Abstract

Functional near-infrared spectroscopy (fNIRS) measures cortical hemodynamic changes, yet it cannot collect this information from subcortical structures, such as the thalamus, which is involved in several key functional networks. To address this drawback, we propose a machine-learning-based approach to predict cortical-thalamic functional connectivity using cortical fNIRS data. We applied graph convolutional networks (GCN) to two datasets obtained from healthy adults and neonates with early brain injuries, respectively. Each dataset contained fNIRS connectivity data as input to the predictive models, while the connectivity from functional magnetic resonance imaging (fMRI) served as training targets. GCN models performed better compared to conventional methods, such as support vector machine and feedforward fully connected artificial neural networks, on both identifying the connections as binary classification tasks, and regressing the quantified strengths of connections. We also propose the addition of inter-subject connections into the GCN kernels could improve performance and that GCN models are resilient to noise in fNIRS data. Our results show it is feasible to identify subcortical activity from cortical fNIRS recordings. The findings can potentially extend the use of fNIRS in clinical settings for brain monitoring in critically ill patients.