Performance Analysis of Explainable Artificial Intelligence Methods in Epilepsy Detection

Epilepsy is a chronic, noncommunicable disease that may cause life-threatening seizures. During a seizure attack abnormal electrical activity results in brain dysfunction. An Electroencephalogram (EEG) can be used to record these abnormalities and monitor seizure patterns. Many Machine Learning (ML) models are proposed to identify epileptic seizure patterns which generated promising results. Interpretable ML models help understand the decisions generated by the model. Therefore, we identified epileptic seizure signals using an interpretable VGG16 model. EEG signals are split into small parts and images are obtained from these small parts. These images are fed to the VGG16 model for classification. Moreover, we analyzed and compared the results generated by three Explainable Artificial Intelligence (XAI) methods: SHAP, Guided backpropagation, and LIME. These methods identified and marked positively contributing features in the image to the classification. This study can help model developers select a better XAI method for epilepsy detection. The VGG16 model achieved seizure and non-seizure classification accuracy of 98.98% on the Bonn University dataset. XAI method SHAP produced the best explainability among all the three methods. Such explainability can help neurologists understand why the model has arrived at a particular decision and identified a person as an epileptic patient.