Case intake and classification
Read free-text reports, extract the suspect product and the reaction, and order the queue so the team reviews urgent cases first.
Pharma · Pharmacovigilance and Medical Affairs
AI in pharmacovigilance: what it can do and what regulators expect
AI in pharmacovigilance helps classify adverse event reports, prioritize them by seriousness and detect signals for human review. Regulators accept it with conditions: the marketing authorization holder validates the model, documents it within its pharmacovigilance system, and a qualified professional still assesses each case and its causality.
Educational guide. Arkangel AI does not offer a pharmacovigilance system: it does not receive, code or report cases to any authority. This page explains where the physician questions shown in Arkangel Insights relate to pharmacovigilance and where they do not.
What is pharmacovigilance?
WHO defines pharmacovigilance as the science and activities relating to the detection, assessment, understanding and prevention of adverse effects or any other medicine- or vaccine-related problem. In practice, it is the system through which marketing authorization holders, health professionals and regulators report, analyze and act on adverse reactions: FAERS in the United States, the EMA framework and its good pharmacovigilance practices (GVP) in the EU, and national centers such as INVIMA, COFEPRIS and DIGEMID in Latin America.
Where AI is used in pharmacovigilance
The EMA reflection paper on AI names adverse event report management, classification and severity scoring, and signal detection. CIOMS frames AI as support for the expert, not a replacement.
Seriousness and timelines
Suggest whether a case looks serious so a professional can confirm it in time. Reporting timelines come from each local rule, not from the model.
Signal detection
Group similar reports and surface new patterns for the safety committee, which decides whether there is a signal and what to do about it.
Literature and other sources
Support the review of publications and other sources the holder must monitor. Causality assessment stays with a professional.
How AI helps with case processing and signal detection
Step by step, what a model can propose and what stays with the pharmacovigilance team. This describes how AI is used across the industry, not an Arkangel AI product.
- 1
Case intake
Read free-text reports and propose the four ICH E2D minimum elements: reporter, patient, reaction and suspect product. Incomplete reports are flagged for follow-up, and a professional confirms each field.
- 2
Coding and seriousness
Suggest MedDRA terms and whether a case looks serious. The coder confirms them, and the reporting deadline comes from each regulator's rules, not from the model.
- 3
Signal detection
Group similar reports and highlight drug-event combinations that appear more often than expected, alongside the team's existing statistical methods. The safety committee decides whether there is a signal.
- 4
Validation and records
EMA expects the marketing authorization holder to validate, monitor and document model performance within its pharmacovigilance system. Logging each accepted or corrected suggestion helps.
What AI does not do
It does not assess causality on its own, does not submit reports to FDA, EMA or any national authority, and does not replace the qualified person responsible for pharmacovigilance. CIOMS warns that automating only for efficiency can remove the value of expert case assessment.
What regulators expect when AI is used
We could not verify an AI-specific pharmacovigilance rule from any Latin American authority. The reference frameworks come from EMA, FDA, CIOMS and ICH.
Responsibility is not delegated to the model
EMA states that the marketing authorization holder must validate, monitor and document model performance and include AI/ML operations in the pharmacovigilance system (adopted September 2024).
EMA reflection paperCredibility for a context of use
FDA's January 2025 draft guidance proposes a risk-based framework to assess the credibility of an AI model for a specific context of use. It is a draft and non-binding.
FDA draft guidanceA human in the loop
CIOMS Working Group XIV published its report on AI in pharmacovigilance on 4 December 2025. It warns that automation aimed only at efficiency can remove the value experts add to each case.
CIOMS Working Group XIVCase criteria do not change
Under ICH E2D, a case needs four minimum elements: an identifiable reporter, an identifiable patient, an adverse reaction and a suspect product. AI does not change that rule.
ICH E2D guidelineAuthorities and reporting channels
Where reports go in each country. Timelines and obligations are those of each authority's current rules; for Mexico, see the NOM-220 guide (in Spanish).
| Authority | Country | What it publishes |
|---|---|---|
| FDA | United States | FAERS, the database of adverse event reports. Official site |
| EMA | European Union | The EU pharmacovigilance framework and good pharmacovigilance practices (GVP). Official site |
| INVIMA | Colombia | Online pharmacovigilance services: VigiFlow eForms to report adverse reactions and enrollment in the national pharmacovigilance network. Official site |
| COFEPRIS | Mexico | Pharmacovigilance guides and guidelines; the binding rule is NOM-220-SSA1-2016, published in the DOF. Official site |
| DIGEMID | Peru | National Pharmacovigilance and Technovigilance Center (CENAFyT), the VigiFlow platform and safety alerts. Official site |
| ANVISA | Brazil | Pharmacovigilance and the VigiMed reporting system, which registration holders use with the ICH E2B(R3) standard. Official site |
| WHO · Uppsala Monitoring Centre | Global | VigiBase, the WHO global database of reports of suspected adverse effects. Official site |
What Arkangel AI does and does not do
Arkangel Insights shows aggregated, anonymized signals from the real clinical questions physicians ask in Colombia, Peru, Mexico and Ecuador, by molecule, specialty, disease area and country. That is medical intelligence, not pharmacovigilance.
What it shows
- What physicians ask about a molecule, a disease area or a specialty, and in which country.
- Recurring doubts and barriers, to prioritize medical education and scientific communication.
- An aggregated view, with no data that identifies patients or physicians.
What it does not do
- It does not receive, record or code individual case reports.
- It does not assess causality or seriousness, and it does not decide what is a safety signal.
- It does not report to FDA, EMA, INVIMA, COFEPRIS, DIGEMID, ANVISA or any other authority.
- It does not replace a company's pharmacovigilance system or its qualified person.
Where physician questions and pharmacovigilance meet
If a frequent question is about an adverse effect or an interaction, Medical Affairs can use it to review materials, training or safety messages. But an aggregated signal is not a case: it has no identifiable patient or reporter, which ICH E2D requires. Any specific safety information that reaches the company through any channel is assessed by its pharmacovigilance team under its own procedures.
Frequently asked questions
What is pharmacovigilance and why does it matter?
WHO defines pharmacovigilance as the science and activities relating to the detection, assessment, understanding and prevention of adverse effects or other medicine-related problems. It lets companies, health professionals and regulators report and analyze adverse reactions and act in time to protect patients, always with a qualified professional assessing each case.
How can AI help with signal detection?
AI can group similar adverse event reports and highlight drug-event combinations that appear more often than expected, alongside the statistical methods safety teams already use. It speeds up triage, but it does not decide: the safety committee assesses whether there is a signal, and the marketing authorization holder documents the model in its system.
What is AI in pharmacovigilance?
It is the use of artificial intelligence models to support pharmacovigilance work: classifying adverse event reports, prioritizing them by seriousness, grouping similar reports to look for patterns, and screening literature. The marketing authorization holder validates the model and documents it in its system, and a professional still assesses each case, its causality and its reporting.
Can an AI report adverse events to FDA or EMA on its own?
Reporting remains the responsibility of the marketing authorization holder and of the professionals each country's rules require to report. AI can prepare or prioritize cases, but an accountable person reviews them and submits them through the official channel. EMA expects AI/ML operations to be documented inside the pharmacovigilance system.
Does Arkangel AI offer a pharmacovigilance system?
No. Arkangel AI does not receive, code or report cases, and it does not assess causality. Arkangel Insights shows aggregated, anonymized physician questions by molecule, specialty, disease area and country. It can guide medical education and scientific communication, but it is not part of a company's pharmacovigilance system and does not replace it.
Do physician questions in Arkangel Insights count as adverse event reports?
No. They are aggregated, anonymized signals with no identifiable patient or reporter, and ICH E2D requires both, plus a reaction and a suspect product, for a case to exist. If the company learns of a specific case through any channel, its pharmacovigilance team assesses it under its own procedures.
What do regulators require to use AI in pharmacovigilance?
EMA asks the marketing authorization holder to validate, monitor and document model performance within its pharmacovigilance system. FDA's 2025 draft guidance proposes assessing a model's credibility for its context of use. CIOMS stresses keeping a human expert in the assessment of each case. None of these sources lets the model decide alone.
Keep reading
See what physicians ask about your molecule
In 30 minutes we review what Arkangel Insights shows for your disease area and country, and what it cannot tell you. Your pharmacovigilance stays in your own system.
Sources
- EMA. Reflection paper on the use of Artificial Intelligence (AI) in the medicinal product lifecycle — EMA/CHMP/CVMP/83833/2023, adopted by CHMP on 9 September 2024.
- FDA. Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products — Draft guidance, January 2025. Non-binding.
- CIOMS Working Group XIV: Artificial Intelligence in Pharmacovigilance — Report published on 4 December 2025.
- ICH E2D. Post-Approval Safety Data Management: Definitions and Standards for Expedited Reporting — Step 4 version of 12 November 2003, sections 3.2 and 4.2.
- WHO. Pharmacovigilance — Definition of pharmacovigilance. Accessed 1 November 2026.
- FDA. FDA Adverse Event Reporting System (FAERS) — United States.
- EMA. Pharmacovigilance: overview — European Union.
- Uppsala Monitoring Centre. About VigiBase — WHO programme.
Last verified: . Informational content for pharmaceutical industry teams. It does not replace each authority's current rules or regulatory advice.