New transformer-based AI system predicts epileptic seizures 30 seconds in advance with over 90% accuracy by adapting to individual patients.

Researchers developed a patient-adaptive transformer framework that predicts seizure onset within a 30-second window using EEG brain signal recordings

The model uses a two-stage training approach: self-supervised pretraining to learn general EEG patterns, followed by patient-specific fine-tuning for seizure prediction

Validation results on the TUH EEG dataset show the method achieves over 90% accuracy and F1 scores exceeding 0.80, addressing the challenge of high variability between patients

The approach preprocesses multichannel EEG signals with noise-aware techniques and converts them into tokenized sequences suitable for transformer-based learning