(P03-12-26) A Dietary Knowledge Base for Precision Nutrition via Multi-Source Large Language Model-Based Data Fusion

Objectives: Precision nutrition requires a structured computable framework to capture the complex relationships between foods, nutrients, and health outcomes. We present NutriKB, a dietary-centric knowledge base (KB) designed to serve both artificial intelligence (AI) and human users. For AI applications, NutriKB provides reliable structured information to ground model outputs and reduce hallucination. For human users, it offers a user-friendly interface to explore concepts and uncover pathways between dietary and health outcomes.

Methods: NutriKB covers core nutritional concepts, from food composition to health outcomes. Food composition data are sourced from the FNDDS, while recommended daily allowances and nutrient–health outcome relationships are derived from the NIH-ODS fact sheets and the USDA DGA. In addition, NutriKB integrates data on dietary patterns and population demographics in the United States by incorporating the NHANES. These authoritative data sources employ heterogeneous formats and naming conventions, posing significant challenges for knowledge base construction. We propose an LLM-based agentic AI system that automates key tasks, including relation extraction, entity alignment, multi-source conflict resolution, and quality control.

Results: The constructed NutriKB contains 7 domains, including dietary patterns, food combos, foods, ingredients, nutrients, health conditions, and person. It captures a rich set of cross-domain knowledge, such as food’s nutrition profiles, recommended nutrient intakes for specific age-sex roup, and the impact of excess nutrient intake on chronic disease risk. To date, NutriKB encompasses over 9.3K food combos, 5.4K foods, 64 nutrients, and more other entities. Similar to existing food KBs such as FoodOn, NutriKB is accessible via a web-based interface at https://lujiaying.github.io/NutriKB, allowing users to browse nutrition concepts and visualize connections between diet and health outcomes.

Conclusions: NutriKB addresses a fundamental challenge in precision nutrition by transforming isolated heterogeneous dietary data into a structured, accessible KB. By bridging the gap between raw data and actionable knowledge, NutriKB serves as a foundational resource to advance precision nutrition.

Funding Sources: N/A