Feature extraction by common spatial pattern in frequency domain for motor imagery tasks classification

Common spatial pattern (CSP) as a feature extraction algorithm has been successfully applied to classify EEG based motor imagery tasks in brain computer interface (BCI). Successful application of CSP depends on the character of input signals and the first and last m eigenvectors of projection matrix. In this study, we proposed a novel and robust feature extraction method designated frequency domain CSP (FDCSP) that the samples in frequency domain obtained by fast Fourier transform (FFT) algorithm and evenly distributed in 8-30Hz were employed as the input signals of CSP. Besides, we made some modifications to classical CSP to address the inconsistent issue and enhance the generalization ability. Cross validation classification accuracy and standard deviation based on training data were employed as the principle to optimize the subject-specific parameter m. Two public EEG datasets (BCI competition IV dataset 2a and 2b) were used to validate the proposed method. Experimental results demonstrated that the proposed method significantly outperformed many other state-of-the-art methods in classification performance. What's more, samples in frequency domain as the input signals of CSP are demonstrated more robust against preprocessing. Based on the two public datasets, the proposed FDCSP method has potential significance to motor imagery based BCI design in practice.