fNIRS Signal Classification Using Graph Representation for Brain-Computer Interface

Abstract

Brain-computer interface, which enables interaction between brain and device by interpreting brain waves, is expanding its scope from the rehabilitation to practical or entertainment. fNIRS is actively researched in that it has complementary characteristics with EEG, and most studies apply convolutional neural networks after converting signals into images to extract spatial features. In this study, we propose that graph representation is more suitable since it can freely reflect the neurovascular association between fNIRS signals. A graph representation method reflecting location and functional connectivity was presented, and performance evaluation was conducted using various graph neural networks. With proposed GCN module and multi-branch model, the ternary classification accuracy of the left-right hand and foot motor task reached up to 75.46%, whereas the highest validation accuracy of the CNN-based model remained at 70.84%.