Graph Attention Based Spatial Temporal Network for EEG Signal Representation

Graph attention networks (GATs) based architectures have proved to be powerful at implicitly learning relationships between adjacent nodes in a graph. For electroencephalogram (EEG) signals, however, it is also essential to highlight electrode locations or underlying brain regions which are active when a particular event related potential (ERP) is evoked. Moreover, it is often im-portant to identify corresponding EEG signal time segments within which the ERP is activated. We introduce a GAT Inspired Spatial Temporal (GIST) net-work that uses multilayer GAT as its base for three attention blocks: edge atten-tions, followed by node attention and temporal attention layers, which focus on relevant brain regions and time windows for better EEG signal classification performance, and interpretability. We assess the capability of the architecture by using publicly available Transcranial Electrical Stimulation (TES), neonatal pain (NP) and DREAMER EEG datasets. With these datasets, the model achieves competitive performance. Most importantly, the paper presents atten-tion visualisation and suggests ways of interpreting them for EEG signal under-standing.