Multi-Class fNIRS Classification Using an Ensemble of GNN-Based Models

Functional near-infrared spectroscopy (fNIRS) is a neuroimaging technique used to estimate brain activity by measuring local hemodynamic changes. Due to its high spatial resolution, fNIRS is being actively researched as a control signal in the field of brain-computer interface (BCI). Extraction of effective features and accurate classification of signals have always been the focus of research. Previous studies have often converted fNIRS data into images based on the relative positions of the measurement channels and utilized convolutional neural networks (CNN) for classification. However, image representation cannot fully express the non-Euclidean characteristics of the brain signal. In this paper, we propose an approach for single-trial, multi-class fNIRS classification using a graph representation and a graph neural network (GNN). Specifically, a class-specific graph was constructed for each class to incorporate both positional and task-dependent functional connectivity (FC) information. The GNN-based models were then trained on each of the obtained class-specific graphs to have specificity for the corresponding class. Finally, the stacking ensemble learning with a gating network was introduced to weight the models for the final prediction. The proposed method was evaluated on a public dataset consisting of three types of overt movements. The results were compared with baseline models based on support vector machine (SVM) and CNN, using different image conversion methods. The best-performing baseline model achieved an average ternary classification accuracy of 68.71%, whereas the proposed model achieved a classification accuracy of 72.31% for the single model, and 75.47% for the ensemble model.