From Statistical Signal Detection to Causal Understanding of Drug Harm through Traceable, Multisource Pharmacovigilance Evidence for Decision-Making

Pharmacovigilance systems can detect unusual patterns of reported adverse events, yet statistical prominence does not establish that a medicine caused the observed harm. Disproportionality measures, spontaneous reports, observational associations, case narratives, mechanistic findings, and computational predictions each address different questions and remain vulnerable to distinct biases. This Original Causal Safety Architecture Article proposes a traceable signal-to-causality evidence architecture for organizing those heterogeneous contributions without collapsing them into a single score or treating detection as causal confirmation. The architecture separates statistical alert generation, safety-signal formulation, case-level clinical and temporal assessment, source-specific causal analysis, mechanistic evaluation, multisource triangulation, causal-strength grading, and decision use. Triangulation is defined as the structured comparison of evidence generated through meaningfully different data sources and bias pathways, with contradictions and dependencies retained rather than averaged away. Proposed causal-strength grades characterize evidentiary support and residual uncertainty without representing validated probabilities. Decision thresholds are treated as context-dependent judgements influenced by harm seriousness, reversibility, preventability, exposure, evidence quality, and consequences of delayed or premature action. A traceability layer records data provenance, coding, analytical choices, evidence dependence, contradictions, version history, and updating triggers. The original contribution is an end-to-end conceptual architecture connecting pharmacovigilance detection with conditional causal reasoning and accountable decision-making. Its components require computational, epidemiological, clinical, temporal, and implementation validation. The architecture cannot eliminate missing data, reporting bias, confounding, measurement error, model drift, or expert disagreement, and it is not a validated clinical tool, regulatory pathway, or deployment-ready decision system.