Development of a Single-Camera Markerless Motion Capture System for Assessing Upper Limb Function Post-Stroke (FULMA)

BACKGROUND: Upper limb impairment is common after stroke. Standardised measures such as the Action Research Arm Test and Fugl Meyer Assessment are widely used but may show inter-rater variability and limited sensitivity to subtle motor changes. OBJECTIVE: To develop and evaluate the Functional Upper Limb Movement Assessment (FULMA) system, a single-camera markerless motion capture tool for objective movement analysis of upper limb function with clinician facing outputs after stroke. METHODS: A prospective observational study included 39 stroke participants and 19 healthy controls. Participants completed 9 functional tasks recorded with a LiDAR depth camera. Kinematic features were analysed using machine learning models to predict an ordinal task performance score anchored to Action Research Arm Test scoring descriptors. Predictions were compared with independent occupational therapist ratings, and automated kinematic commentary was generated. RESULTS: The multilayer perceptron achieved a mean accuracy of 83.3%, with task level accuracies of 75.0% to 91.7% and weighted kappa values of 0.67 to 0.88. The support vector machine achieved a mean accuracy of 78.5%, with task level accuracies of 72.2% to 89.0% and weighted kappa values of 0.59 to 0.83. Performance varied across tasks, with the strongest results for cup to mouth and hair brushing. Commentary highlighted prolonged movement time, reduced smoothness and increased trunk displacement. CONCLUSION: This developmental study showed preliminary agreement between automated ordinal score predictions and therapist ratings within a controlled dataset. External validation and evaluation of measurement properties are required before clinical use. Kinematic commentary should be considered exploratory rather than a validated clinical interpretation.