An Adverse-Event Provenance Graph for Linking Case Narratives, Exposure Evidence, Clinical Context, Analytical Decisions, and Safety Conclusions

Pharmacovigilance evidence is progressively transformed as source reports are received, linked, coded, interpreted, analyzed, corrected, and translated into safety conclusions. Existing systems frequently preserve the resulting records without preserving a complete, inspectable account of how each assertion, exposure estimate, contextual judgment, analytical result, and conclusion was derived. This article proposes an adverse-event provenance graph as an original conceptual architecture for linking source narratives, reporter and channel information, provisional case identities, product and exposure evidence, clinical context, competing explanations, coding operations, analytical decisions, and safety conclusions. The architecture separates original artifacts from extracted assertions, standardized codes, inferred relationships, statistical outputs, evidence interpretations, and decision statements. It further represents corrections, superseded records, human overrides, uncertainty, and historical graph states as first-class objects rather than undocumented workflow events. Evidence interpretation and safety conclusions are therefore modeled as traceable but contestable stages, not automatic consequences of signal detection. Validation would require source-to-assertion fidelity testing, duplicate-linkage assessment, temporal exposure reconstruction, coding reproducibility, clinical-context review, interpretation auditing, subgroup evaluation, change-control testing, and reconstruction of prior conclusions. Interoperability would require preservation of identifiers, uncertainty, semantic meaning, transformation history, and source linkage across systems. The principal contribution is a proposed assertion-centered architecture that treats provenance as part of the scientific evidence rather than administrative metadata. A complete provenance path would improve inspectability and error investigation, but it would not establish causal truth, clinical utility, regulatory acceptance, universal applicability, or deployment readiness.