GLEN-Bench: A Graph-Language based Benchmark for Nutritional Health

TL;DR

GLEN-Bench tackles the need for end-to-end, nutrition-aware interventions by building the first graph-language benchmark that unifies diet, health, and socioeconomic signals from NHANES, FNDDS, and USDA data. By embedding these signals into a heterogeneous knowledge graph and evaluating three interdependent tasks—risk detection, personalized food recommendation, and explainable QA—the work demonstrates that graph-based reasoning combined with language models yields stronger, more interpretable guidance under real-world constraints. The extensive evaluation across 20+ models reveals that explicit relational structure and constraint-aware objectives improve both predictive performance and practical feasibility, with retrieval-augmented QA and constraint-aware recommendations delivering notable gains. Overall, GLEN-Bench provides a reproducible framework to advance nutrition-aware AI and informs end-to-end decision support for interventions beyond opioid misuse, while highlighting areas for improving faithfulness, safety, and causal interpretation.

Abstract

Nutritional interventions are important for managing chronic health conditions, but current computational methods provide limited support for personalized dietary guidance. We identify three key gaps: (1) dietary pattern studies often ignore real-world constraints such as socioeconomic status, comorbidities, and limited food access; (2) recommendation systems rarely explain why a particular food helps a given patient; and (3) no unified benchmark evaluates methods across the connected tasks needed for nutritional interventions. We introduce GLEN-Bench, the first comprehensive graph-language based benchmark for nutritional health assessment. We combine NHANES health records, FNDDS food composition data, and USDA food-access metrics to build a knowledge graph that links demographics, health conditions, dietary behaviors, poverty-related constraints, and nutrient needs. We test the benchmark using opioid use disorder, where models must detect subtle nutritional differences across disease stages. GLEN-Bench includes three linked tasks: risk detection identifies at-risk individuals from dietary and socioeconomic patterns; recommendation suggests personalized foods that meet clinical needs within resource constraints; and question answering provides graph-grounded, natural-language explanations to facilitate comprehension. We evaluate these graph-language approaches, including graph neural networks, large language models, and hybrid architectures, to establish solid baselines and identify practical design choices. Our analysis identifies clear dietary patterns linked to health risks, providing insights that can guide practical interventions.

Figures (6)

• Figure 1: An Overview of GLEN-Bench. Our framework brings scattered nutrition and health signals together into one pipeline for nutritional health assessment.
• Figure 2: The GLEN-bench construction process. (a) Data extraction from NHANES and USDA sources to collect multi-dimensional user and food information. (b) Graph construction of a heterogeneous knowledge graph modeling the relationships between clinical health status, dietary habits, socioeconomic barriers (e.g., food insecurity), and food nutritional profiles.
• Figure 3: Overview of the GLEN-Bench framework. The system builds on a unified nutrition-health Graph to perform three related tasks: (I) personalized food recommendation, (II) health risk detection to identify intervention needs, and (III) nutritional question answering that provides explainable, graph-based explanations.
• Figure 4: Personalized Food Recommendation performance across four model families (GNNs, collaborative filtering methods, contrastive-enhanced CF, and domain-specialized models) on GLEN-Bench.
• Figure 5: A case study of error analysis.