Collective Almost Synchronization Modeling Used for Motor Imagery EEG Classification

Abstract

Classification based on feature extraction is a crucial technique to develop Brain Computer Interface (BCI) systems. The human brain can be considered as a dynamical system, and its behavior measured by EEG signals can be modeled by a group of nonlinear oscillators. Exploring the dynamical nature of EEG signals along with model based approach may improve classification accuracy in BCI. This study proposes a novel feature extraction method for the classification of Motor Imagery (MI) EEG using a dynamical network model operating in a special collective state, so called Collective Almost Synchronization (CAS). The CAS, the nonlinear oscillators set to operate in a weakly coupled regime, can be used to model an EEG. Purpose of this study is to investigate the performance of the CAS model to identify features for the classification of MI states. To achieve this goal, a linear regression method is used and linear coefficients are extracted as feature vectors. Our approach boils down to identifying patterns in the MI-EEG by associating them to the coefficients of a linear regression (or weights of an output function) constructed to model the MI-EEG signals from simulated time-series generated by a dynamical neural network. The dataset 2b from BCI Competition-IV was used to evaluate the performance of the proposed method. Results indicate that the CAS-based classification method is more robust in extracting distinguishable features from EEG signals as compared with other state-of-the-art methods. The proposed method achieved better performance on two-class MI classification. Moreover, the method developed in this study for MI classification across subjects is effective with 74.03% of the accuracy.