Epileptic seizures are one of the most common neurological disorders worldwide, yet accurate prediction remains a major challenge to date. Existing techniques largely rely on either electroencephalography (EEG) or neuroimaging and use a one-size-fits-all paradigm with limited personalization. Wearable physiological signals such as heart rate (HR), electrodermal activity (EDA), and skin temperature (TEMP) provide a non-invasive, real-time monitoring solution, whose seizure prediction potential is yet to be comprehensively explored. This work proposes a proof-of-concept patient-specific seizure risk forecasting model that integrates a Transformer encoder with a Proximal Policy Optimization (PPO) reinforcement learning agent. The system is trained from 30-second segments of wearable data, fusing clinical priors with patient-specific percentiles through an α -blended thresholding and dynamic reward mechanism. Performance is evaluated using both standard frame-level metrics (e.g., precision, recall, F1) and clinically motivated episode-level metrics (e.g., true-positive and false-alarm rates, total episodes per hour). Across ablation studies, we build upon conventional machine learning baselines that over-fit rule-based labels and fail to capture temporal structure. Occlusion-based explanation also evidences that the agent prefers clinically important feature combinations (HR+EDA, HR+TEMP). Our results indicate the promise of reinforcement learning from evolving wearable signals for proactive and individualised seizure prediction. We conclude by emphasizing the limitations of simulated data use and point towards future validation on actual patient datasets. • Wearable HR, EDA, and temperature signals used for seizure risk forecasting. • Transformer-enhanced PPO reinforcement learning enables personalized adaptation. • α -blended thresholds fuse clinical priors with patient-specific baselines. • Episode-level metrics (TP/FA per hour) align with clinical relevance. • Occlusion analysis confirms reliance on physiologically meaningful features.
Kapadia et al. (Fri,) studied this question.