Channel-Aware Self-Supervised Learning for EEG-based BCI

Electroencephalography (EEG) can record brain activity in a non-invasive way, but it is expensive and has high inter-and intra-subject variability. Recent research on EEG analysis using deep learning has shown great performance in dealing with this problem, but there is still the issue of heterogeneity. In this work, we propose a novel self-supervised learning method that can extract robust signal representations of EEG to solve the mentioned problem. To model the paradigm in which different channels exist, we devise a channel-aware encoder. In the downstream task, fine-tuning was performed by adding spatial convolution to consider channel information in the feature obtained by a single channel encoder. For the validity of our proposed framework, we conduct experiments using two public datasets with different paradigms, i.e., sleep staging classification and seizure detection. Further, we compare the proposed method to other comparable methods.