Abstracts

Rationale: Seizure monitoring is essential for epilepsy management. The current outpatient gold standard is based on seizure diaries with attempts to predict seizures based on historical seizure occurrence. Complementing diaries with clinical and wearable data holds the potential to improve seizure monitoring. We aimed to validate biomarkers for seizure detection and seizure prediction in wearable recordings with daily resolution, and to assess the effectiveness of combining wearable and diary data to identify periods of decreased and increased seizure susceptibility, as a step towards improved detection and prediction.


Methods: Patients with focal seizures were enrolled in a pharmacological trial to test the effectiveness of eslicarbazepine acetate as an adjunctive therapy to levetiracetam or lamotrigine (NCT03116828). In a prospective assessment over 30 weeks, 102 patients were enrolled from 2017 - 2019, maintained a seizure diary and wore wrist-worn wearables recording accelerometry (ACC), electrodermal activity (EDA), and peripheral body temperature (TEMP). Based on diary data, we labeled individual days as either “seizure” and “no-seizure”, or “pre-seizure” and “no-pre-seizure”. We compared 24-hour patterns obtained by harmonic 24-hour modeling between seizure day conditions. Best-ranking wearable markers with and without historical seizure diary variables (days-since-last-seizure, seizure frequency) were used as input features for a fully connected neural network that classifies by seizure-day conditions.


Results: The final sample contained 70 patients (median age 42.5 years; 43 females) with 5437 recorded patient-days, including 557 seizure days and 537 pre-seizure days. On average, 24-hour patterns in EDA and ACC differentiated seizure day conditions (no-seizure vs. seizure and no-pre-seizure vs. pre-seizure days; both p < 0.01). Classification between no-seizure and seizure days, based on days-since-last-seizure, seizure frequency, 24-hour EDA level and first harmonic phase shift, and ACC 24-hour level, revealed high performance of the classifier (weighted F1 = 0.86, sensitivity = 0.62, specificity =0.86). Similarly, classification between no-pre-seizure and pre-seizure days, based on days-since-last-seizure, seizure frequency, 24-hour pattern’s level and phase shift of the first harmonic for EDA and level, phase shift, and amplitude for ACC, revealed high performance of the classifier (weighted F1 = 0.84, sensitivity = 0.64, specificity =0.84).


Conclusions: On a group level, wearable data capture seizure-related differences with daily resolution, identifying periods of higher or lower seizure susceptibility. Combining diary-based clinical and wearable data bears the potential for developing a dynamic seizure detection and prediction system with daily resolution.


Funding: The study was supported by the American Epilepsy Society under award number 932267, the Epilepsy Research Fund, and Sumitomo Pharma America, Inc. (formerly Sunovion Pharmaceuticals Inc.) funded NCT03116828.