Brain State Classification With Group l 1 – norm Sparse PDC as Novel Features for EEG

Abstract

Understanding the brain state with their classification for brain-computer interfacing (BCI) and motor imaginary studies have recently been witnessed as significant research activity. Brain state has a direct correlation with the mental state as well as the involved cognitive task. This paper presents a novel three-fold framework to classify the EEG signal for binary classification problems. The threefold includes feature extraction, feature selection, and feature classification. First, in the feature extraction step, a new sparse parametric feature matrix is obtained from the brain coherence connectivity measure known as partial directed coherence (PDC) from a sparse multi-variate auto-regressive (MVAR) model. The sparsity has been introduced to the MVAR model using the group least absolute shrinkage and selection operator (gLASSO) algorithm. The PDC has been estimated from the sparse MVAR model’s coefficient which is named as sparse PDC (sPDC). Secondly, the features: energy, relative energy, entropy, and conditional entropy are extracted from sPDC, and the relevancy of features are checked by mutual information (MI) algorithm as the most relevant features. In the third step, to identify different brain states involved in specific cognitive activities, the sPDC extracted features are used for the classification of brain states for two groups of control and meditation. The classification has been performed by a polynomial kernel-based support vector machine (k-SVM). The proposed framework provides exciting results indicate reasonable robustness and consistency in the distinction between two groups in almost all participants.