An improved feature extraction method for Self-paced brain-computer interface application

EEG based motor imagery is widely used in most practical self-paced brain-computer interface systems where the user conveys his/her intents at will whenever they wish to do so, and the system doesn´t tell the user when to perform the mental tasks that convey their intents to the system. The most important step for the self-paced BCI system is to discriminate between characteristic mental activity changes and ongoing EEG, this step is used to detect the motor imagery task state from on ongoing brain activity, and determine the type of motor imagery task when motor imagery related brain activity is detect. In practical application of online signal processing for SBCI systems, numerous features can be extracted from EEG signals, but only a subset of them are really discriminative. In this case, a better feature extraction and selection method becomes critical and leads to many good performance for SBCIs. This paper shows an improved feature extraction and selection method, features are extracted by stationary wavelet transform and bandpass filtering, and support vector machine classify these extracted feature vectors. Our method selects the EEG features and classifier parameters by genetic algorithm. We test our methods in the BCI competition 2008 dataset I, our informal results indicate that our method is efficient for feature extraction and selection in self-paced BCI system.