Using Brain Network Features to Increase the Classification Accuracy of MI-BCI Inefficiency Subject

Abstract

Motor imagery-based brain–computer interface (MI-BCI) inefficiency phenomenon is one of the biggest challenges in MI-BCI research. BCI inefficiency subject is defined as the subject who cannot achieve classification accuracy higher than 70% since 70% is considered to be the minimum accuracy for communication by BCI. About 15–30% of the people are MI-BCI inefficiency according to the investigation. Most of the existing studies used common spatial patterns (CSP) to extract motor imagery feature and identify MI-BCI inefficiency subject based on the obtained classification accuracy. We think the MI-BCI performance maybe suppressed because CSP mainly extracts event-related desynchronization (ERD) feature, while the features generated by motor imagery are more than that. In this current work, we screened a total of 12 MI-BCI inefficiency subjects by CSP feature firstly, and recorded the motor imagery EEG data of them. Furthermore, we constructed a task-related brain network by calculating the coherence between EEG channels, the graph-based analysis showed that the node degree and clustering coefficient have intensity differences between left and right hand motor imagery. Finally, the two kinds of features were used to discriminate the two tasks. The results showed that both node degree and clustering coefficient features perform better than CSP, and the feature combination of brain network and CSP achieved higher accuracy than a single feature. In particular, a total of four subjects achieved accuracy higher than 70% by node degree and CSP features fusion. This work demonstrates that the accuracy of the MI-BCI inefficiency subject can be increased by using the brain network feature, but the accuracy gains are not high enough; it is worth to try other types of feature extraction algorithms for the MI-BCI inefficiency subject.