A Novel Feature Extraction Method for Motor Imagery Based on Common Spatial Patterns with Autoregressive Parameters

The method of common spatial patterns (CSP) is often used for feature extraction in the electroencephalogram (EEG)-based brain-computer interface (BCI). However, the CSP method requires a large number of electrodes to produce good results. To improve the CSP classification accuracy with a smaller number of electrodes, we introduce a new method of feature extraction named common spatial patterns with autoregressive parameters (CSP-AR). The CSP-AR method not only maximizes the differences between two populations (i.e., right and left motor imagery), but also makes explicit use of frequency information. The data set of BCI Competition II (held by Berlin Brain-Computer Interface in 2003) for motor imagery is used and the experimental results show the CSP-AR has higher classification accuracy of 87.1% than traditional CSP and AR parameters (82.9% and 81.9%, respectively). The method of CSP-AR improves the classification results and has the advantages of high robustness.