Assessment of large language model chatbots for hemodialysis meal planning: a descriptive study

Abstract

Large language models (LLMs) have the potential to improve nutritional counseling for patients with end stage kidney disease (ESKD). This study evaluates the utility of publicly available LLMs in generating meal plans for individuals receiving hemodialysis. Fifty hypothetical patient profiles were generated from United States national data and used to prompt four LLM chatbots, ChatGPT-o3-mini ® (OpenAI), Claude Sonnet 3.7 ® (Anthropic), Gemini 2.5 ® (Google), and Llama 3.1 ® (Meta), to create a single day meal plan accounting for ESKD dietary constraints. The primary outcome was concordance of LLM-generated meal plans with reference nutrition databases and specified nutrition goals. A secondary outcome was a qualitative usability assessment, as judged by three independent reviewers. All models demonstrated substantial limitations in accurately representing nutrient content, particularly in underestimating phosphorus and potassium content in foods. Quantitatively, ChatGPT achieved the highest performance of the models studied with the highest average concordance with gold standard nutrition databases and with the lowest nutrient deviations from prompt goals. Gemini and Llama had less qualitative errors than ChatGPT and Claude, but all models frequently had vague outputs, often recommending composite foods without clear nutrient values. Currently, publicly available LLMs do not readily generate clinically acceptable meal plans for hemodialysis patients. All models misrepresented nutrient content and had significant usability concerns. Improvements in model architecture, knowledge bases, and domain-specific optimization will be required before LLMs can be safely used for dietary counseling for patients with ESKD.

Ähnliche Arbeiten