Single-channel EEG-based seizure prediction using deep learning.

Reliable seizure prediction can improve patient safety by enabling timely protective actions, yet most high-performing approaches depend on multichannel EEG, limiting feasibility in wearable and low-power environments. This study developed an ultralight deep learning model for seizure prediction using only single-channel EEG and evaluated its clinical applicability under predefined, clinically actionable criteria: a seizure prediction horizon (SPH) of 2 minutes and a seizure occurrence period (SOP) of 30 minutes. A 37,985-parameter MobileNet-derived architecture was designed to process STFT spectrograms and validated using patient-specific leave-one-out cross-validation on the SNUH and CHB-MIT datasets. Performance was assessed using segment-based accuracy and false positive rate (FPR), along with event-based sensitivity computed under SPH-SOP constraints. The model demonstrated strong and consistent performance across both datasets. In the SNUH cohort, it achieved 85.97% accuracy, an FPR of 0.130, and 94.93% sensitivity, successfully predicting 95.08% of seizures within the SOP window. In the CHB-MIT dataset, it reached 90.72% accuracy, an FPR of 0.092, and 97.92% sensitivity. These findings provide the first evidence that single-channel EEG can support reliable seizure prediction within clinically meaningful early-warning windows. The proposed lightweight model delivers multichannel-comparable performance while drastically reducing computational demand, demonstrating strong potential for real-time deployment in wearable and patient-centered epilepsy management systems.