Multi-class stationary CSP for optimal feature separation of brain source in BCI system

Electroencephalogram (EEG) record brain activation into electric millivolts. Brain computer interface based on EEG signal recognized brain control limbs movement is one of the main application of BCI system. Classification attention of user when moving limbs find out the brain region activating and translate them into commands which could control outer device. Recently, a famous technique to analyze EEG signals based on spatial filter is common spatial patterns (CSP). It is popular for two-class paradigm when maximize one class in the same time minimize the other one. In the other hand, CSP gets limitation when just working on covariance matrices in which not only stationary brain signal sources but also contaminated non-stationary sources. In this paper, we proposed to applied extension of CSP to separate non-stationary and also apply for multi-class BCI to reduce these disadvantage mentioned above. To solve non-stationary sources problem, we applied stationary CSP (sCSP) to separate signal sources. sCSP method is also powerful for binary paradigm. To improve sCSP for multi-class, we applied joint approximate diagonalization (JAD), which is successful to find efficient spatial filter in the context of multi-class BCI. Since CSP supposed to separate data space linearly, nearest neighbor method was used to classify the multi-class BCI to evaluate the performance of methods. We used 2 sorts of dataset: 1) auditory cue to two patients respond to three kinds of sound for executive movement; 2) auditory spatial from 2 speakers separate sounds. Three kinds of sound are: 500Hz left hand, 2000Hz right hand and noise sound ignored.