An explainable graph retrieval augmented generation framework for personalized nutrition recommendation

Introduction Dietary planning is essential for managing non-communicable diseases, yet many AI-based nutrition systems lack structured knowledge grounding, demographic sensitivity, and explainability. These limitations are particularly evident in culturally diverse contexts such as India, where standard approaches often fail to align clinical dietary requirements with traditional meal patterns. Methods This study proposes a graph-centric decision-support framework using a Graph Retrieval-Augmented Generation (Graph-RAG) architecture. A Neo4j knowledge graph models relationships among diseases, nutrients, food items, and demographic-specific Recommended Dietary Allowances (RDA). A semantic Extract, Transform, Load (ETL) pipeline integrates heterogeneous datasets and resolves terminology inconsistencies using embedding-based alignment. At inference time, nutrient requirements are retrieved from the graph and matched with food composition profiles using a deterministic cosine similarity-based ranking algorithm that prioritizes proportional nutrient balance. A language model is restricted to formatting graph-validated outputs. Results The framework was evaluated across multiple case-study scenarios, including anemia, hypertension, and diabetes, using diverse user profiles. Results indicate improved ranking consistency, greater alignment with nutrient requirements, reduced nutrient-dominance bias, and enhanced demographic sensitivity compared with baseline approaches. Ablation analysis shows that RDA-based normalization significantly improves nutritional balance. Discussion The results suggest that combining graph-native reasoning with constrained language generation supports transparent and knowledge-guided nutrition recommendations. The system improves interpretability while reducing risks associated with unconstrained generative models. However, the evaluation is conducted in a controlled setting, and the framework should be considered a decision-support tool rather than a clinically validated system. Future work includes uncertainty modeling, dataset expansion, and expert validation.