Classification between mental arithmetic and motor imagery based on historical data, a study of brain-computer interface using functional near-infrared spectroscopy

Recent years, the functional near-infrared spectroscopy (fNIRS) has been more and more recognized as an important approach for brain computer interface (BCI) due to its advantages e.g. high spatial resolution compared to EEG, robust to electromagnetic noises and motion artifacts, and etc. In this paper, the feasibility and validity of a fNIRS-BCI binary classification method based on the historical data training strategy were investigated. 10 male participants were enrolled and they carried out two classical paradigms, mental arithmetic and right-hand motor imagery, over three experimental sessions. The mean value of oxygenated haemoglobin (HBO) was chosen as the main classification feature and SVM was the classifier. Results showed that the average classification accuracy for session 2 was 67.2% when data from session #1 were used as training set and that of session #3 was 69.3% when combined data from session #1 and session #2 were used as training set, both significantly higher than binary random probability, indicating that the selected two paradigms may be classified effective using historical data. Moreover, an iterative training method (including historical data) was proposed and the average classification accuracies of session #2 and session #3 were promoted, respectively to be 72.3% and 70.0%, showing that the proposed approach was valid and good, and even better especially when the training set was small. These findings may provide useful theoretical and practical experiences for real applications of fNIRS-BCI.