Machine Learning Approaches for Detecting and Assessing Compensatory Movements in Robot-Assisted Stroke Upper-limb Rehabilitation - PostureCheck

Stroke is one of the leading causes of motor disability worldwide, especially regarding upper limb impairments. As stroke incidence continues to rise and the shortage of physical therapists (PTs) able to assist individuals with disabilities worsens, the burden on healthcare systems is projected to intensify significantly. In this context, robotic technologies for stroke rehabilitation have gained increasing attention as effective tools to enhance patients’ recovery. However, even with robotic assistance, PT supervision remains essential to ensure patients perform exercises safely and correctly. To alleviate the supervision burden, robotic group therapy has emerged as a promising option. This highlights the need for tools that can support therapist supervision during group sessions. One such tool, developed as part of this thesis, is PostureCheck. The goal of this thesis is to develop a novel framework for detecting and classifying compensatory movements in post-stroke upper-limb rehabilitation using a robotic end-effector arm, integrating advanced markerless human pose estimation (HPE) with machine learning. Data were collected from 8 expert therapists and 14 non-specialised subjects. Therapists simulated compensatory movements commonly observed in post-stroke patients, while normative data from all participants established a baseline. The proposed system employs SMPLest-X, a parametric 3D HPE model, to extract high-fidelity skeletal keypoints coordinates, enabling the extraction of biomechanically meaningful features, designed to ensure both biomechanical interpretability and clinical explainability. A hierarchical machine learning model (Random Forest) was developed to perform four classification tasks: (1) distinguishing normative movements from compensatory movements, (2) differentiating single versus multiple compensatory patterns, (3) identifying seven specific compensatory movement types, and possibly (4) grading severity levels within each movement type. Two validation strategies (Leave-One-Subject-Out (LOSO) cross-validation and LOSO by target) and deep hyperparameter tuning were implemented to ensure robustness and generalizability. It was achieved over 85% accuracy in differentiating compensatory movements from normative ones and over 77% in more complex and challenging scenarios (single vs multiple compensatory movements and in classifying individual compensatory movement). To address occlusion challenges, a hybrid setup integrating SMPLest-X, Xsens inertial sensors, and a Vicon motion capture system was designed and is presented in this manuscript. A preliminary data collection was carried out to demonstrate the setup’s feasibility, while comprehensive validation and analysis are planned for future work. The classifier is being preliminarily applied to real patient video data to automatically identify compensatory movements frame by frame, overlaying the results onto the video to assist therapists during manual labelling. In conclusion, this work establishes an innovative, robust, and reliable methodology for enhancing post-stroke rehabilitation assessment (automatic detection of compensatory movements), offering promising perspectives for real-time clinical integration and broader adoption by therapists and clinicians.