Feasibility of an AI-based system to estimate upper extremity motor function in stroke patients from white matter mean diffusivity features

Abstract

Introduction: Upper extremity (UE) impairment after stroke substantially affects patients’ quality of life. Standardized clinical evaluations such as the Fugl-Meyer Assessment for the Upper Extremity (FMA-UE) and the Action Research Arm Test (ARAT) are routinely used to assess UE motor function. However, these evaluations may be influenced by subjective factors and inter-rater variability. Mean diffusivity (MD), derived from diffusion tensor imaging (DTI), reflects white matter microstructural alterations and could enable artificial intelligence (AI)-based estimation of standardized UE motor scores after stroke from white matter biomarkers, providing an objective complement to clinical assessments.

Objective: To evaluate the feasibility of an AI algorithm to estimate UE motor function of stroke patients using white matter MD values and identify the most relevant white matter regions contributing to the estimations.

Methods: A private DTI database comprising 85 observations from 30 stroke patients was used. Nineteen patients had 4 observations each during a clinical trial of a brain-computer interface (BCI)-based UE motor rehabilitation therapy, whereas 11 patients contributed one observation during a feasibility BCI therapy clinical trial. FMA-UE and ARAT scores were obtained at each observation. Data collection and validation were conducted at the Instituto Nacional de Rehabilitación LGII (INR-LGII) and involved stroke patients, rehabilitation clinicians, and researchers. All patients provided informed consent approved by the INR-LGII Research and Ethics Committees. MD values were extracted using FSL (v6.0.4) from homologous white matter regions of interest (ROIs) in the ipsilesional and contralesional hemispheres defined in the JHU ICBM-DTI-81 White Matter Atlas. For each ROI pair, the ratio of ipsilesional to contralesional MD (rMD) was calculated. These ratios were used to estimate, separately, FMA-UE and ARAT scores with regression models computed with the Feature-ranked Self-growing Forest (FSF) algorithm. Model performance was estimated using a leave-one-subject-out cross-validation, repeated 30 times to address FSF variability. After confirming non-Gaussian distributions, linear associations between estimated and actual scores were calculated using Spearman correlations, whereas Wilcoxon signed-rank tests were used to assess differences between actual and estimated scores. Feature relevance was assessed with FSF to determine each ROI pair’s contribution. All computations and statistical analyses, with a 95% confidence level, were performed in MATLAB R2025b.

Results: The estimated FMA-UE scores achieved a significant but weak correlation with the actual scores (r=0.22, p<0.05), and no statistically significant differences were observed (p=0.12). Estimated ARAT scores were not significantly correlated with the actual scores (r=0.09, p=0.40), and no statistically significant differences were observed (p=0.08). For FMA-UE estimation, the superior longitudinal fasciculus and superior corona radiata (SCR) were the most relevant ROIs, whereas the uncinate fasciculus and SCR were most relevant for ARAT estimation.

Conclusions: White matter rMD values showed limited ability to estimate UE motor function in stroke patients, possibly because MD alterations are not exclusively related to motor pathways. While the sample size was limited, the relevance of motor-related white matter regions suggests the potential use of brain structural biomarkers for objective UE motor function estimation. Future developments could support implementation as a complementary tool in rehabilitation assessments after integration into existing imaging workflows.