Similarity-Driven Personalization and Optimization for Long-Horizon EEG Seizure Prediction

Epileptic seizure prediction using an Electroencephalogram (EEG) can improve patient safety by enabling early intervention, yet most existing approaches focus on short prediction horizons with limited personalization or computational efficiency. This study presents a unified deep learning framework evaluated across ten pre-ictal prediction windows up to 300 min before seizure onset, using recordings from 161 patients and 1023 seizure events. At the 5 min horizon, the generalized model achieved 96.30% accuracy and 91.62% sensitivity. Two complementary personalization strategies are introduced: incremental transfer learning, which progressively fine-tunes the generalized model using patient-specific data, and Dynamic Time Warping (DTW)-based similarity personalization, which constructs a morphology-aware training cohort from a single reference seizure. Personalized models consistently outperform generalized baselines, particularly at longer horizons, with the DTW-based approach achieving 89.68% accuracy using only 70 similar patients. Reliable prediction is demonstrated up to 60 min prior to onset, while model optimization reduces computational complexity with minimal performance loss, supporting deployment in resource-constrained clinical environments.