An Efficient Deep Learning Framework for Automated Epileptic Seizure Detection: Toward Scalable and Clinically Applicable Solutions

In this study, we present an efficient epileptic seizure detection framework driven by a graph convolutional neural network (GCNN). Unlike conventional methods that primarily rely on local features or complex feature engineering, our GCNN-based approach explicitly encodes the spatial dependencies among electroencephalogram (EEG) electrodes, thereby capturing more comprehensive spatiotemporal features. A minimal preprocessing pipeline, consisting only of bandpass filtering and segmenting, reduces system complexity and computational overhead. On the CHB-MIT scalp EEG database, our method achieved an average accuracy of 98.64%, sensitivity of 99.49%, and specificity of 98.64% at the segment-based level and sensitivity of 96.81% with FDR of 0.27/h at the event-based level. On the SH-SDU database we collected, the method yielded segment-based accuracy of 95.23%, sensitivity of 92.42%, and specificity of 95.25%, along with event-based sensitivity of 94.11%. The average testing time for 1 h of multi-channel EEG signals is 3.89 s. These excellent results and low-computation design make the framework especially suited for clinical applications, advancing EEG-based epilepsy diagnostics and improving patient outcomes.