[Interpretable dynamic adjacency graph attention network for epileptic seizure detection]

To address the insufficient robustness of existing deep learning methods in modeling cross-channel dependencies in epileptic electroencephalogram (EEG) signals and the limited traceability of their decision-making rationale, this paper proposes an interpretable dynamic adjacency graph attention network (DA-GAT). The proposed model first extracts EEG features through a convolutional module, and subsequently introduces a dynamic graph learning module and a multi-head graph attention module with prior-gated fusion to achieve automatic epileptic EEG detection. Experimental results on three public datasets, namely the Children's Hospital Boston-Massachusetts Institute of Technology (CHB-MIT), Helsinki, and Siena datasets, demonstrate that DA-GAT outperforms the selected comparison methods in terms of accuracy, F1-score, and other evaluation metrics. Furthermore, by combining gradient-weighted class activation mapping (Grad-CAM) with dynamic adjacency matrix visualization, DA-GAT provides interpretable evidence to support clinical review. In summary, this study provides a new methodological approach to cross-channel dependency modeling and interpretability analysis of epileptic EEG signals, and provides a reference for improving the reliability and clinical applicability of deep learning models for automatic epilepsy detection.