Advancement in Graph Neural Networks for EEG Signal Analysis and Application: A Review

Electroencephalography (EEG) can non-invasively measure neuronal events and reflect brain activity at different locations on the scalp. Early studies for EEG signal processing have focused more on extracting EEG temporal features and considered the topology of EEG channels less due to the limitation of rich spatial information. Graph neural networks (GNNs), as a new kind of deep learning method, can use EEG signals as graph vertices, capturing the hidden topological connections between signals. GNNs have made great progress in EEG studies due to the advantage. In this overview, we review the very new and fundamental models of GNNs and their modifications, such as graph regularized neural networks, graph convolutional neural networks, spatial-temporal graph neural networks, graph attention networks, and their variants in EEG signal analysis fields. The applications of GNNs are summarized in the domains of emotion detection, epilepsy seizure detection, stroke rehabilitation, Alzheimer's disease diagnosis, motor imagery detection, neurological disease diagnosis, major depressive disorder, and driving fatigue detection. We employed a Systematic Literature Review (SLR) approach to select 79 papers for a comprehensive review. The current state is analyzed and forecasts are provided based on the available difficulties. We conclude by suggesting potential directions for future research in this rapidly developing topic.