Machine Learning-Driven Vision System for Detecting and Evaluating Compensatory Movements in Robot-Based Stroke Rehabilitation

A stroke, also known as a cerebrovascular accident (CVA), occurs when blood flow to a part of the brain is interrupted or when a blood vessel in the brain bursts. Stroke remains the second-leading cause of death and the third-leading cause of death and disability combined in the world. Only 12% of stroke survivors achieve complete upper limb functional recovery. Robotic rehabilitation has become a pivotal tool in improving the recovery process for stroke survivors, offering targeted support to enhance motor function. However, many patients develop compensatory movements—unintended patterns that allow them to work around mobility limitations but reinforce incorrect motor behaviors. This thesis aims to develop a machine learning framework to detect and evaluate compensatory movements in stroke patients using 3D skeletal data extracted from video frames. By identifying and correcting compensatory movements, our approach promotes more personalized and effective rehabilitation, potentially easing the burden on therapists and healthcare facilities with a scalable solution. A previous study was conducted using 2D reconstruction through OpenPose, in this work we advanced the analysis by exploiting the third dimension. We collected data from 22 subjects, including 8 therapists and 14 non-specialized subjects. Therapists were tasked with simulating compensatory movements, similar to those exhibited by patients, while interacting with a robotic end-effector arm, which was used to assist in performing the movements. Normative data were gathered by considering all the subjects. By extracting biomechanically relevant features from this data, and selecting them, we developed hierarchical models that initially identify compensatory movements and subsequently classify them into specific categories. Three scenarios were considered: (1) differentiation between Normative and Compensatory Movements, (2) Single Compensatory Movements vs Multiple Compensatory Movements, and (3) distinction between seven Single Compensatory Movements (e.g. Trunk Extension, Trunk Flexion, Trunk Rotation, Shoulder Elevation). As first approach, we used Random Forest with K Fold Cross-Validation and Leave One Subject Out (LOSO) methods to evaluate the models' performance. Our goal was to develop robust models that could accurately detect and classify compensatory movements, for this reason we conclude our study by considering a further approach with the introduction of an uncertain class. Our models delivered promising results. Considering the F1 score metric for the LOSO per target study, we reached: (1) a value of 0.83 for the classification of Normative vs. Compensatory Movements, (2) a score of 0.74 for the classification of Single vs. Multiple Compensatory Movements, and (3) a value of 0.79 for the classification of Single Compensatory Movements. This thesis presents an innovative approach to automatically detect compensatory movements using 3D skeletal reconstruction and machine learning algorithms. It provides a framework for improving compensatory movement detection in rehabilitation scenarios and could be extended for real-time analysis to support automated therapeutic assistance.