Patient-Specific Deep Learning for Seizure Prediction: EEG-Based Modeling and Explainability

Project Details

• Student(s): Marc Farah
• Advisor(s): Dr. Lina Abou Abbas
• Department: Electrical & Computer
• Academic Year(s): 2024-2025

Abstract

Epilepsy is a chronic neurological disorder commonly diagnosed using electroencephalography (EEG). Predicting seizures prior to onset can greatly improve patient care by enabling timely interventions. While deep learning has shown promising performance in EEG-based seizure prediction, a major limitation remains its lack of interpretability—often referred to as the “black-box” problem—which restricts its clinical applicability.

This study introduces a patient-specific seizure prediction model based on a lightweight CNN architecture, trained on EEG data from the CHB-MIT dataset. To enhance transparency, we integrated the Integrated Gradients (IG) technique, which enables attribution of predictions to specific EEG input features. The extracted features—corresponding to clinically meaningful brain regions—offer insight into the neural patterns preceding seizures.

Experimental results show high accuracy, sensitivity, and precision, outperforming several benchmark models. The inclusion of explainability not only improves trustworthiness but also facilitates future clinical translation. This work contributes to a reproducible and interpretable deep learning framework for seizure prediction that addresses both performance and explainability gaps in existing literature.