PS42: From Signal to Regulatory Action in Weeks: Glass-Box AI + Causal Inference for Pharmacovigilance

Pharmacovigilance is increasingly digital, but decision-making remains constrained by a persistent trade-off: spontaneous reporting can be fast but noisy and confounded; EHR/claims analyses can strengthen causal interpretation but are operationally slower and sensitive to design choices; and systematic reviews provide essential context but are difficult to keep current. This session proposes a practical "glass-box" framework for moving from signal detection to decision-grade evidence within weeks by explicitly quantifying, auditing, and propagating uncertainty across multiple evidence streams. Rather than treating each data source in isolation, the session focuses on cross-source triangulation-a coherent workflow that aligns (1) spontaneous-report signal detection, (2) EHR-based causal validation (e.g., target trial emulation / "digital twin" analyses), and (3) living evidence synthesis from the literature. A central theme is making the strengths, limitations, and failure modes of each data modality explicit and comparable for the same drug–event pair, enabling faster and more defensible escalation, de-escalation, and communication. What makes this session distinctive Many PV sessions highlight either advanced ML/LLMs for signal detection or causal inference designs for RWE studies. This session bridges both-showing how audited AI extraction and misclassification correction can feed causal models, and how EHR validation and literature synthesis can be integrated into a single decision workflow. The "glass-box" emphasis means each step is interpretable and auditable, with uncertainty quantified (e.g., AI extraction error, confounding, and design sensitivity) rather than hidden. Joint Q&A: A decision-making clinic Moderated discussion focused on real-world friction points and decision gates, including: • What to do when VAERS/FAERS, EHR, and literature disagree • How to set alert thresholds when AI extraction is imperfect • How to pre-specify controls and benchmarks without over-claiming "ground truth" • How to communicate subgroup findings and uncertainty to multidisciplinary decision-makers Intended audience and takeaways This session is designed for statisticians, epidemiologists, safety scientists/physicians, data scientists, and regulatory reviewers. Attendees will leave with: • A concrete, auditable workflow for accelerating signal-to-evidence timelines • A checklist of evaluation metrics and decision gates suitable for regulated settings • A practical template for cross-source triangulation ("evidence fingerprint") to clarify what each data source can-and cannot-reliably contribute to safety decisions