A Semi-Supervised Classification Method for Motor Imagery EEG

A sequential updating semi-supervised classification based on training iterations was proposed for the two-class motor imagery task in brain-computer interface. Firstly, making use of multiple autocorrelation analysis, the training samples with high confidence were selected as initial training set. Then common average reference and common spatial pattern were used for pre-processing and feature extraction, respectively. Lastly, support vector machine was applied to test new samples. The samples with low confidence were removed successively according to the iterations. The remains were used to retrain the model to optimize the parameters of both feature extractor and classifier. The proposed method was applied to Dataset IIa of BCI Competition IV to verify its validity. The results show that the classification accuracy is higher than other algorithms on the occasion where training samples are not enough. It can provide a new solution for the real-time BCIs.