Rehabilitation Engagement Prediction in Pediatric Patients with Neuromotor Disorders: An Adaptive Hierarchical Bayesian Approach

Patient engagement is essential for improving clinical outcomes in pediatric neuromotor rehabilitation. Physiological computing offers a quantitative aid for its assessment, but clinical implementation is limited by intrinsic noise in therapist-reported labels and high inter-subject physiological variability. The study aims to predict absolute Engagement State and relative Engagement Dynamics in 31 underage participants (21 with neuromotor disorders, 10 neurotypical). Multimodal data (heart rate variability, electrodermal activity, and acceleration) and PRIME-SP engagement labels were collected during technology-assisted rehabilitation. Following feature extraction, confident learning reduced noise in therapist annotations, and hierarchical Bayesian models were trained and tested under cold start and adaptive learning evaluation paradigms. Results proved the non-destructive effectiveness of adaptive learning, achieving F1-Scores of 0.75 and 0.79 for Engagement State and Dynamics, respectively. The study marks a first step towards the development of a decision support system for real-time engagement assessment.