The Role of Artificial Intelligence and Professional Expertise in Adapted Physical Activity Prescription for Orthopedic Rehabilitation

TL;DR

Expert professionals create higher-quality physical activity plans for orthopedic rehabilitation compared to AI alone or novices using AI, showing that AI is a helpful tool but not a replacement for expertise.

Contribution

This study demonstrates that AI can partially bridge the expertise gap in physical activity prescription when used by novices, but cannot replace expert judgment.

Findings

Expert protocols scored highest in quality, followed by novice professionals using AI, with AI alone scoring lowest.

AI support improved protocol quality for novices in most conditions but failed to match expert performance in safety and appropriateness.

AI-generated protocols showed greater variability across different orthopedic conditions.

Abstract

Background: Adapted Physical Activity (APA) prescription is a complex decision-making process that integrates clinical guidelines and individual patient characteristics and remains strongly dependent on clinician experience. Generative artificial intelligence (AI) has recently emerged as a potential decision-support tool in exercise prescription; however, its interaction with professional expertise is still unclear. This study compared the perceived quality of APA protocols developed by expert professionals, novice professionals supported by AI, and AI operating autonomously across multiple orthopedic conditions. Methods: In this observational cross-sectional study, five real orthopedic prescriptions (scoliosis, low back pain, osteoporosis, high risk of falls, and osteoarthritis) were used to generate three APA protocols per condition: expert professional (EP), novice professional with AI support (NAI), and AI alone. All protocols were created using an identical standardized prompt and anonymized. A multidisciplinary panel of 135 professionals blindly evaluated the protocols using a structured questionnaire assessing effectiveness, safety, appropriateness, clarity, and progression. Overall quality scores were compared using Friedman tests with post hoc Wilcoxon signed-rank tests. Results: Across all conditions, EP protocols achieved the highest quality scores, followed by NAI, while AI-alone protocols consistently received the lowest ratings (all p < 0.05). NAI protocols showed intermediate performance, partially reducing the expertise gap. Post hoc analyses showed that EP protocols received significantly higher rating than AI protocols in all conditions (p < 0.01). NAI protocols received significantly higher rating than AI protocols in most conditions (p < 0.01) except osteoporosis (p = 0.362). Differences between EP and AI were most pronounced for safety (p < 0.01), appropriateness (tailoring p < 0.01), and progression (p < 0.05), whereas EP–NAI differences were smaller and condition-dependent. AI-alone protocols showed greater variability across pathologies. Conclusions: Professional expertise remains the main determinant of APA protocol quality. AI support can improve protocol structure and perceived quality when used by novice professionals but does not replace expert clinical reasoning. AI-generated protocols without human oversight are not yet suitable for autonomous APA prescription, supporting a complementary, expertise-dependent role of AI in exercise programming.

Figures