Subdomain Adversarial Network for Motor Imagery EEG Classification Using Graph Data
Xingchen Li, Xianlun Tang, Sichao Qiu, Xin Deng, Huiming Wang, Yin Tian
Abstract
Motor imagery Electroencephalogram (EEG) signals have been widely used in the field of brain-computer interface (BCI) due to their advantage of non-invasiveness and easy acquisition. However, due to distortions in the temporal and local information of the EEG signal and inter-subject variability, it is very time-consuming to perform a calibration procedure designed in a subject-specific manner, which requires a large number of labeled samples. To this end, we construct EEG brain graph based on EEG electrode distribution and propose a new subdomain adversarial network for learning domain-invariant features to achieve cross-domain classification of motor imagery EEG signals. Specifically, the proposed method constructs a brain graph based on the spatial distribution of electrodes and their functional connections, and builds EEG graph data. The subdomain adversarial network aligns data distributions under the same class domain and improves the classification performance of the target domain using source domain labeled samples. We have conducted extensive experiments on two publicly available datasets to verify the effectiveness of the proposed method. In addition, we have done ablation study and visualization analysis to explain the contribution of each component to cross-domain classification.