Monitoring functional engagement of upper-limb prostheses using embedded sensors and machine learning in real-world contexts

Abstract Accurately monitoring upper-limb prosthesis use in daily life remains a major challenge in neurorehabilitation, as existing tools primarily measure wear duration without capturing functional engagement or task relevance. To address this gap, this work develops a multimodal sensing transradial prosthesis simulator equipped with embedded pressure, strain, and temperature sensors, enabling simultaneous tracking of mechanical and physiological interactions during natural use. Using data collected from able-bodied participants performing both structured tasks and unstructured daily activities, we train a random forest–based machine learning model to classify three distinct behavioral states: goal-directed active use, non-task use, and passive wear. The classifier trained on six multimodal features achieves 95% precision in distinguishing these states, with pressure-based sensors contributing the most discriminative information. Furthermore, a simplified variability-only model, using rolling standard deviation features from analog signals, accurately differentiates ‘Use’ vs. ‘Rest’ states with a balanced accuracy of 0.83 and an AUC of 0.913, demonstrating robust performance independent of absolute signal calibration. These findings highlight the feasibility of lightweight, interpretable, and scalable prosthesis monitoring systems that operate continuously in real-world settings, providing objective behavioral feedback for personalized rehabilitation and long-term prosthesis adaptation.