Loading
Skip to content
Miracle Minds AI
  • Try “HIPAA”, “model risk”, “FinOps” or “agentic”.

Pharmacovigilance and drug safety

Automating case processing pays off only when the combined human and AI process is at least as accurate as the manual one, costs less effort per case and still meets every reporting deadline. The design of human review and the evidence behind it decide that, not the model.

Part of AI for Life Sciences & Healthcare. 8 sources, checked .

reports processed in FDA's adverse event database (AEMS, formerly FAERS) by its Office of Surveillance and Epidemiology in calendar year 2025 (2,048,059 reports)
2.05M
reports processed in FDA's adverse event database (AEMS, formerly FAERS) by its Office of Surveillance and Epidemiology in calendar year 2025 (2,048,059 reports)
U.S. Food and Drug Administration, Center for Drug Evaluation and Research, 2026(opens in a new tab)
range of F1 scores, by case field, for AI extraction in a published industry case processing pilot, as reported by FDA authors in 2022
0.36 to 0.91
range of F1 scores, by case field, for AI extraction in a published industry case processing pilot, as reported by FDA authors in 2022
Drug Safety, 2022(opens in a new tab)
calendar days from initial receipt to report a serious and unexpected adverse drug experience to FDA under 21 CFR 314.80 (text in force November 2026)
15 days
calendar days from initial receipt to report a serious and unexpected adverse drug experience to FDA under 21 CFR 314.80 (text in force November 2026)
Electronic Code of Federal Regulations (FDA), 2014(opens in a new tab)

The problem

Case volume is why this work gets automated. FDA's Office of Surveillance and Epidemiology processed 2,048,059 reports in its adverse event database in calendar year 2025, and the database held more than 32.4 million reports as of December 2025.1 The system is now the Adverse Event Monitoring System (AEMS), which replaced FAERS in March 2026.1 The duties behind those reports sit with companies. An applicant must promptly review adverse drug experience information from any source, foreign or domestic, and report each serious and unexpected experience no later than 15 calendar days from initial receipt.3

The published evidence says AI helps but cannot run alone. In a 2022 review, two authors from FDA's Office of Surveillance and Epidemiology reported that F1 scores in a published industry case processing pilot ranged from 0.36 for finding the verbatim adverse event text to 0.91 for reporter occupation.2 They wrote that use without human review would require an F1 score approaching 1.0, and that the combined human and AI system must perform at least as well as the human-only system it replaces.2 Those results are from 2022 and earlier, so newer models need the same test on your own cases. FDA describes its own AI-enabled review tool, InfoViP, as "a decision-support tool, not a replacement for reviewers".1

Regulators are still forming their approach, and existing duties apply in full. FDA's Emerging Drug Safety Technology Program, announced in June 2024, meets with companies to learn how they use AI in pharmacovigilance and to inform potential regulatory and policy approaches.4 FDA's January 2025 guidance on AI to support regulatory decision-making for drugs is still a draft. It says sponsors remain responsible for postmarketing safety surveillance and reporting regardless of the technology used.5 In the EU, EMA's reflection paper, adopted in September 2024, says the marketing authorization holder remains responsible for validating, monitoring and documenting model performance and for including AI operations in the pharmacovigilance system.6

The international reference is the CIOMS Working Group XIV report on AI in pharmacovigilance, published on 4 December 2025.8 CIOMS describes it as a framework of principles and good practices rather than technical guidance, covering a risk-based approach, human oversight, validity and robustness, transparency, data privacy, fairness and equity, and governance and accountability.8 FDA's draft notes that pharmacovigilance process documentation is generally not submitted but kept under the sponsor's own procedures and made available on request, for example during an inspection.5 So each automated step needs a risk rating, a measured error rate, a defined human check and a record an inspector can follow.

What we build

  1. 01

    Multi-channel case intake and triage

    We build intake that reads emails, call notes, portal forms, partner reports and literature hits, checks whether each holds the minimum information for a valid case, records the receipt date that starts the reporting clock, and routes likely serious cases to a named reviewer first.

  2. 02

    Case processing with field-level evidence

    Extraction of patient, product, event and reporter fields, proposed MedDRA terms, suspected duplicates and a draft narrative, each linked to the source passage, with thresholds that decide which fields a person must confirm before submission.

  3. 03

    Literature screening

    Scheduled searches of reference databases and local journals, with articles ranked by the likelihood that they contain a reportable case, and sampled human review of the articles the model set aside.

  4. 04

    Signal detection support

    Disproportionality screening and case-series summaries that show safety physicians the underlying reports and timelines. The decision on whether a signal is real stays with them.

  5. 05

    Validation and quality control package

    Intended use, risk assessment, test sets drawn from your own closed cases, acceptance criteria, sampling plans for human review, change control and monitoring reports, written so the system can be defended in an inspection.

Where it goes wrong

The failure modes we design against from the first week, and the control for each.

  • Assuming the model can run unreviewed

    Published extraction results ranged from an F1 score of 0.36 to 0.91 depending on the field, and FDA authors judged that use without human review would require a score approaching 1.0.2 Control: set the level of human review per field and per case type from error rates measured on your own cases, and require the combined process to match the manual baseline before go-live.

  • Quality control that is too heavy or too thin

    The challenge with a human in the loop is to ensure quality without giving back the efficiency gained. FDA authors suggest effort proportional to the consequences of misclassification and sampling that oversamples reports scored close to the threshold.2 Control: a written sampling plan by risk tier, an error limit for each tier, and escalation to full review when a sample exceeds its limit.

  • Losing days on the reporting clock

    The 15 calendar days run from initial receipt by the applicant, not from the end of automated processing.3 In the EU, day zero is the day any personnel of the holder, including contractors, first has the minimum information.7 Control: timestamp receipt in every channel, track queue age against due dates, and keep a manual path for when the automation is unavailable.

  • Changing the model without revalidating

    FDA authors state that each time an algorithm is retrained, a formal validation of system performance will be required.2 EMA expects performance monitoring that detects drift early, with defined thresholds for acceptable performance.6 Control: version every model and prompt, rerun a locked test set before each release, monitor production performance against thresholds, and keep a rollback path.

Rules that apply

  • 21 CFR 314.80, postmarketing reporting of adverse drug experiences

    Primary text on ecfr.gov(opens in a new tab)

    Applicants must promptly review adverse drug experience information from any source, report each serious and unexpected experience within 15 calendar days of initial receipt, keep written procedures, and retain records for 10 years.

  • EU Good Pharmacovigilance Practices, Module VI (Rev 2)

    Primary text on ema.europa.eu(opens in a new tab)

    Serious valid cases are submitted within 15 days of receipt and non-serious cases within 90 days. Holders are expected to screen widely used literature databases at least once a week, quality control data entry and keep an audit trail of all data entered or modified.

  • EMA reflection paper on AI in the medicinal product lifecycle (September 2024)

    Primary text on ema.europa.eu(opens in a new tab)

    The marketing authorization holder remains responsible for validating, monitoring and documenting model performance and for including AI operations in the pharmacovigilance system.

  • CIOMS Working Group XIV report on AI in pharmacovigilance (December 2025)

    Primary text on cioms.ch(opens in a new tab)

    A consensus framework of principles and good practices: a risk-based approach, human oversight, validity and robustness, transparency, data privacy, fairness and equity, and governance and accountability.

How to measure it

  • Share of expedited reports submitted within 15 calendar days of receipt, by intake channel
  • Precision and recall of extraction and coding against a human-verified sample, by field and case type
  • Error rate found in sampled review of auto-accepted fields and of literature the model set aside
  • Handling time per case and the share of cases reworked after human review
  • Model performance on the locked test set after each release, and duplicates found after submission

Questions buyers ask

Can AI fully automate adverse event case processing?

Not on the published evidence. FDA authors who reviewed industry pilots in 2022 found that extraction accuracy varied widely by field and concluded that full automation would need near-perfect scores. They expected a human in the loop to remain necessary for the foreseeable future. Plan for automation that prepares the case and a person who confirms the fields that matter, with sampled checks on the rest.

Does FDA allow AI in pharmacovigilance?

FDA's position so far is that reporting duties stay the same whatever technology you use. Its January 2025 draft guidance says sponsors remain responsible for postmarketing safety surveillance and reporting regardless of the technology. FDA also runs the Emerging Drug Safety Technology Program, where companies present how they use AI in pharmacovigilance. Those meetings are for mutual learning, not compliance advice, so keep your validation evidence ready for inspection.

How do you validate an AI system used for case processing or MedDRA coding?

Define the intended use and what goes wrong if each step is wrong. Test on a held-out set of your own closed cases, by field and case type, against acceptance criteria agreed in advance. Document the human checks, the sampling plan and the audit trail. Lock the version, revalidate after every model or prompt change, and monitor production performance against set thresholds.

Sources

Every figure on this page comes from the sources below. We checked each link on .

  1. 1.Office of Surveillance and Epidemiology - 2025 Annual Report(opens in a new tab)U.S. Food and Drug Administration, Center for Drug Evaluation and Research, September 2026 · fda.gov
  2. 2."Artificial Intelligence" for Pharmacovigilance: Ready for Prime Time?(opens in a new tab)Drug Safety, May 17, 2022 · pmc.ncbi.nlm.nih.gov
  3. 3.21 CFR 314.80: Postmarketing reporting of adverse drug experiences(opens in a new tab)Electronic Code of Federal Regulations (FDA), June 10, 2014 · ecfr.gov
  4. 4.Agency Information Collection Activities; Submission for Office of Management and Budget Review; Comment Request; Emerging Drug Safety Technology Program (91 FR 7501)(opens in a new tab)U.S. Food and Drug Administration (Federal Register), February 18, 2026 · govinfo.gov
  5. 5.Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products: Draft Guidance for Industry and Other Interested Parties(opens in a new tab)U.S. Food and Drug Administration, January 2025 · fda.gov
  6. 6.Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle (EMA/CHMP/CVMP/83833/2023)(opens in a new tab)European Medicines Agency, September 9, 2024 · ema.europa.eu
  7. 7.Guideline on good pharmacovigilance practices (GVP) Module VI: Collection, management and submission of reports of suspected adverse reactions to medicinal products (Rev 2)(opens in a new tab)European Medicines Agency and Heads of Medicines Agencies, July 28, 2017 · ema.europa.eu
  8. 8.Artificial Intelligence in Pharmacovigilance: Report of the CIOMS Working Group XIV(opens in a new tab)Council for International Organizations of Medical Sciences (CIOMS), December 4, 2025 · cioms.ch