Several Issues On High-Performance Visual Event-Related Potential-Based Brain Computer Interfaces

In recent decades, as the development of the human brain project which focuses on neuroinformatics, brain-computer interface (BCI) technique has been brought to the forefront by world-wide researchers. A BCI retrieves consciousness-related information from neural signals coming from human brain, and uses the information to directly control peripheral devices, thus building a new pathway between human brain and ambient environment. High-performance visual event-related evoked potential (ERP)-based BCIs aiming at obtaining high accuracies and high output rates, have always been the research focus of ERP-BCI studies. Besides the performance improvement problem of the ERP BCI, the periodic interference problem, the asynchronous control problem, and the online classification problem are also important issues to be considered in theoretical and applied research of the ERP BCI, especially for long-term BCI use. In this dissertation, we explore to resolve these problems mentioned above, from both experimental and signal processing perspectives.A novel ERP-BCI experimental paradigm which uses a new stimulus type called varied geometric pattern flashing stimulus, was proposed and examined. Experiments were carried out to compare the proposed paradigm with the classical paradigm. Experimental and analytical results showed that, for the proposed paradigm, the BCI performance is remarkably improved in terms of classification accuracy, information transfer rate, and written symbol rate; for the proposed paradigm, amplitudes of N2 and P2a components preponderated over frontal-central area, as well as N1 and P2b components preponderated over occipito-temporal area, are significantly enhanced. Furthermore, from the application point of view, we examined the electrode reduction problem under the proposed paradigm. Results showed that, by applying the varied geometric pattern flashing stimulus, the number of electrodes required for classification could be reduced to half, while the classification accuracy could still remain high, comparing with the classical paradigm.A theoretical model was established to describe the periodic interference problem of the ERP BCI. According to the model, it was revealed that the overlapping effect of the ERP responses elicited by the target stimuli could be an important reason for the periodic interference. To solve the model and to further isolate the underlying ERP responses, a Toeplitz method and a difference-wave method were proposed. Effectiveness of the proposed methods were verified through processing of real BCI signals. Results showed that, the target ERP responses could be isolated from averaged non-target signals by the Toeplitz method, while the periodic interference could be effectively suppressed by the difference-wave method. Although the periodic interference is an adverse effect for analysis of ERP components, from another point of view, these quasi-periodic fluctuations could be used to indicate whether the BCI is under control state or not, which may help to resolve the asynchronous control problem. Potential use of the quasi-periodic fluctuations for asynchronous control was then examined through a simulation experiment. As another related research, a phenomenon called the stimulus onset asynchrony (SOA) perturbation phenomenon, which is similar but not the same to the periodic interference, was also examined through theoretical and experimental studies.The adaptive projections subgradient method (APSM) in reproducing kernel Hilbert spaces, which has been reported recently to resolve the general online classification problem, was studied. A modification of the standard APSM method under the case when using linear kernel function (APSM-LM), was proposed and was applied to resolve the online classification problem of the ERP BCI. Effects on online classification performance when tuning algorithm parameters of both the APSM with Gaussian kernel algorithm and the APSM-LM algorithm, were examined based on real ERP-BCI signals. Meanwhile, online classification performance as well as processing time of the APSM-LM algorithm, were compared to those of the classical batch processing algorithm, the step-wise linear discriminant analysis algorithm. Results showed that, the APSM-LM algorithm obtained a classification performance close to that of the batch processing algorithm, while got remarkably increased computational efficiency.