Representation Before Retrieval: Structured Patient Artifacts Reduce Hallucination in Clinical AI Systems

BackgroundLarge language models show promise for clinical decision support, yet their propensity for hallucination--generating plausible but unsupported claims--poses sub-stantial patient safety risks. Retrieval-augmented generation (RAG) is widely assumed to mitigate this problem by grounding outputs in retrieved documents, but this assumption remains inadequately tested in clinical contexts where information density, temporal complexity, and safety stakes are uniquely high. MethodsWe developed a system that compiles heterogeneous patient data (electronic health records, wearables, genomics, imaging reports) into structured, machine-readable artifacts with explicit provenance tracking across seven clinical domains. We evaluated four conditions: baseline LLM (C0), RAG over raw clinical text (C1), artifact-augmented single-pass generation (C2), and artifact-augmented multi-step agent workflow with verification (C3). Using 100 synthetic patient vignettes evaluated across 3 random seeds (N = 300 per condition, 1,200 total), we measured unsupported claim rates, factual accuracy, temporal consistency, contraindication detection, and clinical safety metrics using GPT-4o-mini with physician-adjudicated safety review. ResultsRAG substantially increased hallucination: unsupported claim rates rose from 5.0% (95% CI: 3.8-6.4%) at baseline to 43.6% (95% CI: 40.1-47.2%) with retrieval--an 8.7-fold increase (p < 0.001, Cohens d = 2.31). Structured artifacts reduced unsupported claims to 8.4% (95% CI: 6.7-10.3%) in single-pass generation, a 40% relative reduction versus baseline (p = 0.02, d = 0.48). The agent workflow achieved 21.1% unsupported claims with the lowest contraindication miss rate (0.04) and highest clinician utility scores. Ablation analysis revealed that citation requirements and constraint checking contributed most to safety improvements. ConclusionsContrary to prevailing assumptions, RAG increases rather than decreases hallucination in clinical text generation. Structured representation with explicit provenance offers a more effective approach to grounding LLM outputs in verifiable patient data. We propose an information-theoretic framework explaining why representation quality determines the ceiling on factual reliability, while agentic verification affects uncertainty handling and safety constraint enforcement.