Research And Design For Online Brain-Computer Interface Based On Decision Theory

Brain-Computer Interface (BCI) is a new path which does not dependent on the neuromuscular organization, and it provides new means for the patients with dyskinesia to interact with the outside world. The emergence of BCI for the purpose of clinical application in recent years has been researched in the field of non-clinical development. Therefore, online BCI system, which has a good performance of real-time feedback in the premise of ensuring the accuracy, has a very important significance.The Design of BCI system includes two parts, offline analysis and online realization. At present, although there are many EEG signal processing methods which obtaining high accuracy in the competition data sets, for the operation speed, complexity of the algorithm or other reasons, some of them were not practical. Therefore, the performance of the algorithm is a guarantee to achieve online system. This article focus on establishing motor imagery-based BCI. and achieving the control cursor in the case of real-time feedback provided to the subjects. To fulfill the task mentioned above, an autoregression (AR) model and adaptive Morlet wavelet basis as feature extraction methods are applied. Based on decision theory we propose two classifiers. Sequence Probability Ratio Test (SPRT) and Sequential Linear Discriminant Analysis (SLDA), both classifiers can conduct the dynamic classification. The data sets from BCI competition show adaptive Morlet wavelet could yield better result than AR model. And with the same feature extraction method, two kinds of classifiers result with equivalent level.To verify the real-time performance and practicality of our algorithm, left or right hand motor imagery-based BCI experiment and paradigm are designed on BCI2000software platform based on event-related desynchronization/synchronization. The offline analysis tool from BCI2000is used for obtaining the most distinguish channels between two kinds of mental tasks. Then, feature extraction, using adaptive wavelet basis in the first experiment data of4subjects, won72.88%of the average accuracy SPRT, SLDA won76.21%of the average accuracy rate. And less decision time are needed for SLDA in SPRT. Based on offline analysis, two subjects with high correct rate are selected participate in cursor control experiment; adaptive wavelet basis and SLDA are used as real-time algorithm. After training, the accuracy rate of two subjects is higher than80%.