An Efficient Dual-Specialist Framework for Real-Time Epileptic Seizure Detection and Prediction

Abstract

Real-time epileptic seizure monitoring requires simultaneous detection and prediction capabilities, yet existing systems address only one task, leading to either delayed response or excessive false alarms. This paper presents a dual-specialist deep learning framework that performs concurrent seizure detection and prediction through three key innovations: (1) seizure-sensitive preprocessing with cross-frequency coupling detection and adaptive channel weighting, (2) complementary dual-specialist architecture where a detection specialist optimizes for rapid response while a prediction specialist optimizes for conservative forecasting, and (3) probabilistic temporal state machine providing interpretable risk assessment through four clinical states with sustained evidence validation. Evaluation on the CHB-MIT database using Leave-One-Seizure-Out cross-validation across 20 patients (98 seizures, 851.1 hours) demonstrates 99.2% sensitivity, 99.7% specificity, and 0.31/h false alarm rate with 7.9-minute average warning time, achieved with 1.2M parameters and 1.58M multiply-accumulate operations. Real-time Raspberry Pi 4 deployment achieves 108.0 ± 13.5 ms inference latency, validating practical edge deployment. The framework demonstrates potential for wearable seizure monitoring applications with improved false alarm control.