Channel Selection for Seizure Detection Based on Explainable AI With Shapley Values

This article proposes an interpretable channel selection method for electroencephalogram (EEG)-based seizure detection, which can significantly reduce the number of channels and thus the computational complexity. To achieve the above goal, feature extraction in three domains and the long short-term memory (LSTM) model are employed for accurate classification. Also, the deep SHapley Additive exPlanations (SHAP), an explainable artificial intelligence (AI) technology, is adopted in conjunction with the LSTM network to perform channel selection based on a backpropagation strategy. The Shapley value, calculated by the deep SHAP, quantifies the individual contribution of each channel to seizure detection, based on which the optimal channel combination is determined. Evaluated by the CHB-MIT dataset, the proposed method can yield a well-balanced performance using only five channels on average, including an area under the curve (AUC) of 0.9387, an accuracy of 95.31%, a sensitivity of 92.42%, and a specificity of 95.32%. Notably, the process of channel selection can be visualized, which enables trust and transparency in decision-making processes.