Usability of Large Language Models for Developing Exercise‐Based Rehabilitation Programs in Patients With Stable Angina

Background While artificial intelligence (AI) and large language models (LLMs) are increasingly used in exercise‐based cardiac rehabilitation (CR), the clinical reliability and adherence to guidelines of AI‐generated prescriptions require systematic comparison with expert‐designed protocols. Objective This study aimed to evaluate the heterogeneity between therapeutic exercise‐based rehabilitation programs generated by LLMs (ChatGPT and Gemini Advanced) and those developed by healthcare professionals for individuals with stable angina and chronic coronary syndrome (CCS). Methods Exercise‐based rehabilitation prescriptions generated by ChatGPT and Gemini Advanced were compared with those developed by an expert panel across key clinical domains, including dietary recommendations, resistance‐training prescription method, lower‐extremity strengthening decisions, and aerobic exercise intensity selection. Results Significant heterogeneity was identified between AI‐generated and expert panel‐prescribed rehabilitation programs across all evaluated domains. Statistically significant differences were observed for dietary recommendations ( χ 2 = 80.00, p < 0.001), resistance‐training prescription method ( χ 2 = 20.16, p < 0.001), lower‐extremity strengthening decisions ( χ 2 = 54.07, p < 0.001), and aerobic exercise intensity selection ( χ 2 = 49.24, p < 0.001). Agreement analyses demonstrated poor‐to‐absent concordance for most clinically relevant decisions. Fleiss’ kappa values indicated significant disagreement for dietary recommendations ( κ = −0.500, p < 0.001), flexibility training ( κ = −0.500, p < 0.001), respiratory muscle training (RMT) ( κ = −0.500, p < 0.001), and rating of perceived exertion (RPE) based safety warnings ( κ = −0.364, p = 0.001). Pairwise analyses showed perfect agreement only for selected isolated parameters (diet recommendations between the healthcare Professional and ChatGPT; flexibility training and RMT between the expert panel and Gemini Advanced), whereas no systematic agreement was observed for lower‐extremity strengthening, depression alerts, or RPE warnings. Overall, LLMs generated structurally consistent outputs in some domains but demonstrated substantial variability in exercise prescription and safety‐related clinical decisions. Conclusion Current LLMs may serve as supportive tools for structuring rehabilitation recommendations; however, their outputs showed limited agreement with expert‐designed programs and inconsistent adherence across key components of comprehensive CR. Given the marked variability in exercise prescription, safety alerts, and guideline‐concordant recommendations, LLM‐generated programs should be interpreted within a human‐in‐the‐loop framework and not used as autonomous prescribing systems. Further development, guideline integration, and prospective clinical validation are required before routine implementation in CR practice.