Spatio-Temporal Variable Structure Graph Neural Network for EEG Data Classification

This paper proposes a strategy for improving the correct diagnosis of epilepsy based on electroencephalogram (EEG) using a spatio-temporal variable structure graph convolutional neural network. Specifically, this method is called the variable-structure graph convolutional neural network (VGCRN), which is derived by combining spatial information and noise removal through variable structured graph convolutional neural network and temporal information through recurrent neural network. Despite the potential benefits of EEG for diagnosing and monitoring neurological conditions, the low signal-to-noise ratio often hinders timely and accurate diagnosis in many clinical cases. Previous research on EEG data classification has mainly focused on extracting features from the time or frequency domain, disregarding the spatial features among electrodes. EEG can be viewed as a structured time series, consisting of multivariate time series data with prior information provided by the spatial location of electrodes on the patient’s scalp. Spatial information is just as crucial as time or frequency-domain information, but introducing unconstrained spatial features in topological map structures can result in noise and the aggregation of irrelevant information by nodes. The proposed method in this paper can better leverage the spatial and intrinsic temporal information of brain waves while reducing noise, thus enhancing the robustness and accuracy of the model.