Canonical correlation analysis of EEG for classification of motor imagery

The performance of classification of various mental states using Electroencephalography (EEG) is often limited by the lack of information regarding the most discriminative channels and frequency bands. The paper proposes a Canonical Correlation Analysis (CCA) of EEG recorded during bilateral imagined hand movement. CCA determines linear transformation of EEG that is maximally correlated with a transformed version of neural response to imagined movement. In the proposed method, the linear weights are identified for the spatial and spectral components of EEG signal. The study investigates how the CCA-based transformation of the signal improves discrimination between two movement classes. Further, two parallel CCA blocks, determine two transformations of EEG, one that maximizes correlation with the corresponding class, and other that minimizes the correlation with the mismatched class. Features derived using this approach are used for single-trial classification of bilateral imagined hand movement. Time-frequency patterns of EEG derived from the proposed approach illustrate their discriminative ability in mu and beta bands. An average classification accuracy of 61.12% (109 subjects) and 73.34% (Best 40 subjects) are obtained. The results indicate the scope of CCA to obtain time-frequency representations of EEG and for single-trial classification of motor imagery (MI).