Epileptic signal classification using convolutional neural network and Shapley additive explainable artificial intelligence method

• Published in: Neural computing & applications Vol. 37; no. 6; pp. 4937 - 4955
• Main Authors: Rathod, Prajakta, Naik, Shefali, Bhalodiya, Jayendra M.
• Format: Journal Article
• Language: English
• Published: London Springer London 01.02.2025; Springer Nature B.V
• Subjects: Accuracy; Artificial Intelligence; Artificial neural networks; Channels; Classification; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Datasets; Electroencephalography; Epilepsy; Explainable artificial intelligence; Image filters; Image Processing and Computer Vision; Neural networks; Neurological diseases; Original Article; Probability and Statistics in Computer Science; Signal classification; Explainable artificial intelligence; Electroencephalogram; Epilepsy; Shapley additive; Machine learning
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-024-10915-7

Neurological disorder epilepsy can cause life-threatening seizures which may produce abnormal patterns in electroencephalogram (EEG) signals. However, manual tracking of abnormal EEG patterns is subjective, and only artificial intelligence (AI)-based identification lacks trustworthiness. Therefore, we proposed an explainable AI (XAI)-based method to classify epileptic and non-epileptic patients. The proposed method has four steps. The first step creatively preprocesses EEG signals into a stack of spectrogram images and enhances them using filters. The second step employs convolutional neural networks (Densenet121, Resnet18 and VGG16) to classify images. The third step incorporates the SHapley Additive exPlanations (SHAP) method to explain the classification results. The fourth step identifies significant pixels and top contributing EEG channels for each classification. Moreover, we comparatively found the best AI and XAI combination among Densenet121, Resnet18 and VGG16, with SHAP using area over perturbation curve (AOPC). We used the Temple University Hospital EEG Epilepsy Corpus dataset (100 subjects: 50 epileptic, 50 non-epileptic) with k-fold cross-validation to evaluate the classification accuracy. Proposed VGG16-, Densenet121- and Resnet18-based approaches achieved 97.55%, 91.29% and 95% accuracy, respectively. Moreover, VGG16-, Densenet121- and Resnet18-based approaches with SHAP reported AOPC values of 197,600, 355,350 and 240,150, respectively. Additionally, we identified the top contributing EEG channels for each patient. The proposed VGG16-based approach achieved the highest accuracy compared to the literature using the same dataset. Densenet121 with SHAP reported the best explainability. Such an approach can help EEG technologists and neurologists to identify epileptic patients automatically by using EEG.