Empowering Wearable Seizure Forecasting with Scheduled Sampling

The unpredictability of seizures imposes a significant burden on tens of millions of individuals with epilepsy worldwide. The ability to continuously monitor and forecast epileptic seizures would lead to a paradigm shift in epilepsy management. In this paper, we propose a novel progressive, personalized two-stage approach for seizure forecasting using 10-minute wearable time series data from wristbands worn by epilepsy patients. Our method effectively tackles the challenges posed by class imbalance and the complex nature of physiological signals. By measuring and ranking the reconstruction error and energy the normal samples present to a deep autoencoder and employing scheduled sampling, we demonstrate superior performance over existing deep learning models, anomaly detection methods, and class balancing during training. The proposed approach offers a promising solution for seizure forecasting and has potential applications in other medical problems characterized by imbalanced data and complex physiological signals.Clinical relevance— The study demonstrates the potential for seizure forecasting using wearable data and individualized treatment planning. Its findings also highlight the value of adaptive learning mechanisms in training deep learning models for imbalanced healthcare data.