Article image

Using articial intelligence for exercise prescription in personalised health promotion: A critical evaluation of OpenAI’s GPT-4 model

Using articial intelligence for exercise prescription in personalised health promotion: A critical evaluation of OpenAI’s GPT-4 model


Ismail Dergaa , Helmi Ben Saad , Abdelfatteh El Omri , Jordan M. Glenn

Cain C. T. Clark , Jad Adrian Washif , Noomen Guelmami , Omar Hammouda

Ramzi A. Al-Horani , Luis Felipe Reynoso-Sánchez , Mohamed Romdhani , Laisa Liane

Paineiras-Domingos 16, Rodrigo L. Vancini 17, Morteza Taheri 18, Leonardo Jose Mataruna-Dos-

Santos 19, Khaled Trabelsi , Hamdi Chtourou , Makram Zghibi , Özgür Eken , Sarya Swed,

Mohamed Ben Aissa , Hossam H. Shawki , Hesham R. El-Seedi , Iñigo Mujika 

Stephen Seiler , Piotr Zmijewski , David B. Pyne , Beat Knechtle , Irfan M Asif , Jonathan

Drezner , Øyvind Sandbakk , Karim Chamari


ABSTRACT: The rise of articial intelligence (AI) applications in healthcare provides new possibilities for personalized health management. AI-based tness applications are becoming more common, facilitating the opportunity for individualised exercise prescription. However, the use of AI carries the risk of inadequate expert supervision, and the efcacy and validity of such applications have not been thoroughly investigated, particularly in the context of diverse health conditions. The aim of the study was to critically assess the efcacy of exercise prescriptions generated by OpenAI’s Generative Pre-Trained Transformer 4 (GPT-4) model for ve example patient proles with diverse health conditions and tness goals. Our focus was to assess the model’s ability to generate exercise prescriptions based on asingular, initial interaction, akin to atypical user experience. The evaluation was conducted by leading experts in the eld of exercise prescription. Five distinct scenarios were formulated, each representing ahypothetical individual with aspecic health condition and tness objective. Upon receiving details of each individual, the GPT-4 model was tasked with generating a30-day exercise program. These AI-derived exercise programs were subsequently subjected to athorough evaluation by experts in exercise prescription. The evaluation encompassed adherence to established principles of frequency, intensity, time, and exercise type; integration of perceived exertion levels; consideration for medication intake and the respective medical condition; and the extent of program individualization tailored to each hypothetical prole. The AI model could create general safety-conscious exercise programs for various scenarios. However, the AI-generated exercise prescriptions lacked precision in addressing individual health conditions and goals, often prioritizing excessive safety over the effectiveness of training. The AI-based approach aimed to ensure patient improvement through gradual increases in training load and intensity, but the model’s potential to ne-tune its recommendations through ongoing interaction was not fully satisfying. AI technologies, in their current state, can serve as supplemental tools in exercise prescription, particularly in enhancing accessibility for individuals unable to access, of ten costly, professional advice. However, AI technologies are not yet recommended as asubstitute for personalized, progressive, and health condition- specic prescriptions provided by healthcare and tness professionals. Further research is needed to explore more interactive use of AI models and integration of real-time physiological feedback