AI in Pharmacovigilance
How AI agents are reshaping ICSR processing, MedDRA coding, and signal detection; what works in production today, and how to deploy it without breaking GVP, 21 CFR Part 11, or your QPPV's audit trail.
BY CAROLINA BOSCH · DIRECTOR, R&D CLINICAL DATA
What AI Pharmacovigilance Is
Pharmacovigilance is the discipline of collecting, assessing, and acting on adverse events across a product's lifecycle, clinical trials, post-marketing spontaneous reports, literature, patient support programs, and social listening. The work is regulated end to end: EMA GVP, FDA 21 CFR 314.80/600.80, ICH E2B(R3), and for computerized systems that touch the record 21 CFR Part 11 and EU Annex 11.
AI in PV means using NLP, LLMs, and now agentic systems to accelerate the parts of that workflow that are high-volume and pattern-heavy: intake, duplicate detection, MedDRA and WHODrug coding, narrative drafting, seriousness and expectedness assessment, and first-pass signal detection. The QPPV, safety physician, and drug-safety associate still own the medical judgment and the submission.
From Intake to E2B
Case intake
Emails, portal submissions, call-center transcripts, literature PDFs, and regulator downloads parsed into structured ICSR fields, patient, reporter, product, event, dates, outcome.
Duplicate detection
Embedding-based matching across the safety database catches near-duplicates a rules engine misses, same event, different reporter, different date format.
MedDRA coding
Verbatim adverse event and medical history terms mapped to MedDRA LLTs with confidence scores and version pinning; overrides feed the training loop.
Narrative drafting
First-draft case narratives generated from structured fields and source text, in the sponsor's house style, ready for medical edit rather than blank-page authoring.
Seriousness & expectedness
Proposed seriousness criteria and listedness against the current RSI, with the reasoning surfaced so the safety physician can accept, edit, or reject.
E2B(R3) submission
Validated safety database still owns the outbound gateway, AI accelerates authoring, but the E2B(R3) message and the audit trail are the regulated deliverable.
Agentic Signal Detection
Traditional signal detection leans on disproportionality, PRR, ROR, EBGM, run on FAERS, EudraVigilance, and the sponsor's safety database. It works, but it drowns reviewers in weak signals and misses combinations that live across product, indication, and demographic slices.
Agentic PV layers on top: an agent monitors new case volume and coding patterns, correlates them with literature and regulator communications, drafts a signal-evaluation memo with cited evidence, and routes it to the medical safety reviewer. The reviewer approves, edits, or dismisses, and the workflow logs every step for the periodic PBRER/PSUR.
"Agents do not replace the safety physician. They collapse the days between 'something looks off in the data' and 'here is a defensible first assessment on your screen.'"
Validation & GxP
Any AI component that touches a regulated safety record is a computerized system under 21 CFR Part 11 and EU Annex 11. That means intended use documented, risk-based validation against a gold-standard dataset, versioned model and prompt, controlled change management, and an audit trail that ties every AI-proposed value to the source text, model version, and human reviewer.
MedDRA and WHODrug versions must be pinned per reporting period, silent dictionary upgrades break comparability across PSURs. Prompts and tools available to an agent are configuration, not code, and belong in your change-control system. The QPPV signs off on the workflow, not on the model.
Metrics That Matter
The wrong metric is "cases processed per FTE." The right ones are compliance and quality: on-time submission rate for 7/15-day expedited reports, coding agreement against the gold-standard panel, narrative acceptance rate on first medical review, and signal false-positive and time-to-first-assessment.
Track them by product, by source, and over time. If AI-assisted cases drift on any of them, you catch it before an inspection finds it.
Getting Started
- 1
Pick one case source
Literature intake or portal submissions, bounded, high volume, low medical ambiguity. Bounded scope is the difference between a pilot that ships and a pilot that becomes a 483 observation.
- 2
Build a gold-standard set
A qualified safety physician and drug-safety associate re-process a stratified sample end to end. That is your ground truth for coding, narrative, and seriousness, not the vendor's benchmark.
- 3
Deploy assistive first
Every AI-proposed value goes through a human. Measure agreement, edit distance, and cycle-time saved. Only expand to autonomous steps where sustained agreement clears your audit bar.
- 4
Validate as a computerized system
Intended use, risk assessment, IQ/OQ/PQ against the gold-standard set, change control for model and prompt versions. Wrap the whole thing in an audit trail before it touches a live case.
- 5
Instrument for inspection
Source span, proposed value, confidence, model version, reviewer, final value, timestamp per field, per case. If EMA or FDA asked tomorrow, you should be able to replay a single case end to end.
Rolling out AI in your PV org?
I am open to senior leadership roles in AI, Healthcare & Life Sciences, and Developer Ecosystems, plus select consulting engagements on AI-assisted safety workflows, MedDRA/WHODrug governance, and validated agentic deployment.