AveragedLIME for general explanations in EEG domain

This paper introduces averagedLIME, a method designed to enable global interpretation of Convolutional Neural Network (CNN) decisions by averaging local explanations generated using the Local Interpretable Model-agnostic Explanations (LIME) algorithm. The method was developed for systems with a relatively stable spatial distribution, such as event related potentials (ERP)-based systems. The method's correctness and its ability to detect patterns in electroencephalography (EEG) data were evaluated through two separate studies. The first study aimed to confirm the method's reliability under conditions where the regions of interest (ROIs) were precisely known and directly observable both in the raw data and in individual saliency maps. The second study, more exploratory in nature, revealed coherent and medically interpretable ROIs that were not visible in the measurement data and nearly absent from individual saliency maps. In that study, the averagedLIME successfully uncovered general patterns that would otherwise remain hidden. To further contextualize the proposed approach, we compared the interpretability results obtained with averagedLIME against those produced by two widely used reference techniques - SHapley Additive exPlanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM). The results demonstrated that averagedLIME yielded more consistent and generalizable activation patterns compared to these reference approaches. The obtained results further support the hypothesis that under certain conditions (relatively stable spatial distribution), the averaged saliency maps can support more general interpretations of CNN model behavior than individual maps. Thus, the averagedLIME method has the potential to significantly enhance the transparency of deep learning-based systems in neuro-informatics and diagnostic applications.