Graph Convolutional Network-Based Fusion of Multi-State fNIRS Data for Assessing Post-Stroke Upper Limb Motor Function.

Conventional neuroimaging tools for post-stroke motor function evaluation (e.g., EEG, fMRI) have some constraints. Conversely, functional near-infrared spectroscopy (fNIRS) offers a viable compromise. Nevertheless, few studies have yet quantitatively assessed the current motor function scores based on fNIRS data. This study proposed a Graph Convolutional Network (GCN) and Support Vector Regression (SVR) fusion model to fit Fugl-Meyer Assessment (FMA) scores by leveraging a multi-state integration of fNIRS metrics and clinical indicators. After preprocessing, brain network features and GCN features were extracted from the fNIRS data. Then a modified forward search method was used for SVR model training and feature selection from three feature sets: resting-state/ task-state fNIRS feature sets and clinical feature set. Finally, the SVR model was employed to estimate the FMA scores. The coefficient of determination (R ${}^{{2}}\text {)}$ , root mean square error (RMSE), and mean absolute error (MAE) were utilized to evaluate the models. The GCN-SVR fusion model demonstrated good goodness-of-fit, and exhibited limited variability. The multi-state model demonstrated better performance than both single-state models (P< 0.001). For cortical cases, Aggregate measures over the nine sparsity levels confirmed both high accuracy and stable fitting (R ${}^{{2}}= 0.8397\pm 0.0487$ , RMSE $= 6.75\pm 1.06$ , MAE $= 4.91\pm 0.94$ ), with R2 consistently above 0.76 across sparsity levels. In subcortical patients, the multi-state model achieved a mean R2 of $0.7562\pm 0.0185$ , with RMSE $= 10.51\pm 0.40$ and MAE $= 7.74\pm 0.53$ . The proposed GCN-SVR fusion algorithm based on fNIRS data achieved high accuracy and stable performance in fitting FMA scores, while subset-based sequential forward selection enhances multi-dataset feature selection.