(P03-18-26) LLM-Based Personalized Nutrition Recommendations From 24-Hour Dietary Recall Data

Objectives: 24-hour dietary recalls are widely used in nutrition research and clinical practice, yet translating recall data into personalized, actionable recommendations remains time-intensive and manual. Few studies have explored automating this workflow. This study developed a system using a large language model (LLM) to convert dietary recall data into personalized reports and nutrition recommendations.

Methods: The system takes two types of inputs: dietary recall data coded using the Food and Nutrient Database for Dietary Studies (FNDDS), including food codes, meal type, and grams consumed; and participant characteristics including age, sex, and activity level. FNDDS provides the nutrient composition per 100 g for each food code, from which total daily intake is computed. Individualized nutrient targets are set using Dietary Reference Intakes (DRI) for micronutrients and Estimated Energy Requirements (EER) for energy needs. A report comparing intake to targets is provided to GPT-4.1-mini, which suggests food substitutions and portion changes. Each suggestion is returned as an FNDDS food code and gram amount. The system recalculates nutrient totals from the new diet and repeats until intake moves closer to targets. The system was evaluated with 103 NHANES participants.

Results: The system generated personalized reports for all participants. Results showed mean percentage changes from the original diet: energy -4.9% (SD = 14.3%), sodium -5.8% (SD = 29.5%), carbohydrates -9.2% (SD = 16.5%), protein +15.7% (SD = 40.1%), and calcium +15.8% (SD = 27.4%). The system reduced excess energy, sodium, and carbohydrate intake while increasing protein and calcium, indicating that recommendations shifted diets toward improved nutrient density. The mean absolute percentage change across all evaluated nutrients was 19.5% (SD = 11.5%).

Conclusions: These findings suggest that integrating an LLM with nutrition databases can automate personalized dietary assessment from recall data, potentially reducing the documentation burden on dietitians. Future work will expand evaluations, assess additional nutrients, and validate outputs with registered dietitians.

Funding Sources: N/A