Integrated Convolutional and Graph Attention Neural Networks for Electroencephalography

In the present study, we propose a novel deep neural network (DNN) model designed to enhance the performance and interpretability of DNN based on the attention module. The proposed EEG-Graph Attention Network (EEGAT) consists of the convolutional neural network (CNN) and graph attention network (GAT). We evaluated the EEGAT using three heterogeneous datasets: Fatigue, DEAP, and BCI Competition IV 2a. Convolutional kernels in the EEGAT extract temporal and spatial features from the minimally preprocessed EEG signals. The extracted spatiotemporal features were then constructed as node features of the GAT layers and subsequently updated. The proposed EEGAT outperformed the alternative DNN models without GAT across the datasets and enhanced the interpretability of the prediction performance.