Extracting features from phase space of EEG signals in brain–computer interfaces

Conventional feature extraction methods based on autoregressive and amplitude–frequency analysis assume stationarity in the Electroencephalogram signal along short time intervals in Brain–Computer Interface (BCI) studies. This paper proposes a feature extraction method based on phase space for motor imagery tasks recognition. To remodel the single nonlinear sequence is to reveal variety information hidden in the original series by its phase space reconstruction, and maintaining the original information continuity. Three phase space features (PSF) datasets are used to classify two Graz BCI datasets. The simulation has shown that the linear discriminant analysis (LDA) classifiers based on the PSF outperform all the winners of the BCI Competition 2003 and other similar studies on the same Graz dataset in terms of the competition criterion of the mutual information (MI) or misclassification rate. The maximal MI was 0.67 and the minimal misclassification rate was 9.29% for Graz2003 dataset by using the combined PSF. According to the competition criterion of the maximal MI steepness, the LDA classifier based on PSF yielded a better performance than the winner of the BCI Competition 2005 and other similar research on the same Graz dataset for subject O3. The maximal MI steepness of O3 is 0.7355.