AI in pharmacovigilance: case intake, MedDRA coding and medical review

AI in pharmacovigilance is the use of machine learning, natural language processing and large language models to help drug safety teams receive, process, code and assess adverse event reports. Its longest-running uses are in individual case safety report (ICSR) processing, where it extracts case data from narratives, finds duplicates and proposes MedDRA terms, and in signal detection. The Council for International Organizations of Medical Sciences (CIOMS) Working Group XIV report on artificial intelligence in pharmacovigilance, published on 4 December 2025, sets seven guiding principles for its use, beginning with a risk-based approach and human oversight.

The practical question is which steps an AI system can perform, which it can propose, and which a qualified person decides.

Where is AI in pharmacovigilance already used?

AI in pharmacovigilance is already used for duplicate detection, encoding to standard terminologies, extracting information from case narratives, translation, triage of incoming reports, screening the scientific literature, and signal detection, according to the CIOMS Working Group XIV report's landscape analysis (section 2.1).

All of it runs on individual case safety reports. ICH E2D(R1), which reached ICH Step 4 on 15 September 2025, defines an ICSR as a description of an adverse event, adverse drug reaction or other observation in an individual patient at a specific point of time (what an individual case safety report contains). The same report traces the history:

• Duplicate detection has been in routine use by some organizations since at least the 2010s, including for WHO's VigiBase, FDA's adverse event reporting database and EudraVigilance (CIOMS, section 2.1.2).
• Signal detection saw the earliest real-world use of simple, narrow, rule-based AI in the late 1990s, when disproportionality analysis was built into triage algorithms to direct specialists' attention (CIOMS, section 2.1.3). Disproportionality counts cases by MedDRA term, which is why coding consistency drives signal detection.

Much of this work is still early-stage: the CIOMS report notes that much AI research and development for pharmacovigilance has been experimental, with no real-world deployment or only pilot use. Its examples of deployed solutions include duplicate detection, automated triage of case reports, and literature surveillance at EudraVigilance and the Netherlands centre Lareb.

For case processing, a pilot by Schmider and colleagues in Clinical Pharmacology & Therapeutics (2019) tested three commercial vendors' systems on adverse event source documents. It confirmed that AI could support extraction from source documents and evaluation of case validity; as a pilot, it measured feasibility, not performance in routine use.

FDA's January 2025 draft guidance on AI for regulatory decision-making lists "identifying, evaluating, and processing for reporting postmarketing adverse drug experience information" among AI uses, and CDER's Emerging Drug Safety Technology Program (EDSTP) focuses on AI in postmarket pharmacovigilance.

How does an adverse event case move from intake to submission?

An adverse event case moves through eight steps: intake and validity, duplicate check, data entry, MedDRA coding, seriousness and expectedness, causality, narrative and medical review, and submission. Each carries a different decision.

Step	What AI can do	What a qualified person decides	Primary source
Intake and validity	Proposes: the four minimum criteria, from any source	Whether the case is valid	ICH E2D(R1), section 2.2
Duplicate check	Performs: scores likely duplicates across sources	Whether two records are one case	ICH E2D(R1), section 6.6
Data entry	Proposes: narrative content mapped to E2B(R3) data elements	Whether the structured case says what the source says	ICH E2B(R3)
MedDRA coding	Proposes: a Lowest Level Term beside its verbatim	Which term is selected, and that nothing was added	MedDRA Term Selection: Points to Consider
Seriousness and expectedness	Proposes: seriousness flags and a label comparison	Serious and unexpected, so the 15-day clock applies	21 CFR 314.80
Causality	Assembles: the evidence a named method asks for	The causality category	WHO-UMC system
Narrative and medical review	Proposes: a draft narrative in the patient's chronology	That the narrative is a correct medical story	ICH E2D(R1), sections 6.2 and 6.3
Submission	Performs: assembles the E2B(R3) message	Sign-off and submission	21 CFR 314.80; ICH E2B(R3)

An illustrative case. An oncology nurse emails the safety inbox: a 64-year-old woman was admitted with fever and very low neutrophils after cycle 2 of a named regimen. The AI system shows the four minimum criteria it found, proposes the Lowest Level Term Febrile neutropenia beside the verbatim, and flags inpatient hospitalization as a seriousness criterion. Expectedness and causality stay with the reviewer. Febrile neutropenia is also a CTCAE v6.0 term.

Where does AI help in case intake?

In case intake, AI reads each inbound source on arrival, scanned documents included, and extracts the four minimum criteria, so a person reaches the validity decision sooner. It does not move day zero.

Under E2D(R1), day zero is the date any personnel of the marketing authorization holder, including service providers acting for it, obtain enough information to show a case meets the four minimum criteria:

• at least one adverse event, adverse drug reaction or other observation;
• at least one suspect or interacting medicinal product;
• an identifiable patient;
• at least one identifiable reporter.

A case missing any of them does not qualify, and the guideline asks for due diligence to collect it.

How is AI used for MedDRA coding?

In MedDRA coding, AI proposes a Lowest Level Term for each reported term, shown with the verbatim it came from and its place in the hierarchy, so a coder can accept or change it on one read.

In the five-level MedDRA hierarchy, each Lowest Level Term (LLT) links to one Preferred Term (PT), and each PT has one primary System Organ Class, which decides where it appears in cumulative outputs. Two coders choosing different LLTs for one verbatim can split an event across two PTs, so an aggregate table shows two smaller counts.

The ICH-endorsed MedDRA Term Selection: Points to Consider, Release 4.26 of March 2026, sets the rules any coder follows, human or machine:

• Always select a Lowest Level Term.
• Use medical judgment when no exact match exists.
• Select terms for all reported information, and do not add information.

An AI coder has to be built to keep the last rule: a term the source text does not support is an addition, however plausible.

For oncology trials, NCI states that each CTCAE v6.0 term is a MedDRA v28.0 LLT, so an adverse event graded on CTCAE v6.0 already names its LLT (MedDRA v28.0 LLT coding alongside the grade).

Where does human judgment stay in pharmacovigilance AI?

Seriousness, expectedness, causality and the final narrative each decide a reporting obligation or make a regulatory statement, so they stay with a qualified person.

Seriousness and expectedness. Under 21 CFR 314.80(a), an adverse drug experience is serious when it results in death, a life-threatening experience, inpatient hospitalization or its prolongation, persistent or significant disability, or a congenital anomaly; important medical events can qualify on medical judgment. AI can flag each criterion and set the event beside current labeling. A person decides, because that decision determines whether the 15-day expedited clock applies.

Causality. The WHO-UMC causality system uses six categories, from certain to unassessable (WHO-UMC compared with Naranjo and Kramer). Its documentation says causality assessment cannot prove the connection between drug and event or change uncertainty into certainty. AI can assemble what each category asks for (time to onset, response to withdrawal, other explanations) and show where it found each piece. The category is a medical judgment.

Medical review. E2D(R1) calls the narrative a stand-alone medical story in the chronology of the patient's experience, and medical review the step that ensures correct interpretation. AI can draft the chronology; the reviewer confirms it.

How much oversight, and where. The CIOMS report scales oversight to risk (section 3.2.1): a missed duplicate in a very large database has little or no consequence for patient safety, while a very serious signal detected late, under mass exposure, is very rare but can have dramatic public health consequences. FDA's draft guidance calls this model risk: model influence (how much the AI output contributes) combined with decision consequence (the significance of an incorrect decision).

The report also names automation bias (excessive trust in AI output) and warns that generative AI can produce coherent, convincing output that is wrong. Showing the source text beside every output lets a reviewer check the case against the record, not the system's summary.

What do CIOMS Working Group XIV and regulators expect?

The CIOMS Working Group XIV report sets out seven guiding principles, which members of the working group also summarized in Drug Safety on 26 August 2026:

1. Risk-based approach
2. Human oversight
3. Validity and robustness
4. Transparency
5. Data privacy
6. Fairness and equity
7. Governance and accountability

In postmarketing pharmacovigilance, FDA's draft guidance notes that process documentation is generally kept under the sponsor's procedures and made available on request, for example during an inspection.

In the EU, the European Medicines Agency's reflection paper on AI in the medicinal product lifecycle, adopted by CHMP on 9 September 2024, makes the marketing authorization holder responsible for validating, monitoring and documenting model performance and for including AI operations in its pharmacovigilance system.

Five questions for any AI system proposed for case work:

1. Does every output show its source text?
2. When the source is silent, does it flag the gap or return a term anyway?
3. Which MedDRA version does it code to, and how are updates handled?
4. Is every acceptance and override recorded, with who and when?
5. How is performance monitored, and what triggers more human review?

Where does Burna AI fit in the adverse event workflow?

Burna AI works upstream of the safety database, at the trial encounter where an adverse event is first documented. Under 21 CFR 312.64(b), an investigator must immediately report any serious adverse event to the sponsor, with an assessment of whether the drug could have caused it.

Burna AI grades each event against the CTCAE version the protocol names, and attribution is scored against every drug in the regimen using WHO-UMC and Kramer, with the reasoning shown. Its answers to the questions above:

• Source text. Every grade traces to the line in the medical record that produced it.
• Gaps. A grade whose citation set is incomplete is flagged before it is signed.
• MedDRA version. Each grade carries its MedDRA LLT code; on CTCAE v6.0, that is the MedDRA v28.0 LLT NCI assigned to the term.
• Acceptance and override. A grade becomes a record only when a qualified investigator signs it with a 21 CFR Part 11 electronic signature, and the signature carries what was accepted, what was overridden and the citation set as signed.

For a sponsor, each event arrives graded, coded and attributed with its evidence, which helps elucidate the safety profile the label will rest on. The platform's scope runs from first-in-human trials through postmarket pharmacovigilance.

Burna AI has submitted a Q-Submission to FDA, and its application to the CDER Emerging Drug Safety Technology Program (EDSTP) is in progress.

Common questions about AI in pharmacovigilance

Will AI replace pharmacovigilance case processors?

Not in the CIOMS Working Group XIV framing: AI takes on tedious, repetitive pharmacovigilance tasks so safety experts can focus on work needing their full expertise, such as causality assessment (CIOMS Working Group XIV). The working group warns that pharmacovigilance automation aimed only at efficiency and financial targets risks removing the value humans in the loop add.

Does AI change the expedited reporting clock?

No. Under E2D(R1), the 15-calendar-day clock for an expedited postmarketing report runs from day zero, the date anyone acting for the marketing authorization holder obtains enough information to show the case meets the four minimum criteria. In the US, 21 CFR 314.80(c)(1)(i) requires each serious and unexpected adverse drug experience to be reported no later than 15 calendar days from initial receipt of the information. For an IND safety report, the 15 days under 21 CFR 312.32(c)(1) run from the sponsor's determination that the information qualifies (how IND safety reports are timed).

What is CIOMS Working Group XIV?

CIOMS Working Group XIV is the Council for International Organizations of Medical Sciences working group on artificial intelligence in pharmacovigilance. Its report, published on 4 December 2025, sets seven guiding principles that start with a risk-based approach and human oversight. Members of the working group summarized those principles in Drug Safety on 26 August 2026.

What is the difference between human-in-the-loop and human-on-the-loop?

The CIOMS Working Group XIV report (section 4.2.2) defines human-in-the-loop as the capability for human intervention in every decision cycle, and human-on-the-loop as intervention in a system's design and the monitoring of its operation. Under its risk-based approach (section 3.3), a deployment may start with a high share of human-in-the-loop review and reduce human controls gradually as confidence in routine performance grows.

If your team is deciding where AI belongs in its adverse event workflow, book a 15-minute conversation.