Improving the Recognition Accuracy by Solving the Inherent Data Imbalance Problem of ErrP with Generative Adversarial Network

Brain-computer interface (BCI) has broad application prospects in rehabilitation, neural prosthesis, and exoskeletons. Current electroencephalography (EEG) based BCIs, especially motor imagery (MI) based BCIs, suffer from low recognition accuracy due to their limited signal-to-noise ratio (SNR) and high non-stationarity, which hinders their practical applications. Integrating error-related potential (ErrP) to construct a hybrid BCI and correct the recognition results of the main BCI modal is an effective way to improve the overall performance of BCI system. However, the inherent data imbalance of ErrP leads to the unbalanced classification accuracy, in which the recognition accuracy is low in error trials that makes the system cannot efficiently correct the classification results of the main BCI mode. This study constructed a generative adversarial network (GAN) and used it to generate new data to address the data imbalance of ErrP for the first time. An EEGNET was realized to assess the classification result of the proposed method. The quantitative assessment indicates that the constructed GAN works well in generating new ErrP data. Statistical analysis shows that the proposed method simultaneously improves the degree of inter-class balance of the accuracy and the overall accuracy. The proposed method enhances the self-correction ability of BCI and facilitates its practical application.