Algorithms For Brain Signal Analysis And Multi-model Brain Computer Interfaces

The goal of a Brain-Computer Interface (BCI) consists of the development of a uni-directional interface between a human and a computer to allow control of a device onlyvia brain signals. Based on the brain activity signals during the execution of one or moremental tasks by a user, a computer system translates those signals into control commandsfor external device control through preprocessing, feature extraction and classifcation.BCI not only has signifcant practical value for severely paralyzed patients, but also its re-search process and the solution of key technology problems can promote the developmentof diferent discipline areas. This may stimulate new ideas and explore new techniquesfor a new direction. Hence, the BCI technology research has both the scientifc and thepractical value. In this thesis, the main objective is the algorithm design and develop-ment to improve the performance of BCI. The contribution of this thesis mainly consistsof the following two points:First, the algorithm study on the two brain signals of EEG and fMRI: The algorithmstudy is the most important factor in the BCI. Considering the algorithm study, weproposed a semi-supervised learning method for the EEG signal analysis and a multi-variable pattern analysis (MVPA) based on sparse representation for voxel selection inthe fMRI data analysis.(a) For improving the stability and adaptation of the BCI system, we proposed asemi-supervised learning algorithm for joint parameter selection, feature extraction andclassifcation. The parameters of EEG-based BCI consist of the time-frequency windowand the channel. The problem of these parameters selection is hard to tackle as optimalparameters may vary between subjects and even between sessions with the same subject,specially when the training data is small. The semi-supervised learning algorithm pro-posed in this thesis can use the test dataset without labels for parameters selection andhas been applied in the on-line BCI. Finally, we applied this algorithm in both motorimagery and P300based BCI. The analysis results show that the efective parameters canbe selected to improve the performance of BCI especially in the case of small trainingdataset. Furthermore, the stability and the adaptation of the BCI system have beenimproved with the proposed semi-supervised learning method.(b) Compared with the single-variable analysis method, MVPA can handle a spatial pattern consisted of multiple voxels signal and it provides higher sensitivity for detectingcognitive representations in the brain. Hence, a MVPA method based on sparse repre-sentation is proposed for the voxel/feature selection and decoding in this thesis. Themain benefts of this feature selection method include: being able to identify all of theinformative features, instead of a highly discriminative few, and being able to separatelymap features onto the class to which they are diferentially active. Finally, we presenttwo experiments and demonstrate the efectiveness of our method through data analysis.The frst contains two simulation, the second is an fMRI experiment for face recognitionin human brain.Second, the application of hybrid brain computer interface using diferent singlemodal EEG component: A hybrid BCI through combining motor imagery and P300in aspecial guided user interface (GUI) has been proposed and successfully applied to solvethe problems in the single modal BCI system, e.g., few independent control signals orlong interval for two consequent control signals output. We also applied this hybridBCI for the mouse control and the simulated wheelchair control. In the mouse control, ahybrid EEG-based BCI system that combines motor imagery and P300for sequential2-Dcursor movement control and target selection was successfully implemented. Specifcally,the horizontal and vertical movements of the cursor are separately and simultaneouslycontrolled by motor imagery and the P300potential. Whereas the target selection isperformed through combining the patters of motor imagery and P300. In the simulatedwheelchair control, subjects can steer the simulated wheelchair efectively by controllingthe direction and speed with our hybrid BCI system. Specifcally, our paradigm allows theuser to control the direction (left or right turn) of the simulated or real wheelchair usingleft-or right-hand imagery. Furthermore, a hybrid manner can be used to control speed.To decelerate, the user imagines foot movement while ignoring the fashing buttons onthe GUI. If the user wishes to accelerate, then he/she pays attention to a specifc fashingbutton without performing any motor imagery. The hybrid BCI system described herewon frst place from the AUTOMATIC CAR CONTROL event at the BCI competitionheld in China (2010), which was organized by Tsinghua University..