CSP filter calculation of Single Training Pairs and its application in EEG training set optimization

In the research of motor imagery braspatial pattern(CSP), as a supervised spatial filtering method, has been widely used to extract the in-computer interface(MI-BCI) system, common feature of motor imagery electroencephalogram (EEG). But in the process of training data acquisition, the EEG signals may be influenced by different kind of interferences as well as subjects’ distraction and fatigue, thus abnormal training trials with low quality will be generated inevitably. If all training trials are used indiscriminately for CSP filter calculation and classifier training, the performance of BCI system may be badly degraded. To resolve this problem, this paper proposes a novel and practical method for the training trial selection. The basic steps of the method are, firstly, a pair of two single-trials(one left and one right), which are selected from training dataset in turn, are used to calculate the single-trial-based CSP filters (s_CSP). Then, each s_CSP combined with the zero training classifier is employed to construct the single-trial-based CSP-BCI system, which will be used for self testing on training dataset, and the recognition rates of self testing are used as the index to assess the quality of the single trials for the purpose of bad trial rejection and training dataset optimization. Seventy-five datasets of two-class MIEEG from six subjects collected during different days are used to test the proposed method. Experimental results reveal that the performance of CSP-BCI system is improved significantly compared to the traditional method. The average recognition rate of the six subjects’ testing sets are increased by 5.04%, 6.42%, 13.15% ,15.51%,1.94% and 8.26% respectively.