Med-Diet: evaluation of an LLM-based system for clinically guided nutrition care in chronic diseases

Scientifically grounded and clinically applicable dietary management is essential for patients with chronic diseases. However, in routine practice, nutritionists frequently lack efficient and scalable tools to deliver targeted, guideline-consistent nutritional guidance across diverse and complex clinical scenarios.

To develop and conduct an exploratory expert-rating evaluation of Med-Diet, a large language model (LLM)-based agent for generating dietary plans for chronic diseases.

We built Med-Diet using DeepSeek-R1, integrated clinical dietary guidelines, evaluated it on 79 real cases covering common, rare, and complex noncommunicable diseases, and compared it with four general-purpose LLMs (DeepSeek-R1, GPT-4o, GLM-Z1-32B, and Llama-3.3-70B). Fourteen clinical experts from different fields conducted blinded, multidimensional ratings of generated dietary plans. Furthermore, an exploratory comparative experiment assessed nutritionists' efficiency and output quality without and with Med-Diet assistance.

Med-Diet received higher mean preference scores from expert evaluators compared to all baseline LLMs (mean score of 4.09 ± 0.64). Expert ratings suggested superior performance for Med-Diet in dimensions including accuracy, safety, nutritional balance, personalization, practicality, and overall recommendation (all p < 0.05). DeepSeek-R1 ranked second with an overall average score of 3.60 ± 0.68. This model performed the strongest in rare disease scenarios but lagged behind Med-Diet in common diseases and complex cases. GPT-4o (3.33 ± 0.66) and GLM-Z1-32B (3.35 ± 0.75) showed moderate and inconsistent performance, while Llama-3.3-70B performed the worst (3.00 ± 0.69). When nutritionists used Med-Diet to assist in dietary plan generation, the median time required decreased from 17.5 min to 13.0 min (p < 0.05). Expert scores for accuracy, personalization, practicality, and overall recommendation were higher in the Med-Diet-assisted group (adjusted p < 0.05).

In this expert-rating study, Med-Diet-generated dietary plans received higher preference scores from clinical experts compared to those from general-purpose LLMs. These preliminary findings suggest that knowledge injection and framework constraints of Med-Diet may improve expert-perceived quality of AI-generated dietary plans. Med-Diet shows potential as an adjunctive tool in the dietary management of chronic diseases, but its clinical safety and effectiveness require prospective validation.