Spatio-temporal Graph Neural Networks for Motor Imagery EEG Classification

EEG has several applications in neuroscience, neural engineering, and diagnosis of neurological disorders. It records the electrical activity of the brain tissues resulting from the interaction of distinct neuronal populations. Various methods have been developed for the automatic classification of EEG signals. These algorithms operate by acquiring unique and non-redundant features from EEG data. However, a large portion of the proposed algorithms usually use temporal features while ignoring the dense spatial network structure generating the EEG. We propose a classification pipeline that utilizes the network structure of EEG data for a simultaneous use of spatial and temporal features. First, we project the scalp EEG data onto Schaefer's brain atlas consisting of hundred parcels which are distributed into seven networks. Second, we propose an architecture for a spatio-temporal graph convolutional neural network that combines a graph convolution for learning spatial features and a single dimensional convolution for learning temporal features. Additionally, we examine as to how much each individual network contributes to the classification accuracy. We demonstrate that our model outperforms the standalone temporal convolution models used for EEG classification.