DEEP LEARNING FOR MOTOR IMAGERY-BASED BCIS USING SEEG SIGNALS

Motor imagery (MI) is the most popularparadigm for brain-computer interfaces (BCIs) based onscalp electroencephalography (EEG), while thisparadigm is missing for stereo-electroencephalography(sEEG)-based BCIs. Recently, the first public dataset ofsEEG has become available for MI-based BCIs.However, the performance using traditional methods isstill inferior. In this study, we employed some state-of-the-art methods based on deep learning to improve theclassification accuracy of MI for sEEG-based BCIs. Sixdifferent deep learning models were developed, whichinclude Shallow ConvNet, DeepNet, ResNet20,conformer, vision transformer (ViT) and ViT with pre-trained parameters. Among six deep learning models, weachieved an average accuracy of 0.83 in the handopen/closed binary classification task with the conformermodel. Compared to the available work, our approachdemonstrated a remarkable 16% increase in accuracy.