Resting-state and task-state fNIRS fusion assessment method for upper limb motor function in stroke patients.

INTRODUCTION

Post-stroke rehabilitation strategies rely heavily on accurate motor function assessment. Scale-based assessments face inherent limitations, such as inter-rater variability and long assessment time. Neurophysiological biomarkers are emerging as more objective, complementary assessment tools. However, few studies have integrated multi-states functional near-infrared spectroscopy (fNIRS) data to fit the Fugl-Meyer Assessment (FMA) scores of stroke patients with machine learning.

METHODS

This study enrolled 57 stroke patients and acquired fNIRS signals covering the sensorimotor cortex during both resting-state and grip-task conditions. A modified sequential forward selection algorithm was employed to select optimal features from three domains: functional connectivities (FCs) and graph theory metrics of resting-state fNIRS and task-state fNIRS, and clinical characteristics. Support vector regression was then applied on the multi-domain feature set to predict FMA scores.

RESULTS

Both for cortical and subcortical lesion-type stroke patients, the resting-state + task-state fNIRS fused model demonstrated significantly reduced prediction errors compared to single-state models. Optimal performance was achieved at 45% network sparsity [R2 = 0.8434 (0.6889–0.9311), RMSE = 6.75 (4.26–8.78), MAE = 5.11 (3.17–7.00)], with R2 values exceeding 0.73 across all network sparsity levels for cortical stroke. In subcortical stroke patients, the fused model demonstrated superior performance at 35% network sparsity [R2 = 0.7546 (0.5255–0.8728), RMSE = 10.55 (7.51–13.83), MAE = 7.39 (5.35–10.17)]. And the fusion model demonstrated a better fit than models incorporating either clinical or fNIRS datasets alone.

CONCLUSION

Our findings demonstrate that the fusion model integrating resting-state and task-state fNIRS significantly improves motor function assessment accuracy in stroke patients compared to single-state or single-dataset models.