Region-Aware Genetic Feature Selection with Demographic Meta-Integration and Ensemble Learning for EEG epilepsy Seizure Detection

Abstract

Accurate electroencephalogram (EEG)-based seizure detection is important for early epilepsy diagnosis and timely intervention, yet existing methods often trade predictive performance for interpretability. Deep learning models can achieve high accuracy but function as black boxes, limiting clinical trust. Conventional machine learning models are more transparent, but they often ignore neurophysiological structure and patient-specific metadata, which can reduce performance. To address this gap, this study proposes an interpretable framework that combines region-aware genetic algorithm feature selection, demographic meta-integration, and ensemble learning. EEG channels are first grouped into neurophysiological regions, including frontal, central-parietal, temporal, and occipital areas. A genetic algorithm is then applied within each region to identify the most informative channels while preserving clinically meaningful brain topology. The selected EEG features are combined with patient demographic and clinical metadata, including age, gender, medication status, and seizure history, to create a compact feature vector. This feature set is used to train a soft-voting ensemble of RandomForest, ExtraTrees, and XGBoost classifiers. The framework was evaluated on a dataset of 50 drug-resistant epileptic patients and achieved 99.28% accuracy with a very low false-negative rate. In addition to strong predictive performance, the proposed method remains inherently interpretable by indicating which brain regions contribute to seizure prediction, making it suitable for practical clinical deployment.