Dynamic Graph Contrastive Learning for EEG-Based Epileptic Seizure Detection

Conventional methods for diagnosing epilepsy using EEG struggle to capture the complex spatiotemporal relationships inherent in brain activity. This paper introduces Dynamic Graph Contrastive Learning (DGCL) for detecting epileptic seizures. DGCL combines dynamic graph neural networks (DGNNs) with self-supervised contrastive learning, enabling the model to capture the intricate spatiotemporal dependencies in EEG data while minimizing the need for labeled data. The model utilizes DGNNs to represent the evolving state of the brain network and applies contrastive learning techniques to extract robust feature representations from unlabeled data. Experiments conducted on the CHB-MIT dataset show that DGCL outperforms existing methods, achieving accuracies of 96.95% for patient-specific data and 73.67% for cross-patient data. Comparisons with other state-of-the-art techniques further validate DGCL’s effectiveness in classification accuracy and its ability to generalize. These results confirm that DGCL offers a promising solution for automated epilepsy detection, especially in cases where labeled data is scarce.