EEG-Based Patient Independent Epileptic Seizure Detection Using GCN-BRF

It is estimated that 50–70 million people worldwide suffer from epilepsy, making it one of the most common neurological disorders. The detection of seizures, a primary symptom of epilepsy, presents significant challenges due to the limitations of patient-specific models and the complex nature of EEG signal analysis. Traditional methods, focusing on isolated EEG channels, fail to capture the dynamic interconnections within the brain's network, essential for accurate seizure detection. This study proposes a novel, patient-independent approach leveraging Graph Neural Networks (GNNs) and the Graph Convolutional Network (GCN) layer to process graph-structured data that represents the brain's network. Our methodology utilizes a comprehensive feature set derived from Random Forest selection, including 37 node-specific features and two global features: the GCN's classification output and an eigenvector from the correlation matrix. This rich feature representation enables a detailed analysis of the EEG data's structural properties.

Evaluated on unseen data from four patients, our model demonstrated exceptional generalizability and performance, achieving significant metrics such as 91.70% accuracy, 91.32% precision, 88.71% sensitivity, and a 91.57% F1-Score in patient-independent settings. Further patient-specific evaluations on a subset of 14 patients underscored the model's efficacy, with near-perfect scores across all key metrics. Our findings offer a promising direction for epilepsy care, showcasing the potential of GNNs to surmount existing challenges in seizure detection and contribute valuable insights into neurological disorder analysis.