An Unweighted Exhaustive Diagonalization Based Multi-Class Common Spatial Pattern Algorithm in Brain-Computer Interfaces

In binary brain-computer interfaces (BCI) based on motor imagery, common spatial pattern (CSP) successfully discriminates two-class EEG data. However, low information transfer rate is an intrinsic drawback of binary BCIs that limits their practical applications. It's essential to extend binary CSP algorithm to multi-class paradigms. In this paper, a new approximate joint diagonalization (AJD) method, named unweighted exhaustive diagonalization with Gauss iterations (UEDGI) is proposed for the extension. The UEDGI based multi-class CSP algorithm is applied to five data sets recorded during motor imagery of left hand, right hand, foot or tongue. The performance of the algorithm is accessed by classification accuracy and convergence speed, and compared with other two multi-class CSP algorithms, one versus one (OVO) and one versus the rest (OVR). Experimental results show that the UEDGI based multi-class CSP performs best in both classification rate and running speed.