A Feature Selection and Hyperparameter Optimization Framework for Multiclass EEG-based Epilepsy Detection

Abstract

Epilepsy is a neurological disorder characterized by recurrent seizures, and EEG continues to be the gold standard for diagnosis in clinical practice. However, the nonlinear nature of the EEG and its sensitivity to noise make manual interpretations difficult. This study presents a feature selection and hyperparameter optimization approach for automatic multi-class EEG-based epilepsy detection. Five classes representing two healthy states and three epileptic states were analyzed using EEG signals from the Bonn database. Temporal, statistical, fractal, and spectral features were extracted, and their importance was ranked according to ANOVA F-values. Classifications were performed using support vector machines, random forests, nearest-neighbor algorithms, decision trees, multilayer perceptrons, and linear discriminant analysis; hyperparameters were optimized using Grid Search Cross-Validation. The robustness of the proposed model was ensured using grouped layered k-fold cross-validation. With this approach, it achieved an F1-score of 0.8126 in 5-class classification and 0.9894 in binary classification, outperforming existing methods on the same dataset. Hjorth complexity, standard deviation, and Hjorth mobility emerged as the most discriminative features for epileptic states. The study demonstrates that combining feature selection with hyperparameter optimization provides a robust, interpretable framework for multi-class EEG classification. This approach has the potential to support neurologists in their epilepsy assessments by improving diagnostic accuracy and providing clinically meaningful insights.