Evaluation of AI-generated exercise prescriptions for diverse cardiac conditions in rehabilitation: a simulation study using the DeepSeek model

Background: Cardiac rehabilitation (CR) significantly improves outcomes for patients with cardiovascular disease, but delivering consistently personalized exercise prescriptions remains challenging due to resource limitations and practice variability. Artificial intelligence (AI), particularly large language models (LLMs), may support the generation of guideline-adherent baseline exercise programs. Objective: This study aimed to assess the adherence to guidelines, expert-evaluated safety, and clinical plausibility of AI-generated exercise prescriptions for cardiac rehabilitation across common cardiac conditions in simulated clinical scenarios. Methods: Exercise prescriptions generated by DeepSeek were evaluated using five purpose-designed simulated clinical profiles: recent myocardial infarction (<1 month), prior myocardial infarction (>1 month), atrial fibrillation, heart failure with reduced ejection fraction (EF 35%), and heart failure with preserved ejection fraction (EF 50%). For each profile, the model was prompted to generate a 30-day CR exercise program in accordance with FITT-VP principles (Frequency, Intensity, Time, Type, Volume, Progression), standard safety considerations, and current American Heart Association/European Society of Cardiology (AHA/ESC) guideline recommendations. Two experienced CR specialists independently evaluated the generated programs for adherence to guidelines, safety, and clinical plausibility. Results: The model successfully generated unique 30-day exercise programs for all five profiles, incorporating warm-up and cool-down phases, aerobic and resistance training components, intensity guidance using rating of perceived exertion and/or heart rate targets, and progressive adaptation over time. The programs demonstrated appropriate condition-specific modifications, such as lower-intensity exercise for recent myocardial infarction and heart failure with reduced ejection fraction, and avoidance of high-intensity exercise bursts in atrial fibrillation. Expert evaluation confirmed broad adherence to the FITT-VP principles and the major safety recommendations of the current AHA/ESC guidelines. No overtly unsafe prescriptions were identified. Conclusion: DeepSeek was able to generate structured, guideline-informed exercise prescriptions across various simulated cardiac rehabilitation scenarios. These results suggest the model's capacity to generate guideline-informed exercise prescriptions aligned with current guideline principles in simulated settings, but they do not prove clinical effectiveness, safety, or suitability for routine practice. Additional studies involving real patients and direct comparisons with clinician-made exercise prescriptions are needed before clinical use can be considered.