On The Information Accumulative Algorithm In The Motor Imagery BCI System

As a new method of human-computer interface, brain-computer interface (BCI) technology based on EEG has become a research hotspot in the fields of rehabilitation engineering and biomedical engineering. BCI technology doesn’t depend on normal output channels like peripheral nerves and muscles but builds a direct communication channel between the human brain and external environment. How to extract and classify the electroencephalogram (EEG) signal features fast and accurately is a key issue for the BCI systems, so the study of EEG-based classification and recognition algorithms has great practical significance.Currently, most of the BCI classification algorithms are used to classify EEG signal in a specific time interval, although the statistical analysis of the used time interval may have the best discrimination, but for the whole single trial EEG signals, the EEG information in adjacent time segment are ignored. However, the EEG is non-stationary signal; it cannot take full advantage of EEG information if only consider the specific time interval signals, and cannot achieve a compromise between the decision time and classification accuracy. For this situation, the information accumulative classification algorithm is researched to improve the performance of BCI system.First, based on the events related desynchronization/synchronization phenomenon the frequency spectrum of each dataset are analyzed. Considering the individual differences and the advantages of wavelet in time-frequency analysis, the morlet wavelet was used to extract band power as the classification features.Second, the LDA, SVM and Bayesian classification methods were extended to sequential algorithm based on the information accumulation. The experimental results showed that for the left and right hand motor imagery EEG classification problems, sequential classification algorithm can get a higher classification accuracy, mutual information and Kappa coefficient, but the above sequential algorithms cannot achieve the dynamic classification. In order to achieve dynamic classification, an EEG dynamic classification algorithm based on the sequential probability ratio testing (SPRT) was proposed to classify the motor imagery trails. The unique advantage of the SPRT method lies in its monotone evidence accumulation process, the ability to increase the discriminative power with more evidence observed over time. The results show that the SPRT cannot only improve the classification accuracy and decision speed, but also can balance the time-accuracy trade off.Third, based on the multi-class motor imagery dataset in BCI competition Ⅲ, whether the SPRT algorithm is suitable for multi-class BCI system was studied. The SPRT which combined with Morlet wavelet and common spatial pattern (CSP) feature extraction method was used to classify multi-class EEG signal. Useful experience was gained for the multi-class EEG classification problem.