Related Experiment Video
Updated: Jun 19, 2026
13:19
Deep Neural Networks for Image-Based Dietary Assessment
Published on: March 13, 2021
Evaluation of Large Language Models for Mapping Dietary Data to Food Databases
Danielle G Lemay1, Michael P Strohmeier2, Richard B Stoker3
1United States Department of Food and Agriculture, Agricultural Research Service, Western Human Nutrition Research Center, Davis, CA, United States; Department of Nutrition, University of California, Davis, Davis, CA, United States.
The Journal of Nutrition
|June 17, 2026
Summary
Precision nutrition requires mapping foods to food composition databases (FCDs). A hybrid approach using semantic similarity and large language models (LLMs) achieved high accuracy in matching food descriptions to FCDs.
Area of Science:
- Nutritional Sciences
- Bioinformatics
- Computational Biology
Background:
- Standard food composition databases (FCDs) lack comprehensive biochemical data.
- Emerging food databases offer detailed biochemical information crucial for precision nutrition.
- Novel methods are essential to link dietary data foods to advanced FCDs.
Purpose of the Study:
- To establish benchmark datasets for evaluating food-matching methods.
- To assess the efficacy of large language models (LLMs) in mapping dietary foods to FCDs.
- To develop accurate methods for food text description matching.
Main Methods:
- Developed two benchmark datasets: ASA24-to-FooDB (large FCD) and NHANES-to-DFG2 (small FCD).
- Evaluated matching techniques including fuzzy matching, TF-IDF, semantic embedding, and LLMs.
- Implemented a hybrid approach combining semantic embeddings with LLM reranking.
Main Results:
- Semantic embedding achieved higher accuracy (87.8% and 48.0%) than traditional methods on both datasets.
- A hybrid semantic mapping and LLM reranking approach yielded the highest accuracies (90.7% on ASA24-to-FooDB, 65.4% on NHANES-to-DFG2).
- Prompt strategy optimization was necessary for different LLM sizes.
Conclusions:
- A combined approach of semantic similarity for initial matching and LLMs for reranking provides the most accurate food-to-FCD mapping.
- The FoodMapper application integrates this strategy into a user-friendly interface for nutrition scientists.
- This facilitates manual review and enhances the utility of biochemical food data for precision nutrition.
