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

No real patients were involved; all five profiles were hypothetical and constructed by the study team, so the findings say nothing about how the AI performs with actual clinical data, comorbidities, or patient-specific context. Only five scenarios were tested, covering common but not the full range of cardiac rehabilitation referrals; the authors themselves note the profiles do not encompass the full spectrum of clinical complexity seen in routine practice. There was no comparison group: the AI-generated plans were never compared side-by-side with prescriptions written by experienced cardiac rehabilitation clinicians, so it is unknown whether the AI matches, exceeds, or falls short of standard clinical practice. The evaluation relied on two reviewers' subjective Likert-scale ratings; no objective or standardised assessment framework was used, and no patient outcomes (adherence, functional capacity, adverse events) were measured. Only a single AI model (DeepSeek-V3) was tested, so the findings cannot be generalised to other large language models. The AI worked from static, one-time prompts and could not incorporate real-time physiological data, longitudinal training response, or dynamic clinical changes, which are central to ongoing exercise prescription. The authors acknowledge that large language models remain susceptible to hallucinations — generating inaccurate or fabricated information presented with confidence — which in a cardiac context could lead to inappropriate exercise recommendations if accepted without clinical review.

The authors declared that no financial support was received for the work or its publication. No industry sponsorship, study design involvement, or manuscript-writing contribution by a commercial entity was reported.

Do not read the 4.5-to-5.0 expert ratings as evidence that an AI can safely prescribe cardiac rehabilitation exercise in place of a clinician. The study used five hypothetical patient descriptions, no real patients, no comparison to clinician-written plans, and subjective Likert-scale scoring by two reviewers. The authors themselves state that no conclusions can be drawn about clinical effectiveness, patient adherence, or long-term outcomes, and that clinician oversight remains essential for any potential use of these tools in practice.