P751 - COMPARISON OF CHATGPT-3.5, CHATGPT-4O AND DEEPSEEK IN PERSONALIZED DIETARY PRESCRIPTION: INSIGHTS FROM A CLINICAL DIETITIAN

P751

COMPARISON OF CHATGPT-3.5, CHATGPT-4O AND DEEPSEEK IN PERSONALIZED DIETARY PRESCRIPTION: INSIGHTS FROM A CLINICAL DIETITIAN

Q. You^1,*, X. Li¹, Y. Ma¹, L. Shi¹, Z. Rao¹, W. Hu¹

¹Department of Clinical Nutrition , West China Hospital Sichuan University, Chengdu, China

Rationale: Our previous study revealed the incompetency of ChatGPT-3.5 in generating nutritionally accurate dietary plans for patients with CKD¹. Given the subsequent release of ChatGPT-4o and the emergence of the DeepSeek, further evaluation of their capabilities in prescribing appropriate dietary plans is warranted.

Methods: Standardized dietary prescription prompts were designed with a fixed energy target (1800kcal) and five protein levels (30~70g), representing diets for different stages of CKD. Twenty participants were recruited and instructed to input these pre-established prompts into ChatGPT-3.5, ChatGPT-4o, and DeepSeek to generate dietary plans. The actual energy and protein content of the output meal plans were recorded and compared.

Results: A total of 300 meal plans were generated, with 100 allocated to each LLM. The two ChatGPT models consistently generated diets with higher protein levels than requested (ps<0.05), while DeepSeek adhered most closely to the target values. Regarding energy content, none of the LLMs reached the 1800kcal target, with ChatGPT-3.5 providing the lowest-calorie plans (1281.7±187.9kcal) and DeepSeek the highest (1487.1±199.8kcal). ChatGPT-3.5 exhibited a high degree of repetition, while the other two LLMs demonstrated minimal.

Conclusion: None of the invested LLMs demonstrated the capability to prescribe nutritionally accurate diets, with either deficient energy or inappropriate protein content. DeepSeek outperformed both ChatGPT models in achieving the target energy and protein levels. Patients with strict protein and other particular nutrient restrictions are not recommended to rely on the dietary plans prescribed from the investigated LLMs to avoid potential health risks.

References: 1. You Q, Li X, Shi L, et al. Still a Long Way to Go, the Potential of ChatGPT in Personalized Dietary Prescription, From a Perspective of a Clinical Dietitian. J Ren Nutr 2025.

Disclosure of Interest: None declared