An Interpretable Machine Learning Framework for Interictal EEG-based Epilepsy Classification

Reliable detection of interictal epileptic activity from electroencephalography (EEG) signals remains a challenging problem, as interictal patterns lack the distinctive characteristics observed during seizure episodes. Although many existing studies report high classification performance, this is often achieved by including ictal EEG signals, which substantially simplifies the task and limits clinical realism. In addition, the growing use of complex deep learning models has raised concerns regarding interpretability and practical deployment in clinical settings. To address these limitations, an interpretable machine learning framework is proposed for discriminating normal EEG signals from pathological interictal activity. EEG recordings from healthy subjects and epilepsy patients were analysed using a comprehensive set of features from the time and frequency domains designed to characterise spectral characteristics and temporal dynamics. Multiple machine learning classifiers were evaluated under a unified experimental protocol using five-fold cross-validation to ensure robust performance assessment. The experimental results demonstrate that the k-nearest neighbour classifier achieved the most balanced performance, with an accuracy of 95.5% in distinguishing interictal EEG segments from normal recordings. Furthermore, a Shapley-based feature contribution analysis revealed that a compact subset of influential features can retain high classification accuracy, indicating that model complexity can be significantly reduced without compromising performance. These findings demonstrate that effective interictal EEG discrimination can be achieved using a transparent and computationally efficient approach, providing an interpretable methodological benchmark for EEG-based analysis under controlled experimental conditions.