Abstracts

Rationale:
Wrist-worn wearable devices that monitor autonomous nervous system and movement have shown promise in providing non-invasive applicable seizure forecasts. Nevertheless, challenges related to small number of enrolled patients, insufficient training data and lacking patient seizure cycles information hinder its clinical implementation. In this work, we prospectively validated a previously implemented seizure forecasting algorithm using a larger cohort of pediatric patients with epilepsy. We sought to prove that the performance increases predictably with the dataset size.




Methods:
We recorded continuously electrodermal activity (EDA), peripheral body temperature (TEMP), blood volume pulse (BVP), accelerometery (ACC), with the 24-hour cycles as two sinusoidal cycles with shifted phases. We used a previously implemented Long Short-Term Memory (LSTM) deep model with additional hyperparameters [1] to classify 30-seconds data segments as preictal or interictal periods. We identified preictal periods by defining lead seizures that have at least 2-hour preceding duration of seizure-free recording, then labelling the 1-hour period ending 1-minute before the EEG-onset as preictal. We identified interictal periods as seizure-free recording with a minimum 2-hour duration to any nearby seizure. We applied leave-one-subject-out nested-cross validation across 20 runs. During training we used balanced data from each patient, and for testing we used the entire test-out patient’s data. We tested the better than chance significance of the 20 forecasts in each patient using two-sided Wilcoxon signed-rank test. In addition, we evaluated how the performance scales with dataset size by assessing the performance from the test patients by applying log10 to the x- and y-axis values then fitting a straight line, and its slope is used to estimate the power-law exponent [2].




Results:
We investigated wearable signals from 450 patients undergoing multi-day monitoring in the EMU, 167 patients (age 10±6 years, 78 females, recording duration 7253 hours) had 447 lead seizures. We evaluated 166 patients, 441 lead seizures, with sufficient data. We achieved Improvement over Chance (IoC) in 68% of the patients, with a 19.1% IoC, 78.4% Sensitivity (Sen), and a 64.5% Time in Warning (TiW) in all patients. In the 68% of patients with significant seizure forecasting, we achieved a 27.3% IoC, 90.3% Sen and 63.0% TiW (Figure 1). Patients with significant forecasting had higher reported prior seizure frequency than others (p=0.02, Mann-Whitney U test). Including the 24-hour cycle led to an overall better performance, where tonic-clonic seizures had better IoC by comparing the same number of patients in training (Figure 2).




Conclusions:
Seizure forecasting performance using wearable signals improves predictably with a larger dataset following precise power-law scaling. In addition, including information on 24-hour seizure cycles further improves performance. Tonic-clonic seizure may be particularly suitable to forecasting.





Funding: This study was in part supported by the Epilepsy Research Fund.