Research on epilepsy detection methods based on interpretable features and machine learning

Epilepsy is a prevalent neurological condition that impacts a significant number of individuals worldwide. Patients’ physical and mental health, as well as their daily activities, are significantly affected by seizures, necessitating prompt diagnosis and treatment. The automatic detection of epilepsy using electroencephalogram (EEG) signals has been a significant area of research. Nevertheless, the majority of current methods are based on intricate feature engineering processes that require the extraction and selection of a large number of features to identify the most discriminative feature sets. This results in a high level of algorithmic complexity, inadequate robustness, and inadequate interpretability, which complicates the provision of theoretical support to clinicians. This paper proposes a pathophysiology-driven, interpretable machine learning algorithm to address the limitations of current EEG-based epilepsy detection methods, which include poor interpretability and complex feature engineering. We developed a low-dimensional, interpretable feature combination consisting of only five features and systematically validated its discriminative capability across various epilepsy phases by innovatively integrating electrophysiological markers of epileptic seizures with nonlinear dynamical properties. In the binary classification of seizure versus non-seizure EEG segments, the XGB classifier achieved the highest accuracy of 98.73% and an F1 score of 98.57%. Classification accuracy for interictal versus ictal periods reached 95.33%, with an F1 score of 95.27%. In the challenging ternary classification task encompassing preictal, interictal, and ictal periods, the model achieved a respectable accuracy of 86.3% and an F1 score of 85.79%. Cross-database validation yielded a maximum accuracy of 82.17% and an F1 score of 81.99%, confirming the proposed features’ robust generalization capability and transformative potential. This feature set exhibits outstanding and stable performance across all models, as demonstrated by evaluations across two public datasets using five machine learning classifiers. In addition, SHAP values quantified the contribution of each feature to predictions, thereby providing a transparent decision-making rationale that substantially improves the algorithm’s interpretability and clinical utility.