Self-Supervised Learning with Adaptive Graph Modeling for EEG-Based Epileptic Seizure Classification

Objective: Epileptic seizure classification using EEG signals remains a significant challenge due to complex spatial-temporal dependencies, limited labeled data, and severe class imbalance. Methods: We propose a self-supervised learning framework, ASGPF (Adaptive Spatio-Graph Pretraining Framework), for EEG-based seizure classification. At its core is a novel Spatio-Graph Learning Cell (SGLC), which integrates a Graph Learning module to dynamically construct EEG topology, a Gated Graph Neural Network to extract spatial features across EEG channels, and a Gated Recurrent Unit to capture longterm temporal dependencies. ASGPF uses self-supervised sequence-to-sequence pretraining on unlabeled EEG to learn robust representations, enabling accurate seizure classification with a lightweight model that consists of the pretrained encoder and a simple prediction layer. Results: Extensive experiments on the TUSZ dataset demonstrate that our method significantly outperforms current state-ofthe-art approaches, achieving weighted F1-scores of 83.8% for four-class and 73.5% for eight-class seizure classification tasks, respectively. Notably, with only 25% of labeled data, the proposed model achieves comparable performance to the best baseline trained on 75% of data, validating the effectiveness of ASGPF under data scarcity and class imbalance. Conclusion: ASGPF effectively learns discriminative EEG representations through adaptive spatialtemporal modeling and self-supervised pretraining, enabling accurate seizure classification with minimal labeled data. Significance: This work introduces a data-efficient EEG analysis framework for seizure classification, enabling accurate prediction with minimal labeled data and showing strong potential for clinical application in resource- and label-constrained environments.