The Few-Channel BCI System Research Based On EEG-NIRS

Brain-computer interface (BCI) is currently a widespread research direction aimed to establish a communication bridge between the brain and the external environment. Current BCI studies are increasingly focused on practical and wearable technologies. BCIs based on motor imagery have received much attention because these are an independent type of BCI, which means that they do not require the aid of external stimuli. Furthermore, because their signals are independent and complementary, bimodal BCIs are a possible platform for few-channel portable BCI systems.To examine the portability and wearability of BCIs, this study investigated the whole process of building electroencephalography (EEG) - near-infrared spectroscopy (NIRS) bimodal few-channel BCIs, from experimental paradigm to pattern classification. The following findings were obtained in this study:(1) First, brain activation during left and right hand motor imagery was located with EEG and NIRS signals. Brodmann area 6 in the supplementary motor area was the most activated. Channels from this location were chosen for the EEG-NIRS BCI system. The EEG channels were C3, CZ and C4 in the 10-20 system, while three pairs of emitter and receiver poles were arranged surrounding C3 and C4 for NIRS.(2) To better extract features for few-channel BCIs, the phase space reconstruction (PSR) method was used on 3-channel EEG extensions before the common spatial pattern (CSP) method of spatial filtering. To test the validity of the PSR + CSP method, the study used dataset Ⅲa from BCI Competition 2005. The accuracies for 3-channel data via CSP feature selection were 43.0% while the accuracies for 3-channel data via PSR + CSP feature selection were 73.9% with the same classifier. As for the data achieved in this paper, the accuracies for 3-channel data via PSR+CSP feature selection were 15.0% higher than accuracies for 3-channel data via PSR + CSP feature selection. This indicated that the PSR + CSP method can effectively extract features from few-channel BCIs.(3) This paper discussed bimodal signal fusion for the EEG-NIRS BCI, both the feature level fusion with support vector machine (SVM) and the decision level fusion with back propagation (BP) neural networks were tried. The method SVM which was used as a fusion and classifier for EEG-NIRS signals did a better job. The average accuracy of the fusion feature method was 81.2%, which was higher than either single-mode correction rate. The highest accuracy was 100%. Additionally, owing to the data fusion technology, the subject who had little response to one brain imaging technique could be replenished in another brain imaging technique. The lowest accuracy rate for EEG single mode was 58.3% and for NIRS single mode was 43.1%. It is surprising that the lowest accuracy rate increased to 75.0% for bimodal BCIs. Feature fusion and classification were done at the same time as SVM. With this method, the time of signal processing was reduced and the response speed of the system was improved. Therefore, using fusion data from bimodal brain imaging technology to establish a BCI system improved the spatial coverage of the system and reduced the degree of fuzzy information in system. In summary, the bimodal BCI system is robust, highly accurate and effective in reducing BCI blind.