Research On The Neural Mechanisms And Recognition Algorithms Of Motor Imagery-based Brain-computer Interface

Motor imagery-based brain-computer interface(MI-BCI) provides an substitutive motor output pathways without peripheral neural system, and it has great value in motor assisting and neurological rehabilitation. Many labratories have developed online MI-BCI system which support real-time control commard output, but due to the infuence of several factors such as recording electrode technique, non-stationary electroencephalogram(EEG) signal, artifacts, classification stability, inter-subject variations, and etc, most of the developed online MI-BCI are only used in laboratory environments. In the current study, we first investigate the neural mechnisms of MI-BCI control based on resting-state EEG and functional magnetic resonance imaging(f MRI), then develop a more robust feature extraction method and a high performance pattern recognition method. The main contents of this dissertation are as follows:1. The MI-BCI control performance varies across subjects, we propose to use resting-state EEG spectral entropy as a biomarker to predict individual SMR-BCI performance, which derived from 2 minutes eyes closed resting-state EEG of channel C3. The correlation coefficient between spectral entropy predictor and MI-BCI control performance is 0.65. Intra- and inter-session classification results demonstrate that spectral entropy predictor provide good classification capability for high and low aptitude MI-BCI users, and the corresponding accuracies are 82% and 89%, respectively. To our knowledge, there has been no discussion about the reliability of inter-session prediction in previous studies. The proposed predictor could help to identify subject’s potential MI-BCI performance at the very beginning, avoiding the frustrating and costly training procedures for those inefficiency users.2. We further investigate the MI-BCI control performance variations through the perspective of resting-state EEG network. Results show that the spatial topologies of the network have close relationships with MI-BCI control performance that the mean functional connectivity, node degrees, edge strengths, clustering coefficient, local efficiency, and global efficiency arepositively correlated with MI-BCI control performance, whereas the characteristic path length is negatively correlated with MI-BCI control performance. The above results indicate that an efficient resting-state EEG network facilitates MI-BCI control performance. Specifically, a multiple linear regression model was adopted to predict subject’s MI-BCI control performance based on the efficiency measures of the resting-state EEG network, the correlation coefficient between the predicted MI-BCI control performance and actual performance is 0.75, and the root mean square error is 10.5%.3. Based on resting-state f MRI data, the individual functional connectivity density(FCD) maps with voxelwise spatial resolution are calculated. High MI-BCI control performance group shows stronger long-range FCD in bilateral middle occipital gyrus, whereas low MI-BCI control performance group exhibits stronger long-range FCD in insula and putamen. These findings may help understanding the brain network mechanisms of MI-BCI control.4. Local temporal correlation common spatial patterns(LTCCSP) is proposed for feature extraction of MI-BCI. LTCCSP introduces LTC information to the covariance matrices estimation procedures of CSP, and improves robustness of the estimated spatial filters. Results on simulation dataset show that LTCCSP achieves the highest average classification accuracy in all the outliers occurrence frequencies compared with other two methods. On the real MI-BCI dataset, LTCCSP also achieves the highest average classification accuracy. The above results indicate that LTCCSP could extract motor imagery related EEG features effectively, and it has a higher robustness.5. Z-score linear discriminant analysis(Z-LDA) is proposed for classification method of MI-BCI. Classical LDA only uses mean of the projected training samples to define the classification boundary, whereas Z-LDA uses mean and standard error of the projected training samples simultaneously to define the classification boundary. Z-LDA can adaptively adjust the classification boundary to fit for the heteroscedastic distribution situation. Results from both the simulation dataset and two real MI-BCI datasets consistently showthat Z-LDA achieves significantly higher average classification accuracy than conventional LDA, indicating the superiority of the new proposed classification boundary definition strategy for the classification problem of MI-BCI.In conclusion, the current dissertation establishes a spectral entropy biomarker based on resting-state EEG for predict subject’s MI-BCI control performance, then investigates the neural mechanisms of MI-BCI control from the perspective of resting-state EEG/f MRI brain network; and for the signal processing methods in MI-BCI, we first develop a more robust feature extraction algorithm based on CSP, then propose a new classification boundary definition strategy to improve the accuracy. The above work may helpful for transferring MI-BCI beyond the laboratory.