Multi-classification of fNIRS Signals in Four body parts Motor Imagery Tasks Measured From Motor Cortex

Functional near-infrared spectroscopy (fNIRS) which is known as a new brain imaging technology has been widely used in brain-computer interface (BCI) because of its convenience and anti-interference capability. Many studies concentrate on two-category motor imagery (MI) tasks classification (left hand and right hand) or four directions (up, down, left and right) measured from prefrontal cortex. In this study, we have designed the experimental paradigms to collect fNIRS data from 10 subjects when they execute four-category MI tasks (left hand, right hand, feet and tongue) measured from motor cortex. We analyze the fNIRS signals by extracting dynamic features under three different time windows and used machine learning method to build Support vector machine (SVM), K-Nearest Neighbor (KNN) and AdaBoost (Ada) classifiers to classify. The results show that the average accuracy of ten subjects was up to 48.5% and for single subject, the highest accuracy was up to 60.62%. Moreover, average hemodynamics responses of each subject measured from motor cortex that reflect different parts of motor cortex were activated when subjects executed different types of MI tasks. To our knowledge, the accuracies we get are higher than the previous research which involved four-direction MI tasks measured from prefrontal cortex. Our results prove that classification of MI tasks based on motor cortex were more effective than that based on prefrontal cortex, which can promote the development of fNIRS-based BCI.