Orthogonal matching pursuit-based feature selection for motor-imagery EEG signal classification

This paper focuses on a framework that uses a small number of features to obtain high-quality classification accuracy of left/right-hand movement motor-imagery EEG signal. Motor-imagery EEG signal has been filtered, and suitable features are extracted using a temporal sliding window-based approach. These extracted features from overlapping and non-overlapping approaches are further compared based on three different types of feature extraction techniques: band power, wavelet energy entropy, and adaptive autoregressive model. The overlapping segments with wavelet energy entropy provide the best classification accuracy over other alternatives. The obtained classification accuracy is 91.43%, the highest ever reported accuracy for BCI Competition II data set III. Subsequently, Orthogonal Matching Pursuit (OMP) technique is used to select the subset of most discriminating features from the entire feature-set. It reduces the computation cost but still retains the quality of the classification results with only 1.43% information loss (that is, 90% classification accuracy), whereas the features-set size reduction is 75% for the same. It is found that the wavelet energy entropy technique performs consistently well in all the variants of our experiments and obtains a mean accuracy difference of 0.95% only.