An Attention Mechanism-Based Interpretable Model for Epileptic Seizure Detection and Localization With Self-Supervised Pre-Training

Epilepsy, characterized by unpredictable seizures caused by abnormal brain activity, presents significant diagnostic challenges due to the complexity of interpreting electroencephalogram (EEG) signals. Current EEG analysis methods rely heavily on manual annotations by neurophysiologists, limiting scalability and efficient utilization of vast clinical datasets. In this study, we propose a self-supervised learning (SSL) model for seizure detection and localization, designed to reduce dependence on labeled data while maintaining high performance. The pretext task of signal transformation recognition enables the model to learn generalized spatiotemporal features from EEG data, effectively leveraging unlabeled recordings. This is followed by fine-tuning with a limited set of labeled data for precise seizure detection. Our model integrates an attention mechanism that enhances interpretability by focusing on the most relevant EEG electrodes, improving both detection accuracy and seizure localization. The proposed solution demonstrates competitive performance on the CHB-MIT dataset, achieving a sensitivity of 93.1% and an AUC of 91% with only 30% labeled data, and consistent localization accuracy of 83.33%. Extensive experiments across multiple datasets and under various configurations validate the robustness and generalizability of the proposed approach. This work not only introduces a data-efficient and interpretable solution for EEG-based seizure detection but also sets the stage for future research on applying SSL to broader neurological disorders, opening new avenues for trustworthy and optimal medical diagnostics.