AI-Generated Diet and Exercise Recommendations for Cardiovascular Health Compared to Established Cardiology Society Guidelines

Comprehensive Summary

Nduka et al. evaluated whether four large language models – ChatGPT (GPT-4), Claude AI (Claude 2), DeepSeek AI (DeepSeek-LLM 67B), and Google Gemini (Pro) – produce cardiovascular diet and exercise recommendations consistent with American Heart Association/American College of Cardiology (AHA/ACC) and European Society of Cardiology (ESC) guidelines. Each model was asked 15 standardized questions (five exercise-related and ten diet-related), and responses were graded by a primary care physician and a cardiology fellow with cardiologist adjudication. ChatGPT, Claude, and DeepSeek had 90% of diet responses rated appropriate, while Google Gemini had 80%; all exercise responses across models were appropriate. Most answers reflected guideline targets such as ≥150 minutes/week of moderate or ≥75 minutes/week of vigorous exercise, but dietary responses often lacked precise thresholds, particularly for carbohydrate and added sugar intake (e.g., <10% of total daily energy intake). Some responses showed slight alignment toward AHA/ACC over ESC guidance. The authors conclude that LLMs can provide accessible cardiovascular health information but should only complement expert medical advice, with future systems focusing on improving quantitative specificity and guideline balance.

Outcomes and Implications

The study suggests that current LLMs can reliably provide high-level cardiovascular prevention advice, particularly for exercise recommendations that closely match guideline targets. However, their tendency to omit precise quantitative dietary thresholds limits their usefulness for detailed clinical counseling with patients. In real-world practice, these tools may function best as patient education or preliminary information sources, helping users understand basic lifestyle prevention concepts before seeking guidance from clinicians. Because the study evaluated only four models, fifteen questions, and a small reviewer panel, and used fixed model versions that evolve rapidly, broader validation will be needed before AI-generated lifestyle recommendations can be considered reliable components of preventive cardiology workflows.