Brain Network Features Analysis And Extraction Of Motor Imagery EEG Signals

Brain-computer interface is a control technology that facilitates the communication between brain and external device,which has been widely applied in the fields such as motor dysfunction repairment,medical diagnosis and so on.Motor imagery-based brain-computer interface(MI-BCI)does not depend on external stimuli,but rather on the subject's spontaneous brain activity to finish the required task.Due to the characteristics of MI-BCI,many labratories have worked on it for decades.In the current thesis,brain network is constructed to calculate and analyze the network characteristics based on two-type MI electroencephalograph(EEG),then the related features are extracted based on the above network,finally support vector machine(SVM)is used to classify the two tasks.The main contents of this thesis are as follows:(1)EEG signal acquisition and pre-processing.EEG signals are obtained from twelve subjects that initially participate in the MI experiment.Signal pre-processing such as reference electrode standardization,baseline removement,bandpass filtering,data segmentation,and artifact rejection are performed on the EEG data.(2)Brain network construction and feature analysis.In this section,15 electrodes are selected as network nodes.Coherence coefficients and phase-locking values between nodes are calculated and regarded as connection edges of the network.The graph theory analysis is used to obtain the network properties including node degree,clustering coefficient,characteristic path length and global efficiency.Results show that node degrees and clustering coefficients calculated from electrodes on the right sensorimotor cortex are significantly larger than those on left sensorimotor cortex.The node degrees and clustering coefficients of left-hand MI are higher than the right-hand MI on the right sensorimotor cortex for 67% of the subjects.The characteristic path length of left-hand MI is shorter than the right-hand MI for 67% of the subjects.The global efficiency of left-hand MI is higher than the right-hand MI for 58% of the subjects.The above results indicate that there are differences in the network properties between the left/right-hand MI.(3)Classification based on brain network feature.Seven feature sets that obtained from the brain functional network and common spatial pattern(CSP)feature extraction algorithm are used respectively as the input features to SVM for classification.The obtained classification accuracy results are compared and analyzed.The results indicate that the average classification accuracies of the network-based feature sets and joint feature sets are higher than feature set that extracted by CSP.Particularly,the average classification accuracy using features combining network-based features and adjacent matrix spatial features is 66.97%,which is7.01% higher than that using CSP features,and the highest classification accuracy of individual subject is 81.49%,which is 15.13% higher than that using CSP features.In conclusion,the results in this thesis indicate that MI EEG features based on brain network has potential to increase the classification accuracy of MI-BCI.