The Safety Signal: AI in Pharmacovigilance

Pharmacovigilance is the continuous work of collecting, assessing, understanding, and preventing adverse effects and other medicine-related problems. AI can help teams triage reports, find duplicates, code concepts, monitor literature, and prioritize statistical patterns. It cannot turn an association into causation, replace medical judgment, or decide a product's benefit-risk balance on its own.

Medical and regulatory notice: This article is operational education, not medical, regulatory, or legal advice. Patients and caregivers should report concerns through their clinician or the official channel in their country and seek urgent care when appropriate. Reporting obligations, timelines, terminology, and responsible roles vary by product, study phase, and jurisdiction; the applicable current rules and competent authority control.

A safety report is not yet a safety signal

An individual case safety report, or ICSR, records information about a patient, one or more suspect or interacting products, one or more adverse events, a reporter, and the clinical course. A valid report can still be incomplete. It may contain an uncertain exposure date, missing dose, ambiguous seriousness, duplicate submission, or a narrative translated through several systems.

A signal is information suggesting a new potentially causal association, or a new aspect of a known association, that warrants further investigation. One case can be important, especially for a rare, serious, and biologically plausible event, but report count alone does not establish incidence or causality. Spontaneous-report databases are affected by under-reporting, stimulated reporting, missing denominators, duplicates, confounding, and notoriety bias.

This distinction is the first product requirement. A model can recommend “possible duplicate,” “needs expedited review,” or “unexpected term candidate.” It should not label a medicine unsafe or tell a patient to change treatment.

The current regulatory baseline is multi-jurisdictional

The ICH E2D(R1) Step 4 guideline, adopted in August 2025, provides harmonized definitions and standards for post-approval safety data management. Regional implementation still matters. Teams should map each requirement to effective local law, guidance, implementation dates, and product commitments rather than treating an ICH guideline as self-executing everywhere.

In the European Union, the EMA's Good Pharmacovigilance Practices page identifies final modules and current transition notes. As of July 2026, EMA says GVP modules affected by E2D(R1) and ICH M14 will be revised; in the interim, the new ICH guidance and EU implementation material should be applied where they affect GVP. GVP Module IX remains the central signal-management reference, while the 2025 implementing-regulation changes and EMA's 2026 questions-and-answers affect current EU practice. That is a live transition, not permission to cite an obsolete module in isolation.

In the United States, FDA is implementing the Adverse Event Monitoring System, or AEMS, formerly FAERS, while continuing to publish FAERS database information. FDA also operates Sentinel, a distributed active-surveillance system. These sources serve different purposes; a spontaneous report, literature finding, registry result, and active database analysis should not be blended without provenance.

Build intake around data lineage

Reports can arrive from patients, clinicians, partners, market-research programs, social or digital channels, literature, clinical programs, product-support services, and regulators. Each channel has different consent, privacy, follow-up, and clock-start implications.

The intake record should preserve:

• original content and attachments, not only the extracted fields;
• received time, awareness time, channel, sender role, and market;
• product as reported and mapped medicinal-product identifier;
• event verbatim and coded terminology;
• patient and reporter elements required to assess validity;
• seriousness, listedness or expectedness, and causality assessments with their authors;
• follow-up attempts, merged duplicates, translations, redactions, and submission acknowledgements;
• model, prompt, ruleset, terminology version, and reviewer for every AI-assisted transformation.

If the system translates a narrative, extracts dose, or proposes a MedDRA term, store both source and proposal. Never overwrite the reporter's words. The patterns in our AI data-quality observability guide are directly relevant: missingness, latency, schema drift, and silent mapping failures are safety risks, not mere analytics defects.

Automate case processing in bounded stages

The safest automation is modular. One component detects whether a message may contain a report. Another extracts candidate entities. Another finds possible duplicates. Another suggests coding or case priority. Each output should have confidence, evidence span, and escalation behavior.

Useful applications include:

• detecting reportable content in high-volume inboxes;
• extracting patient, product, event, dose, dates, outcomes, and reporter details;
• comparing fuzzy identity, clinical sequence, and source metadata to identify duplicates;
• suggesting MedDRA coding while retaining the verbatim event;
• translating narratives with medical-term preservation;
• prioritizing possible serious or time-critical cases for trained review;
• generating follow-up questions from missing material fields;
• checking structured fields against the narrative before submission.

Do not let a language model silently assign final seriousness, expectedness, causality, reportability, or submission timing. Those determinations are governed and context-dependent. The approval interface should show the source passage, candidate field, reason, confidence, and any conflicting evidence. Our human-approval design guide explains why a meaningful review needs time, authority, and the ability to reject—not just a decorative confirm button.

Detect signals across complementary evidence

Disproportionality methods can highlight drug-event pairs reported more often than expected in a database. AI can also cluster narratives, recognize syndromes, connect product-quality complaints, monitor literature, and help define cohorts in longitudinal data. None of these methods is a universal signal detector.

The FDA's spontaneous-report systems are valuable for hypothesis generation but do not provide reliable exposure denominators. Sentinel can support active analyses in distributed healthcare data. The WHO Programme for International Drug Monitoring, operated with Uppsala Monitoring Centre, uses VigiBase, a global database of individual case reports. Each data source has different coverage, coding, access, and bias.

A signal workbench should preserve the method, population, time window, background comparator, inclusion and exclusion rules, deduplication state, terminology version, and every analyst decision. It should also show negative or contradictory evidence. Safety scientists then validate, prioritize, assess, recommend action, and track the decision through closure or continued monitoring.

Prevent false confidence from text models

Safety narratives contain negation, uncertainty, temporality, family history, reported speech, and competing causes. “No evidence of liver injury before treatment” must not become “liver injury before treatment.” A model may also confuse indication with event, concomitant medicine with suspect product, or hospitalization for observation with a seriousness outcome.

Test specifically for:

• negation and hypothetical statements;
• pregnancy exposure and maternal-infant linkage;
• dechallenge and rechallenge sequences;
• multiple products, doses, formulations, and routes;
• reporter corrections and follow-up versions;
• translated medical concepts and local brand names;
• tables, scanned forms, handwriting, and malformed PDFs;
• prompt injection embedded in emails or documents;
• rare designated medical events and low-frequency terms.

Generative summaries must cite narrative spans. If evidence conflicts, the output should say so. If the input is incomplete, “unknown” is safer than a plausible fill. A safety system should reward calibrated abstention.

Validate the human-AI process

Case-level performance should include sensitivity for report detection, field precision and recall, seriousness false-negative rate, coding agreement and clinically meaningful disagreement, duplicate precision and recall, case completeness, correction rate, and time to validated intake. Break results down by language, channel, product type, event rarity, scan quality, and follow-up status.

Signal-level evaluation is harder. Measure time from evidence availability to review, enrichment of true review-worthy pairs, false-alert burden, stability across database updates, documentation completeness, and whether important cases were delayed or missed. Back-testing must respect what was knowable at the historical cutoff; using later labels leaks the answer.

Operational validation should also cover audit trails, access control, electronic records and signatures where applicable, business continuity, supplier change notification, terminology updates, and reconciliation with regulatory acknowledgements. The goal is not an impressive aggregate accuracy. It is controlled performance on the cases where error could delay reporting or distort a safety assessment.

A worked example: an incomplete serious report

A patient-support email says that a person taking Product A “collapsed and spent two nights in hospital,” but gives no patient initials, country, dose, event date, or reporter contact preference.

1. The intake detector flags potential reportable safety content and preserves the original email and receipt time.
2. Extraction proposes Product A, collapse, and hospitalization, linking each field to text.
3. A rule-model combination prioritizes the item because hospitalization may meet a seriousness criterion, but it does not make the final determination.
4. A trained reviewer confirms the case status under the applicable procedure and sends approved follow-up questions.
5. Duplicate search compares available information without merging on product and event alone.
6. New details are appended as follow-up, not substituted for the original report.
7. Coding, seriousness, expectedness, causality, and reporting clock decisions are approved by authorized personnel.
8. Submission and acknowledgement are reconciled, and the case remains available for aggregate and signal review.

The model saved time by finding and structuring the report. It did not diagnose the collapse, establish causality, or decide treatment.

Governance and release gates

Before production, require:

• an approved intended use and explicit prohibited uses;
• a jurisdiction, product, and process applicability matrix;
• representative validation with pre-agreed acceptance thresholds;
• documented human roles and escalation for uncertainty;
• original-to-output traceability and immutable audit history;
• privacy, security, data-residency, and least-access controls;
• vendor qualification and model-change notification;
• terminology and ruleset version control;
• downtime, backlog, reconciliation, and manual-processing procedures;
• ongoing drift, error, complaint, and missed-case monitoring.

Changes should be risk-classified. A user-interface fix, new language model, modified prompt, terminology release, new intake channel, and new product population do not have the same validation need. Retest the affected intended-use slice and keep approval evidence. For additional control design, see our AI audit evidence assurance framework.

Source notes

Reviewed 2026-07-30. Main sources are ICH E2D(R1) Step 4; EMA's current GVP index and transition notes; FDA's AEMS overview, FAERS database page, and Sentinel overview; and Uppsala Monitoring Centre's description of VigiBase and the WHO programme. ICH adoption does not eliminate regional implementation requirements, and EMA's page explicitly describes an ongoing GVP transition.

Questions safety leaders should ask

Can AI submit cases directly?

That depends on applicable requirements and the validated system design, but high-consequence determinations should remain under authorized human control. Even a technically valid transmission can contain a clinically material extraction or coding error.

Does a statistical alert prove causality?

No. It is a screening result that may trigger validation and assessment. Reporting biases, confounding, duplicates, and chance must be considered alongside clinical and epidemiological evidence.

Can public adverse-event data estimate risk?

Usually not by simple report counts. Spontaneous systems often lack reliable exposure denominators and have substantial reporting biases. Appropriate epidemiological methods and other data may be needed.

What is the safest first deployment?

Source-linked intake extraction, quality checks, or duplicate recommendation under expert review often provide measurable value without delegating final regulatory judgment.

Keep judgment visible

Good pharmacovigilance AI makes safety work more observable. It preserves what was reported, shows what the model inferred, records what a qualified person decided, and connects the case to later aggregate assessment. The standard is not whether automation can fill a form. It is whether the organization can demonstrate that automation helped it find, evaluate, report, and learn from safety information without hiding uncertainty or weakening accountability.