Initial-dip-based classification for fNIRS-BCI

Abstract

In this paper, the effect of various channel selection strategies on the initial dip phase of the hemodynamic response (HR) using functional near-infrared spectroscopy (fNIRS) is investigated. The strategies using channel averaging, channel averaging over a local region, t-value-based channel selection, baseline correction, and vector phase analysis are examined. For t-value-based channel selection, three gamma functions are used to model the initial dip, the main HR, and the undershoot in generating the designed HR function. The linear discriminant analysis based classification accuracy is used as performance evaluation criteria. fNIRS signals are obtained from the left motor cortex during righthand thumb and little finger tapping tasks. In classifying two finger tapping tasks, signal mean and minimum value during 0~2.5 sec, as features of initial dip, are used. The results show that the active channel selected using t-value and vector phase analysis yielded the highest averaged classification accuracy. It is also found that the initial dip in the HR disappears in case of averaging overall channels. The results demonstrated the importance of the channel selection in improving the classification accuracy for fNIRS-based brain-computer interface applications. Furthermore, the use of three gamma functions can also be useful for fNIRS brain imaging for detecting the initial dip in the HR.