Background/Aims: Upper extremity (UE) impairments post-stroke significantly affects quality of life, necessitating accurate and consistent assessment tools. Current methods often rely on subjective interpretation. The FULMA system integrates machine learning (ML) to enhance assessment accuracy and objectivity. The aims of this study were to develop and evaluate a laptop-based optical motion capture system that applies ML models to classify functional UE movements in stroke patients.
Methods: Motion data were collected from 39 stroke patients and 21 age- and sex-matched healthy controls performing nine functional UE tasks (M1–M9) using everyday objects. Four ML models; Multilayer Perceptron (MLP), Support Vector Machine (SVM), k-Nearest Neighbour (KNN), and Random Forest (RF) were developed to predict Action Research Arm Test (ARAT) scores. Model performance was benchmarked against scores provided by two experienced occupational therapists.
Results: The MLP model outperformed all others, achieving movement classification accuracy between 73.1%-90.8%, and a mean accuracy of 81.3%. SVM models showed slightly lower accuracy (70.4%–89.0%; mean: 77.8%), followed by RF (mean: 73.7%) and KNN (mean: 70.7%). Movements with higher kinematic complexity, such as M5 (reach, grasp, transport, release), consistently yielded better classification accuracy. Cohen’s Kappa values for MLP ranged from 0.61-0.85 across eight of nine movements, indicating substantial to strong agreement with therapist scores. Additionally, the FULMA system generates automated, patient-specific commentary by identifying deviations in kinematic performance relative to age- and sex-matched healthy controls.
Conclusion: FULMA demonstrates strong potential as a scalable tool for objective UE assessment post-stroke, supporting clinical interpretation and laying groundwork for broader rehabilitation applications.