Comparative Performance of Four LLMs in Generating Exercise Prescriptions

Four large language models were evaluated on their ability to generate personalized exercise prescriptions using the FITT-VP framework. Claude 3.7 achieved the highest total score of 50.23 out of 60, while DeepSeek R1 scored the lowest at 40.30. The dataset, authored by Huan Feng and last updated in May 2026, contains the study results and analysis in a 374.6 KB document.

Use Cases

• Benchmarking LLM performance in clinical text generation based on the FITT-VP framework
• Analyzing model strengths and weaknesses in handling specific prescription dimensions like Progression and Intensity
• Supporting research into human-AI collaborative frameworks for personalized exercise medicine

Strengths

• Scores for four specific LLMs (GPT-4o, Claude 3.7, DeepSeek R1, Grok-3) are provided with statistical significance (p < 0.001, η² = 0.896)
• Prescriptions were evaluated by three certified exercise specialists using a validated 0–10 scale across six dimensions
• The study is based on 30 synthetic patient profiles designed from epidemiological data and clinical guidelines

Limitations

• The data is contained in a single 374.6 KB DOCX file, which is a tiny dataset with limited scope
• Column-level documentation is absent; field semantics must be inferred after download
• Results are based on a single-run evaluation of static synthetic profiles rather than real-world clinical outcomes

Provenance

Source

figshare

Collection Method

Generated by evaluating four LLMs on 30 synthetic patient profiles using the FITT-VP framework.

Freshness

Last updated 2026-05-25 04:34:41