Explainable AI based dilated convolutional self-attention based bidirectional gated network for Epilepsy seizure detection

Epilepsy affects 50 million people worldwide and has a significant influence on their quality of life. Epilepsy prevalence varies depending on a variety of conditions, although it is more frequent in underdeveloped countries. This underscores the crucial need for advances in treatment and prevention approaches to improve people's quality of life around the world. Therefore, the proposed study introduces a novel Explainable AI (XAI) based hybrid deep learning model for detecting epilepsy seizures. The most important processes in this study are collecting data, preprocessing, extracting features, and detection. Initially, input EEG signals are acquired from a publically available dataset, then pre-processing is performed to improve signal quality by removing artifacts using the Windowing-based Modified Notch Impulse Response Filter (WMNIRF) approach. Then, the necessary features are extracted in the feature extracted stage using an Improved Pseudo-Wigner Ville Distribution (IPWVD) and Extended Fast Fourier Transform (EFFT) methods. From the extracted features, the most optimal feature sets are selected to reduce the feature dimensionality problem using Maximum Information Coefficient based Frilled Lizard Optimization (MIC-FLO) approach. Finally, based on the selected features, the epilepsy seizures are detected by proposing a novel Dilated Convolutional Self-Attention based Bidirectional Gated Network (DC-SABiGN) model. Finally, for making the proposed deep learning based detection model as more interpretable, Shapley Additive Explanations (SHAP) is utilized and it can analyse the features that more necessary for seizure detection. In addition, the proposed model attain accuracy of 96%, precision of 97%, recall of 96%, and specificity of 96%.