Lightweight and Interpretable Channel Selection for EEG Measurement Systems in Epileptic Seizure Prediction

Utilizing multichannel electroencephalogram (EEG) signals to predict epileptic seizures is crucial for reducing patient distress. However, devices equipped with numerous electrodes are often cumbersome, expensive, and impractical for long-term or home-based monitoring. To address this challenge, we propose a novel EEG channel selection framework that combines LayerCAM with reinforcement learning, aiming to enhance the accuracy and interpretability of epilepsy prediction models. In our framework, LayerCAM is integrated into a specially designed convolutional neural network to accurately align class activation mapping with physical electrodes, thereby ensuring precise localization. This localization serves as the foundation for effective channel selection. Subsequently, we employ a transformer architecture termed Decision Transformer to treat channel selection as a sequence prediction task, and class activation mapping is utilized as a state within a Markov decision process. This strategy captures the dependencies between different channels and the decisions to activate or deactivate them, facilitating the learning of temporal and causal relationships. To validate our approach, we conducted extensive experiments on two datasets. The results demonstrated significant improvements in prediction accuracy across three distinct neural network architectures compared to utilizing all available EEG channels. Specifically, our approach achieved an accuracy of 93.69% with eight channels on the CHB-MIT dataset and 95.41% with only five channels on the Renji dataset. These results demonstrate that reliable seizure prediction is achievable with simplified sensing configurations, thus enabling the design of portable and cost-effective EEG monitoring systems for practical use.