Machine Learning Method Evaluates Upper-Limb Motor Function by fNIRS in Stroke Patients

The upper limb motor impairment after stroke seriously affects patients’ quality of life, and the evaluation of upper limb motor function depends on professionals, which is subjective, complicated, and time-consuming. Neurophysiological information evaluation Fugl-Meyer assessment of the upper limb (FMA-UL) is expected to solve this problem. Our study aims to estimate FMA-UL using the functional connectivity measures and network properties of task state functional near-infrared spectroscopy (fNIRS). FNIRS was recorded when the subjects performed the hand-grasping task, representing the oxyhemoglobin concentration changes. Functional connectivity was gauged via Pearson correlation coefficients, selectively retained within a 10% to 50% threshold. We calculated two macroscale and two mesoscale network properties based on graph theory. Then the support vector regression (SVR) was applied to estimate the FMA-UL. Cross-validation with leave-one-out, resulted in an R$^{\mathbf{2}}$ of 0.49, with a root mean square error of 4.96, 5 of the 15 subjects had absolute errors less than 0.75, when the sparsity of functional connectivity is 40%. Functional connectivity and network properties by fNIRS is a promising method to predict motor impairment. Our study demonstrated that functional connectivity and network properties are correlated with motor impairment and brain reorganization in stroke patients. The method proposed reflects prospects for finding neural biomarkers to predict FMA-UL and is feasible for using fNIRS to predict the motor function of stroke.