combination of both were compared. Forecasting performance was evaluated in 6 of 13 participants in a prospective setting, using long-term data collected after algorithms were developed. FINDINGS:The results showed that the best forecasts achieved a mean area under the receiver-operating characteristic curve (AUC) of 0.73 for 9/13 participants showing performance above chance during retrospective validation. Subject-specific forecasts evaluated with prospective data showed a mean AUC of 0.77 with 4/6 participants showing performance above chance. INTERPRETATION:The results of this study demonstrate that cycles detected from multimodal data can be combined within a single, scalable seizure risk forecasting algorithm to provide robust performance. The presented forecasting method enabled seizure risk to be estimated for an arbitrary future period and could be generalised across a range of data types. In contrast to earlier work, the current study evaluated forecasts prospectively, in subjects blinded to their seizure risk outputs, representing a critical step towards clinical applications. FUNDING:This study was funded by an Australian Government National Health & Medical Research Council and BioMedTech Horizons grant. The study also received support from the Epilepsy Foundation of America's 'My Seizure Gauge' grant.

Multiday cyclic patterns underlying the timing of seizures are well-established in adults with epilepsy and are critical to the development of seizure risk forecasting models. As cycles underpinning these models are yet to be explored in paediatric cohorts, the current study applies methods drawn from seizure risk forecasting to identify and compare multiday seizure cycles between paediatric and adult cohorts. This is followed by the first validation of personalised forecasts of seizure likelihood in a paediatric cohort. Multiday seizure cycles were extracted retrospectively from 325 (71 paediatric) electronic seizure diary users (more than 28 days of app use) with confirmed epilepsy. Cycles were grouped (k-means clustering), and seizure cycles quantified (synchronisation index), with significant cycles identified by Rayleigh test of periodicity ( p <0.05). Wilcoxon rank-sum test assessed differences in prevalence and strength of cycle groups between paediatric and adult cohorts. 34,402 seizures (paediatric: M =101, SD =103, adult: M =107, SD =156) were analysed and seizure cycles were grouped according to circadian (0.5-1.5 days), about-weekly (2-12 days), about-fortnightly (13-22 days) and about-monthly (23-32 days) periodicities. Significant cycles were identified in each cycle group, with no differences in prevalence or cycle strength between paediatric and adult cohorts for any multiday cycle group. Similar effects were observed across clinical and demographic features (sex, epilepsy-type, medication). These multiday patterns formed the basis for cycle-based estimates of seizure likelihood. Receiver operating characteristic (area under the curve: AUC) was applied and demonstrated that these seizure forecasts performed better than chance (i.e. shuffled seizure times). Multiday seizure cycles are therefore similar in paediatric and adult cohorts, and this study provides the first validation of cycle-based seizure risk forecasting models as a promising approach for paediatric epilepsy.### Competing Interest StatementH.K., P.J.K., R.E.S., E.S.N., D.E., D.F. and M.J.C. have employment or a financial interest in Seer Medical Pty. Ltd. The remaining authors have no conflicts of interest.### Funding StatementThis research was supported by NHMRC Investigator Grant (1178220), My Seizure Gauge Grant (Epilepsy Foundation of America) and the BioMedTech Horizons 3 program (initiative of MTPConnect). Funding partners were not involved in the study design, collection, analysis, interpretation of data, the writing of this article or the decision to submit it for publication.### Author DeclarationsI confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.YesThe details of the IRB/oversight body that provided approval or exemption for the research described are given below:Ethics committee of St Vincent's Hospital Melbourne gave ethical approval for this work.I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.YesI understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).YesI have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.YesAll data produced in the present study are available upon reasonable request to the authors.

Seizures beget seizures is a longstanding theory that proposed that seizure activity can impact the structural and functional properties of the brain circuits in ways that contribute to epilepsy progression and the future occurrence of seizures. Originally proposed by Gowers, this theory continues to be quoted in the pathophysiology of epilepsy. We critically review the existing data and observations on the consequences of recurrent seizures on brain networks and highlight a range of factors that speak for and against the theory. The existing literature demonstrates clearly that ictal activity, especially if recurrent, induces molecular, structural, and functional changes including cell loss, connectivity reorganization, changes in neuronal behavior, and metabolic alterations. These changes have the potential to modify the seizure threshold, contribute to disease progression, and recruit wider areas of the epileptic network into epileptic activity. Repeated seizure activity may, thus, act as a pathological positive-feedback mechanism that increases seizure likelihood. On the other hand, the time course of self-limited epilepsies and the presence of seizure remission in two thirds of epilepsy cases and various chronic epilepsy models oppose the theory. Experimental work showed that seizures could induce neural changes that increase the seizure threshold and decrease the risk of a subsequent seizure. Due to the complex nature of epilepsies, it is wrong to consider only seizures as the key factor responsible for disease progression. Epilepsy worsening can be attributed to the various forms of interictal epileptiform activity or underlying disease mechanisms. Although seizure activity can negatively impact brain structure and function, the "seizures beget seizures" theory should not be used dogmatically but with extreme caution.

Epilepsy patients often experience acute repetitive seizures, known as seizure clusters, which can progress to prolonged seizures or status epilepticus if left untreated. Predicting the onset of seizure clusters is crucial to enable patients to receive preventative treatments. Additionally, studying the patterns of seizure clusters can help predict the seizure type (isolated or cluster) after observing a just occurred seizure. This paper presents machine learning models that use bivariate intracranial EEG (iEEG) features to predict seizure clustering. Specifically, we utilized relative entropy (REN) as a bivariate feature to capture potential differences in brain region interactions underlying isolated and cluster seizures. We analyzed a large ambulatory iEEG dataset collected from 15 patients and spanned up to 2 years of recordings for each patient, consisting of 3341 cluster seizures (from 427 clusters) and 369 isolated seizures. The dataset's substantial number of seizures per patient enabled individualized analyses and predictions. We observed that REN was significantly different between isolated and cluster seizures in majority of the patients. Machine learning models based on REN: 1) predicted whether a seizure will occur soon after a given seizure with up to 69.5% Area under the ROC Curve (AUC), 2) predicted if a seizure is the first one in a cluster with up to 55.3% AUC, outperforming baseline techniques. Overall, our findings could be beneficial in addressing the clinical burden associated with seizure clusters, enabling patients to receive timely treatments and improving their quality of life.