Abstract: FR-PO0450
From Prompt to Plate: Can ChatGPT Plan a Safe and Clinically Appropriate Diet for Patients on Hemodialysis?
Session Information
- Dialysis: Measuring and Managing Symptoms and Syndromes
November 07, 2025 | Location: Exhibit Hall, Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Dialysis
- 801 Dialysis: Hemodialysis and Frequent Dialysis
Authors
- Wang, Lin-Chun, Fresenius Medical Care, Clinical Research, New York, New York, United States
- Yueh, Sheng-Han, Renal Research Institute, New York, New York, United States
- Rao, Sharon, Fresenius Medical Care, Clinical Research, New York, New York, United States
- Zhang, Hanjie, Renal Research Institute, New York, New York, United States
- Nandorine Ban, Andrea, Renal Research Institute, New York, New York, United States
- Bergling, Karin, Renal Research Institute, New York, New York, United States
- Ginsberg, Nancy, Fresenius Medical Care Holdings Inc, Waltham, Massachusetts, United States
- Moore, Linda W., Houston Methodist Hospital, Houston, Texas, United States
- Biruete, Annabel, Purdue University, West Lafayette, Indiana, United States
- Kotanko, Peter, Renal Research Institute, New York, New York, United States
- van der Sande, Frank, Maastricht Universitair Medisch Centrum+, Maastricht, LI, Netherlands
- Kooman, Jeroen, Maastricht Universitair Medisch Centrum+, Maastricht, LI, Netherlands
Background
Diverse food needs and preferences among hemodialysis patients are challenging for renal dietitians but large language models (LLMs) could support renal dietitians and patients (Berling, CKJ 2025). We evaluated the real-world feasibility of LLM-generated meal plans by preparing recommended dishes and assessing their clinical appropriateness and nutritional accuracy.
Methods
We created 20 virtual patients using Monte Carlo simulation and selected three with Western, Chinese, and Mexican preferences. Their demographics, clinical characteristics, labs, and food budgets were provided to ChatGPT (model GPT-4o), which generated three recipes per patient. The resulting nine dishes were prepared in a professional kitchen, and the nutritional content was analyzed using ESHA software (ESHA research, Salem, OR, USA), including USDA Foods Database. Separately, three experienced renal dietitians created daily nutrition recommendations for the same three selected patients. We further compared ChatGPT’s recommendations with dietitians’ recommendations.
Results
ChatGPT underestimated energy and macronutrient intake. Average daily recommendations from ChatGPT were as follows: Energy (Fig. 1a): 1,077 kcal (vs. 2,339 kcal from dietitians and 1,494 kcal from ESHA software); Protein (1b): 67 g (vs. 95 g and 87 g); Total carbohydrate (1c): 102 g (vs. 311 g and 121 g); Total fat (1d): 46 g (vs. 80 g and 75 g); Potassium (1e): 1,917 mg (vs. 2,694 mg and 2,779 mg); Sodium (1f): 1,147 mg (vs. 2,300 mg and 2,285 mg). ChatGPT produced low-calorie (10–19 kcal/kg/d) and low-protein diets (0.65–0.77 g/kg/d), potentially appropriate for weight loss but concerning for nutrient deficiency. Dietitians’ plans aligned with guidelines (22–30 kcal/kg/day, 1 g protein/kg/d).
Conclusion
LLMs like ChatGPT show promise for dialysis nutrition planning but require refinement. Clinical oversight remains essential to ensure patient safety and guideline alignment.