Conditional random fields as classifiers for three-class motor-imagery brain-computer interfaces

Conditional random fields (CRFs) are demonstrated to be a discriminative model able to exploit the temporal properties of EEG data obtained during synchronous three-class motor-imagery-based brain-computer interface experiments. The advantages of CRFs over the hidden Markov model (HMM) are both theoretical and practical. Theoretically, CRFs focus on modeling latent variables (labels) rather than both observation and latent variables. Furthermore, CRFs' loss function is convex, guaranteeing convergence to the global optimum. Practically, CRFs are much less prone to singularity problems. This property allows for the use of both time-and frequency-based features, such as band power. The HMM, on the other hand, requires temporal features such as autoregressive coefficients. A CRF-based classifier is tested on 13 subjects. Significant improvement is found when applying CRFs over HMM- and LDA-based classifiers.