ProcessLife Sciences

AI pharmacovigilance case processing, from inbox to E2B(R3) submission

AI can take most of the reading, extraction and coding out of pharmacovigilance case intake, while the decisions that start or shape the regulatory clock stay with qualified safety staff. This page walks through six intake stages for individual case safety reports, from digitizing an email to generating an E2B(R3) file, and marks in each one what an agent prepares and what a person decides.

Reviewed 7 min read

On this page
  1. The path an individual case safety report takes through intake
  2. Why intake decides whether safety reports go out on time
  3. Six intake stages and the decision that stays human in each
  4. A call-center note that changes a case's due date
  5. Measures that show an automated intake workflow is working
  6. Failure modes specific to safety-case automation
  7. Controls to have in place before automated intake goes live
  8. Questions and answers
  9. Sources

The path an individual case safety report takes through intake

01Report sources02Intake anddigitization03Validity check04Seriousness triage05Coding and extraction06Duplicates andfollow-up07Narrative and ICSR
  1. Report sources

    Spontaneous reports, literature, patient support programs, call centers, partners and company digital channels.

  2. Intake and digitization

    Text, attachments and metadata pulled from emails, PDFs, call notes and articles.

  3. Validity check

    Four minimum criteria confirmed, or a follow-up request drafted.

  4. Seriousness triage

    Seriousness and expectedness proposed per event against reference safety information.

  5. Coding and extraction

    MedDRA terms for events, WHODrug entries for products, structured fields filled.

  6. Duplicates and follow-up

    Possible duplicates linked and targeted questions drafted for missing data.

  7. Narrative and ICSR

    Narrative drafted, E2B(R3) message built, checked and submitted.

Conceptual intake flow for one case. Human decisions sit at validity, seriousness, coding exceptions and submission; it does not depict a measured process or a live system.

Why intake decides whether safety reports go out on time

Cases arrive from many directions: spontaneous reports from healthcare professionals and consumers, published literature, patient support and market research programs, call centers, licensing partners under safety data exchange agreements and company-run digital channels. Most of it is unstructured, written for another purpose and incomplete.

The regulatory clock does not wait for intake. Under EU good pharmacovigilance practices, day zero is the date any company personnel, medical representatives and contractors included, first hold information meeting the minimum criteria, counted in calendar days1. Serious valid cases are due within fifteen days and non-serious ones within ninety1. In the US, serious and unexpected postmarketing cases are due no later than fifteen calendar days from initial receipt2. Every day a case sits in an inbox comes out of the time left for medical review.

Six intake stages and the decision that stays human in each

  1. Digitize every source into one case record

    The agent extracts text from emails, scanned forms, call transcripts and full-text articles, detects language, splits documents that mention several patients and keeps a link to the original. Nothing is interpreted yet: the aim is a complete record of what arrived and when.

    Output
    Source-linked intake record
    Owner
    Case intake specialist

  2. Check the four minimum criteria

    The agent looks for an identifiable reporter with qualification and country, a single identifiable patient, at least one suspect product and at least one suspected reaction, as GVP Module VI requires in line with ICH E2D1. It flags gaps and drafts a follow-up request. A person confirms validity and records day zero.

    Output
    Validity decision and day zero
    Owner
    Safety case processor

  3. Triage seriousness and expectedness

    For each event, the agent proposes seriousness criteria and compares the event with reference safety information, such as the company core data sheet or, for trial cases, the investigator's brochure. E2B(R3) records seriousness per event rather than per case4, so proposals are made event by event. A qualified safety professional makes both calls.

    Output
    Confirmed seriousness and expectedness
    Owner
    Safety physician or qualified reviewer

  4. Extract the data and code it

    The agent fills structured fields and proposes MedDRA lowest level terms for events, history and indications, plus WHODrug entries for suspect and concomitant products56. Each proposal carries a confidence score; low-confidence terms, and any term that would change seriousness, go to a coder.

    Output
    Coded case data
    Owner
    Medical coder

  5. Detect duplicates and plan follow-up

    The agent compares the case with existing ones on patient descriptors, product, event, dates and reporter, then proposes links for a person to accept or reject. It drafts follow-up questions only for information that would change the assessment, so reporters are not sent generic questionnaires.

    Output
    Duplicate decision and follow-up letter
    Owner
    Safety case processor

  6. Draft the narrative and build the ICSR

    From confirmed, coded data the agent drafts a chronological narrative and builds the E2B(R3) message for EudraVigilance or FDA's Adverse Event Monitoring System, formerly FAERS. FDA requires E2B(R3) for postmarketing ICSRs sent through its gateway from October 20263. A reviewer checks the narrative against the source before submission.

    Output
    Submitted ICSR
    Owner
    Case reviewer

A call-center note that changes a case's due date

Measures that show an automated intake workflow is working

MeasureWhat it tells youWarning sign
Cycle time per stageWhere cases wait between receipt, validity, triage, coding and submissionValidity or triage queues growing while coding speeds up
Late submissionsWhether serious and non-serious deadlines are metAny late serious case traced back to intake or triage
Coding agreementHow often coders accept MedDRA and WHODrug proposals unchangedNear-total acceptance with no sampled check, a sign of automation bias
Seriousness overturnsHow often reviewers change the proposed seriousness, in each directionA pattern of proposals less serious than the final decision
Rework rateCases reopened after quality review or a regulator queryRework concentrated in one source type or language
Follow-up yieldHow many follow-up requests return information that changes the caseReporters chased for data that would not alter the assessment

Track each measure by source type: a workflow that handles structured partner exchanges well can still fail on literature or call notes.

Failure modes specific to safety-case automation

Under-triage of seriousness

Early signalReviewers mostly correct proposals upward, from non-serious to serious.

MitigationUse asymmetric thresholds so uncertain events route as potentially serious, and sample cases the agent labeled non-serious.

Double reporting of literature cases

Early signalICSRs created for articles that EMA's medical literature monitoring already covers.

MitigationCheck substance and journal against the monitored list first; for that literature, marketing authorization holders should not submit the cases themselves1.

Silent model change

Early signalCoding acceptance or seriousness agreement shifts with no release on your side.

MitigationPin model versions, route updates through change control and re-run a reference case set before each release.

Patient identifiers spreading beyond the safety system

Early signalSource documents copied into prompt logs or third-party tools.

MitigationProcess cases inside your own environment and keep identifiers out of logs, since health data is special category data under the General Data Protection Regulation7.

Controls to have in place before automated intake goes live

0 of 5 checked

Questions and answers

Can AI make the seriousness call on an adverse event case?

It can propose one, and a good proposal saves review time, but the decision should rest with a qualified safety professional. In the EU, seriousness decides whether a valid case is due within fifteen or ninety days1, so an error has a direct regulatory consequence. Keep the proposal and the decision in the audit trail and monitor how often reviewers change it.

Can AI screen medical literature for adverse event reports?

Yes, it suits the task: the agent reads abstracts and full text, flags articles that mention a company product with an identifiable patient and reaction, and proposes which qualify as cases for a person to confirm. For substances and journals covered by EMA's medical literature monitoring, check the monitored list first so cases are not reported twice1.

How is patient data protected when AI processes safety cases?

Run models in an environment you control, keep identifiers out of logs and training data unless there is a documented need, and limit access to the people and services involved. Safety data is health data, which the General Data Protection Regulation treats as a special category in the EU7. Contracts with any model or hosting provider should fix data use, retention and location.

Does EMA allow pharmacovigilance models that keep learning?

EMA's reflection paper says pharmacovigilance may allow a more flexible approach, including incremental learning for classifying and scoring adverse event reports, while the marketing authorization holder stays responsible for validating, monitoring and documenting performance8. Releasing updates as controlled versions remains the simpler route to defend, because each change is tested before it touches live cases.

Sources

  1. Guideline on good pharmacovigilance practices (GVP) Module VI – Collection, management and submission of reports of suspected adverse reactions to medicinal products (Rev 2) — European Medicines Agency · checked 10 October 2026
  2. 21 CFR 314.80 — Postmarketing reporting of adverse drug experiences — Legal Information Institute, Cornell Law School · checked 10 October 2026
  3. FDA Adverse Event Monitoring System (AEMS) Electronic Submissions — U.S. Food and Drug Administration · checked 10 October 2026
  4. Electronic Submission of Postmarketing Individual Case Safety Reports to the FDA Adverse Event Monitoring System Using ICH E2B(R3) Data Standards — Federal Register · checked 10 October 2026
  5. MedDRA: Medical Dictionary for Regulatory Activities — MedDRA MSSO / ICH · checked 10 October 2026
  6. WHODrug — Uppsala Monitoring Centre · checked 10 October 2026
  7. Regulation (EU) 2016/679 (General Data Protection Regulation) — EUR-Lex · checked 10 October 2026
  8. Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle — European Medicines Agency · checked 10 October 2026

More in Life Sciences

Back to Life Sciences

Next step

Map your safety intake before automating any stage

Send us your case sources, rough volume by source and where cases wait today. We will reply with the stages worth automating first, the controls each needs and what validation would cover.

Discuss case intake automation