Graph-Augmented Retrieval for Digital Evidence-Based Medical Synthesis: A Proof-of-Concept Study on Topology-Aware Mechanistic Narrative Generation

This study presents a domain-constrained evolution of retrieval-augmented generation (RAG) tailored to biomedical evidence synthesis. While conventional RAG systems rely primarily on semantic similarity, they often lack structural safeguards required for high-stakes scientific reasoning, including mechanistic completeness, temporal governance, and explicit discrimination between evidentiary absence and retrieval insufficiency. To address these limitations, we developed a two-layer architecture integrating metadata-constrained dense vector retrieval (RAG01) with a graph-augmented overlay (Graph-RAG, RAG02). The system operates on a closed, version-controlled corpus of 11,861 peer-reviewed text chunks derived from 627 publications on iron deficiency. Entity extraction at the chunk level enabled construction of a directed weighted co-occurrence graph (30 nodes, 118 edges), representing structural relationships among key biomedical concepts. Retrieval was conducted under deterministic constraints (top-k = 5, cosine threshold = 0.50, publication year ≥ 2023), ensuring reproducibility and temporal consistency. Graph-augmented scoring combines semantic similarity with topological reinforcement, incorporating connectivity, induced subgraph density, modular overlap, and multi-hop stability diagnostics. Mechanistic topic planning is organized through predefined biological axes functioning as epistemic probes. Each axis is evaluated not only for semantic relevance but also for structural embedding within the corpus-derived knowledge graph. In a case study of obesity-associated iron deficiency, the audited topology demonstrated a centralized regulatory core centered on hepcidin. Graph-aware retrieval preserved semantic alignment while increasing mean cosine similarity (0.673 to 0.694) and reducing similarity dispersion. Only the inflammation-mediated hepcidin pathway exhibited stable multi-hop reinforcement, whereas alternative mechanisms lacked consistent structural embedding and were classified as weakly supported or unsupported. The proposed framework advances RAG beyond similarity-based summarization toward topology-aware biomedical evidence interrogation. By integrating deterministic retrieval, structural validation, causal scaffolding, and specialist oversight, the system operationalizes selected principles of evidence-based medicine within a controlled digital synthesis environment, with implications for reproducible AI-assisted systematic reviews.