A Cross-Session Feature Calibration Algorithm for Electroencephalogram-Based Motor Imagery Classification

The identification of motor intention by analyzing electroencephalogram (EEG) signals has become an important issue in brain-computer interface (BCI) applications. Recently, the features based on the channel covariance measured in Riemannian space have been widely used in doing the motor intention recognition tasks. However, the EEG signal variations across time make the features unstable. In this paper, a two-step feature calibration method for dealing with multiple sessions is proposed to reduce this kind of influence on the classification performance. During feature extraction, the features generated by tangent space mapping are adaptively calibrated with the adjustment of Riemannian center. Then the transfer component analysis is performed to further calibrate the features between two sessions. Experimental results on the BCI Competition IV 2a database show that the classification performance of the proposed method for four types of motor imagery BCI tasks outperform state-of-the-art approaches. It provides a new idea of feature calibration for EEG-based BCI applications.