Short-Time Fourier Transform Covariance and Selection, A Feature Extraction Method for Binary Motor Imagery Classification

The brain-computer interfaces (BCI) technology is able to help dysfunctional people recover their motor functions. Electroencephalography (EEG) is an effective noninvasive method to construct BCI. Motor imagery (MI) paradigm can directly reflect the motor intention of user without additional stimulation equipment in EEG-BCI. The feature extraction methods are the key components to improve the accuracy of MI classification. Traditional feature extraction methods like CSP that only extract features in a single domain or two domains. In this study, we propose two novel feature selection method, short-time Fourier transform covariance and its selection method, which are aiming to extract spatial-time-frequency features simultaneously. In order to evaluate the proposed features, the BCI Competition IV Data Set IIb is employed to test the classification accuracy. By comparing the average accuracy of novel TSGSP method, the proposed method is more stable than TSGSP about 6% and the accuracy is just decrease about 0.5% at the same time. The average accuracy of 83.8% over all subjects is obtained. Superior classification performance results show that our proposed method has great potential, which is helpful for the further development and application of BCI technology for motor imaging in the field of neurorehabilitation.