Hand Movement Recognition Using Dynamical Graph Convolutional Neural Network in EEG Source Space

Brain-computer interface (BCI) has been widely used in the field of medical rehabilitation related to hand movement. However, the single-handed multi-class movement recognition has remained mostly unexplored. In order to improve the accuracy of hand movement classification, an algorithm using dynamical graph convolutional neural network (DGCNN) in electroencephalogram (EEG) source space (sDGCNN) was proposed in this paper. The algorithm firstly maps EEG signals to source space by spatial source localization method. Secondly, time-domain features are extracted from each brain region. Finally, the graphs with brain regions as nodes and extracted features as node values are input into DGCNN for four-classification. The signals in γ band (30−100 Hz) reached the highest accuracy of 90.16 ± 6.8%, which indicates that the high-frequency components of brain may have important significance for hand movement decoding. The result shows that the sDGCNN method significantly improves the accuracy of hand movement classification. The high accuracy also proves the effectiveness of the method in hand movement recognition.