Sparse optimal score based on generalized elastic net model for brain computer interface

Brain computer interface (BCI) offers disabled people a nonmuscular communication pathway. Event-related potential (ERP) is an efficient way to achieve the BCI system. One of important issues for ERP classification is the under sample problem, that is the feature dimension is very high while the sample number is very strictly limited. In this paper, we introduce a P300 feature extraction and classification framework using the sparse optimal score method for discriminative analysis by generalized elastic net model. In order to break the curse of dimension, regularized estimation of within-class covariance matrix is achieved and ℓ1 penalty is applied to learn sparse discriminant vectors. The optimization problem is solved by the alternating least square procedure. We test the proposed framework on P300 target detection task and experimental results indicate that it is able to improve the classification accuracy in P300-based BCI system. The efficient features extracted by our proposed framework provide overall better P300 classification accuracy than several baseline methods especially in the single trial and few training samples case.