Ai-driven lightweight motor assessment embedded in an exergame for post-stroke rehabilitation

Abstract

Introduction: Stroke is a leading cause of long-term disability worldwide and frequently results in persistent upper-limb motor impairments. Accurate motor assessment is essential for rehabilitation planning and monitoring recovery. This assessment is traditionally performed using standardized clinical scales, such as the Fugl-Meyer Assessment (FMA), which require trained healthcare professionals, can be time- and resource-intensive, and involve subjectivity, as scores may vary across evaluators. Moreover, they cannot be easily administered remotely, and repeated in-person evaluations can be difficult for individuals with mobility limitations. Faster and more objective assessment approaches could enable more frequent evaluations and improve the tracking of motor recovery. Although recent advances in digital health, exergames, and Artificial Intelligence (AI) have enabled alternative solutions, many existing systems rely on specialized motion capture sensors, time-consuming software, or non-interpretable AI models, limiting accessibility, scalability, and clinical adoption. Objectives: This thesis aims to develop and evaluate a sensor-free, lightweight, and interpretable approach for post-stroke upper-limb motor assessment embedded into a rehabilitation exergame. The proposed system simultaneously provides therapeutic exercises and automatically estimates upper-limb motor performance during gameplay using only a standard camera, enabling supervised remote use. Specifically, the objectives are to (i) review existing automated motor assessment methods and their relationships with kinematic game parameters, and (ii) propose and preliminarily validate a low-cost exergame framework capable of estimating clinical motor impairment using simple and transparent kinematic features. Materials and Methods: A systematic review was conducted to analyze current technologies and computational approaches for automated upper-limb motor assessment after stroke, identifying the most used game parameters. Based on these findings, an experimental study was developed using a Unity-based rehabilitation exergame controlled by arm movements captured with a standard camera and processed using the MediaPipe framework. Sixteen kinematic and spatiotemporal features were extracted from two-dimensional hand and arm trajectories during gameplay. Twelve individuals with chronic stroke (24 upper limbs) participated, with bilateral FMA scores used as the clinical reference. Correlation analyses, exhaustive feature selection, and multiple linear regression modeling were performed to estimate FMA scores, with exploratory comparisons to alternative machine learning models. Results: The systematic review indicated that most existing approaches depend on external sensors or computationally complex methods, often limiting interpretability and real-world applicability. In the experimental study, some gameplay-derived and clinically interpretable features, including average hand aperture, and spatial exploration area, were significantly correlated with FMA scores. A lightweight multiple linear regression model demonstrated strong predictive performance for affected limbs (Spearman ρ = 0.92, R² = 0.89, RMSE = 4.42) and accurately stratified motor impairment severity, achieving accuracies between 86% and 93%. More complex machine learning models did not outperform the interpretable regression approach. Conclusion: This thesis demonstrates that sensor-free, low-cost, interpretable and lightweight computation models for post-stroke motor assessment can be embedded into rehabilitation exergames using kinematic features derived from a standard camera, and be potentially valid for motor function assessment. Further tests with more games and populations are needed to generalize the result. Integrating assessment into gameplay reduces clinical workload, enables high-frequency monitoring, and improves accessibility for telerehabilitation.