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

Frontiers in rehabilitation sciences · 2026 · Other · PT
Luca Perrero, Menada Gardalini, Tatiana Bolgeo, Patrizia Valorio, Calogero Malfitano, Valter De Michelis

An AI model wrote exercise plans for five made-up cardiac patients that two experts rated as safe and guideline-consistent. No real patients were involved and no comparison to clinician-written plans was made. This is a proof-of-concept, not evidence the AI can prescribe exercise.

Key findings

1

The AI generated a structured 30-day exercise program for every one of the five hypothetical cardiac scenarios, and two experienced cardiac rehabilitation specialists found no manifestly unsafe recommendations in any of them.

2

Expert mean ratings on a 5-point scale were 4.5 for guideline adherence, 5.0 for safety, and 4.3 for clinical plausibility. Safety was the highest-rated domain (range 4.8–5.0, with four of five scenarios scoring the maximum), while clinical plausibility showed the most variability (range 4.1–4.4).

3

The study involved no real patients, used only five hypothetical profiles, relied on subjective Likert-scale ratings by two reviewers, and did not compare the AI-generated plans to prescriptions written by experienced cardiac rehabilitation clinicians.

MethodologyResultsClinical Relevance

Single-blind randomized controlled trial. Sixty-four participants were randomized across two intervention arms and assessed at twelve weeks for the primary outcome…

Sign in to read the deep summary

Create a free account to read the methodology, results and clinical relevance — three full summaries a month, no card.

Summarised by AI from the abstract, not the full paper. Check it against the source before it changes what you do. Read it on PubMed