Group-level Interpretation of Electroencephalography Signals Using Compact Convolutional Neural Networks

Abstract

Despite the excellent performance of deep learning models for decoding electroencephalography (EEG) signals, the lack of explainability hinders the implementation of deep learning techniques in neuroscience. Although recently developed solutions ensure physiologically plausible interpretations, the robustness against subject variability and artifacts requires further improvement. This study presents a method for obtaining the group-level interpretation of EEG signals using a compact convolutional neural network (CNN). The convolutional filters of the CNN were clustered, and the clusters with high task-relevant scores were selectively interpreted. The proposed group-level analysis method was validated using a motor imagery dataset, and the results were visually and quantitatively compared with those obtained from the individual-level analysis. The cortical sources interpreted using the proposed group-level analysis exhibited a significantly smaller root mean square error of the source location from the task-relevant cortical area than those interpreted via the individual-level analysis. Furthermore, the cortical sources in the group-level analysis were concentrated in the sensorimotor area and denoted the event-related desynchronization in the alpha band, which is associated with motor imagery tasks. Conversely, the individual-level analysis resulted in unclear spatial and spectral properties of cortical sources. The findings of this study verify the feasibility of group-level analysis based on compact CNNs, which can robustly handle subject variability and artifacts.