TraceGraph: versioned, provenance-linked clinicalevidence graphs for guideline drift detection and safetysignal auditing

Abstract Clinical knowledge changes continuously: clinical guidelines are revised, drug la-bels are updated, and new safety signals emerge. However, most clinical knowledgerepresentations are static snapshots that lack verifiable provenance and cannot answera basic governance question: what changed, when did it change, and what evidencesupports the change? We present TraceGraph, a versioned clinical evidence graph thatlinks each clinical assertion to cryptographically verified provenance, supports point-in-time querying, and enables automated detection of guideline drift and safety-signal contradictions. TraceGraph initializes its concept layer from standardized clinical terminologies(SNOMED CT, RxNorm, UMLS) and maintains an evidence layer populated from3,842 guideline revisions (27 specialty societies), 12,417 FDA label updates, and ade-identified EHR medication and laboratory event stream (4.8 million events). Eachassertion is stored as an immutable, versioned record with deterministic assertionid,timestamp, evidencehash (SHA-256), and rule-derived confidence. A clinical-logicengine with 61 validation rules detects contradictions (e.g., coexisting indication andcontraindication without subgroup qualifiers) and triggers drift alerts when guidelinerecommendations change across versions. Across a curated benchmark of 2,000 guideline-change events and 1,200 safety-contradiction cases annotated by two clinical informaticists (Cohen’s κ = 0.87), Trace-Graph achieved 96.4% precision (95% CI: 95.1–97.5%) and 92.1% recall (95% CI: 90.4–93.7%) for drift detection (F1: 94.2%), and 97.8% precision (95% CI: 96.7–98.7%)and 90.6% recall (95% CI: 88.5–92.5%) for contradiction detection (F1: 94.0%). Per-specialty stratification showed no specialty below F1 89.4%. Ablation studies con-firmed that scope alignment contributed +12.5 percentage points to drift F1. Thesystem sustained 127,381 assertion ingestions per second with 32 concurrent writersand supported point-in-time rollback to any prior version in 5.13 s (p95) at 78 mil-lion assertions. These results demonstrate that provenance-linked versioning enablesauditable, continuously updated clinical knowledge representations suitable for digitalhealth governance and evidence monitoring.