An Interpretable Deep Learning Framework for EEG-Based Seizure Classification Using the Bonn

Abstract

This study investigates electroencephalography (EEG)-based seizure classification using the Bonn EEG dataset. A binary classification task was formed by grouping the Z and O subsets as the non-seizure class and the S subset as the seizure class. To improve reproducibility and reduce the dependence of the results on a single train–test split, the revised experiment adopts stratified five-fold cross-validation. Within each training fold, a validation subset is used for model selection, and the test fold is kept unseen for final evaluation. A one-dimensional convolutional neural network (CNN1D), a temporal convolutional network (TCN), and a convolutional neural network–long short-term memory (CNN-LSTM) model are compared with a RandomForest baseline under the same preprocessing setting. Class-weighted training is used to address the mild class imbalance. CNN1D achieved the best cross-validated performance, with an accuracy of 0.9967 plus or minus 0.0075, weighted F1-score of 0.9967 plus or minus 0.0074, sensitivity of 1.0000 plus or minus 0.0000, specificity of 0.9950 plus or minus 0.0112, and an area under the receiver operating characteristic curve (AUC) of 1.0000 plus or minus 0.0000. TCN also achieved strong performance, with an accuracy of 0.9767 plus or minus 0.0435 and AUC of 0.9990 plus or minus 0.0022. Shapley additive explanations (SHAP) are used as a baseline-level interpretability analysis for RandomForest to identify potentially informative temporal points. The results indicate that convolution-based models provide reliable performance for small-scale EEG seizure classification, while the revised evaluation protocol provides a more reproducible assessment.