Spatiotemporal modeling of EEG for robust epileptic seizure detection using an edge-centric Graph Convolutional Network and BiGRU

Automated detection of epileptic seizures from Electroencephalography (EEG) is crucial for optimizing neurologists’ diagnostic workflows and advancing the clinical management of epilepsy. Effective seizure detection requires analytical frameworks capable of jointly modeling spatiotemporal patterns, as seizures exhibit distinctive spatial propagation and temporal dynamics. Conventional Graph Convolutional Network (GCN) approaches for spatial feature extraction are typically designed for sparse, node-centric graphs, which severely limits their efficacy when applied to the dense, edge-dominated network structures inherent in EEG channel connectivity. To address this methodological constraint, we introduce GPC-RES-SE, a novel edge-centric graph framework integrated with Bidirectional Gated Recurrent Units (BiGRU) for robust seizure detection. The GPC-RES-SE architecture is specifically designed to accurately extract pertinent spatial features by aggregating information across multiple relational paths, concurrently preserving original data characteristics and amplifying the contributions of clinically salient EEG channels. The refined spatial representations are subsequently processed by the BiGRU module to effectively capture underlying temporal dependencies. When rigorously evaluated on the widely recognized CHB-MIT EEG dataset, the proposed framework demonstrates exceptional performance, achieving a sensitivity of 98.78%, a specificity of 96.34%, and an overall accuracy of 97.78%. These results significantly surpass existing state-of-the-art methodologies. Furthermore, the model exhibits robust cross-dataset generalization capabilities on both the Bonn and TUH datasets, affirming its reliability for clinical deployment.