fNIRS Feature Extraction and Classification in Grip-Force Tasks

It has been of increasing interest to decode the movements based on the changes in hemoglobin concentration in the cerebral cortex that can be measured by functional near-infrared spectroscopy (fNIRS). Existing work were mainly focused on motor imagery tasks for the advantages of fNIRS on cognitive activation detection, and motor tasks were not well studied. This paper investigated fNIRS feature extraction and classification methods for grip-force tasks. Two methods were implemented and compared. The first method took the classic machine learning approach. Features were extracted with respect to the physiological characteristics of fNIRS signals, including the mean, area and slope features. Then, support vector machine was used for classification of grip-force tasks with the extracted features. The second method took the deep learning approach without extracting features. An LSTM-based network was constructed to take in the fNIRS data and output the classified task type. The two methods were tested and compared with an grip-force task fNIRS dataset. The classic machine learning method with extracted features achieved the accuracy of 71.70%, and the LSTM-based deep learning method achieved 82.86%. The results demonstrate the feasibility and potential of fNIRS-based decoding of motor tasks.