A Least-Squares Approach for Extending Common Spatial Pattern Algorithm to Multi-class in Brain-Computer Interfaces

Common spatial pattern (CSP) algorithm achieves great success in two-class motor imagery based brain-computer interfaces (BCIs). Low information transfer rate is an intrinsic drawback of binary BCIs that limits their practical application. This paper generalizes the CSP algorithm from two-class to multi-class using a least-squares approach. The multi-class CSP algorithm is implemented by approximate joint diagonalization of multiple covariance matrices based on Frobenius norm formulation. Five subjects participated in a BCI experiment during which they were instructed to imagine movement of left hand, right hand or foot. The multi-class CSP algorithm is applied to the five data sets recorded during the BCI experiment. The averaged classification accuracy of 85.8% is acquired and the operating speed of the algorithm is fast, verifying the usefulness and effectiveness of the method.