FoodScribe: an open-source semantic framework for nutrient estimation from free-text dietary records

Area of Science:

• Nutritional epidemiology
• Computational nutrition
• Bioinformatics

Background:

• Summarizing dietary records at scale is a major challenge in nutritional epidemiology.
• Manual nutrient analysis is time-consuming and resource-intensive.

Purpose of the Study:

• To develop and evaluate FoodScribe, an automated system for quantifying nutrient profiles from free-text dietary records.
• To assess FoodScribe's efficiency and accuracy compared to manual methods.

Main Methods:

• FoodScribe combines ingredient parsing with nutrient retrieval from the USDA FoodData Central (FDC) database.
• The system was benchmarked using three large language model (LLM) providers on the Nutribench dataset.
• Nutrient estimation accuracy was evaluated using F1 scores for macronutrients.

Main Results:

• FoodScribe processed 3,807 meal descriptions in 2.5 hours, significantly faster than manual annotation.
• High accuracy was achieved in macronutrient estimation (F1=0.79-0.89), with better performance for protein than fat.
• Analysis of a Mediterranean diet cohort showed dietary shifts consistent with intervention patterns.

Conclusions:

• FoodScribe offers an efficient and accurate solution for large-scale dietary record summarization.
• The tool has potential applications in nutritional epidemiology, clinical trials, and personalized nutrition.
• Integration with metabolomics data revealed associations between dietary components and gut metabolites.