Kidney Week

Abstract: FR-PO0055

Longitudinal Benchmarking of GPT-4o and GPT-5.4 for Nutrient Estimation in Culturally Diverse Hemodialysis Meal Plans

Session Information

• Artificial Intelligence: Clinical Prediction, Decision Support, and Digital Kidney Care
  October 23, 2026 | Location: Exhibit Hall A, Convention Center
  Abstract Time: 10:00 AM - 12:00 PM

Category: Artificial Intelligence, Digital Health, and Data Science

• 300 Artificial Intelligence, Digital Health, and Data Science

Authors

• Wang, Lin-Chun, Fresenius Medical Care, New York, New York, United States

Lin-Chun Wang, MS

• Yueh, Sheng-Han, Renal Research Institute, New York, New York, United States

Sheng-Han Yueh, MS

• Rao, Sharon, Fresenius Medical Care, New York, New York, United States

Sharon Rao, MD

• Bergling, Karin, Renal Research Institute, New York, New York, United States

Karin Bergling, MD, PhD

• Nandorine Ban, Andrea, Renal Research Institute, New York, New York, United States

Andrea Nandorine Ban, MS, MSc

• Ginsberg, Nancy, N/A, New York, New York, United States

Nancy Ginsberg, MS, RD

• Moore, Linda W., Houston Methodist Hospital, Houston, Texas, United States

Linda W. Moore, PhD, RDN

• Biruete, Annabel, Purdue University, West Lafayette, Indiana, United States

Annabel Biruete, PhD, RD, FASN

• van der Sande, Frank, Maastricht Universitair Medisch Centrum+, Maastricht, LI, Netherlands

Frank van der Sande, MD, PhD

• Kooman, Jeroen, Maastricht Universitair Medisch Centrum+, Maastricht, LI, Netherlands

Jeroen Kooman, MD, PhD

• Kotanko, Peter, Renal Research Institute, New York, New York, United States

Peter Kotanko, MD, FASN

• Zhang, Hanjie, Renal Research Institute, New York, New York, United States

Hanjie Zhang, PhD

Background

Large language models (LLMs) may support renal dietitians by generating culturally tailored meal plans for patients receiving hemodialysis, but their nutritional accuracy and clinical reliability require validation. We evaluated the real-world feasibility of LLM-generated recipes and compared nutrient-estimation performance between GPT-4o and GPT-5.4.

Methods

We created 20 virtual patients using Monte Carlo simulation and selected one each with Western, Chinese, and Mexican food preferences. Their demographics, clinical characteristics, laboratory values, and food budgets were provided to ChatGPT (GPT-4o), which generated three recipes per patient. The resulting nine dishes were prepared in a professional kitchen, and nutritional content was analyzed using ESHA software, a nutrition analysis software used by the food industry. and the USDA Food Data Central, for energy, protein, total fat, carbohydrates, phosphorus, potassium, and sodium. GPT-4o estimates from August 2024 were compared with ESHA values per dish. The same nutrient-estimation analysis was repeated in April 2026 using GPT-5.4.

Results

GPT-4o underestimated energy, total fat, and potassium by 23–38%, while phosphorus and sodium estimates varied widely, with confidence intervals of 53–210% and 27–172%, respectively (Figure 1). GPT-5.4 showed improved estimation for several nutrients, including energy, total fat, phosphorus, potassium, and sodium; however, protein was overestimated by 27%.

Conclusion

LLMs show promise as supportive tools for dialysis nutrition planning, particularly for culturally tailored meal suggestions. However, nutrient estimates remain model- and nutrient-specific, and newer LLM versions should not be assumed clinically reliable without task-specific validation. Longitudinal benchmarking across model versions may be necessary to assess reproducibility and stability of AI-supported nutritional tools. Clinical oversight remains essential.