Why is a report count not a risk rate?
Records in the FDA dashboard, formerly FAERS and now AEMS, may include incomplete or duplicate reports. An entry does not establish medical confirmation or causation by the suspected medicine. FDA also explains that these reports cannot establish how often an event occurs. Declaring a medicine more dangerous solely because it has more reports is therefore an unsound interpretation of these data.
Research example: what task does AI perform?
In a 2022 French study, models pre-coded 11,633 adverse-reaction reports. The task was extracting coded information and assessing seriousness, rather than proving that a medicine caused harm.
How does a signal lead to a decision?
EMA distinguishes a safety signal from proof of causation. Assessment may lead to requests for information, continued monitoring or regulatory action such as updated product information.
Editorial assessment: define the task before the model
We suggest specifying the output instead of using a broad label such as “medicine safety AI”: does the system extract a term from text, rank potentially serious reports or suggest a pattern for investigation? Acceptance testing should match that task. Success at coding information should not be presented as success at assessing causality.
What should an institutional pilot measure?
Our editorial recommendation is to record missed important reports, unnecessary alerts and expert corrections together on a separate test set. Separating results by language, report source and missing information, and retaining the model version and review date, can make the assessment traceable. These recommendations do not imply that the cited study validated every language or institution.
Research, workflow and regulatory permission are separate questions
A positive research result alone does not establish safe use in a particular institution or regulatory permission for a product. This article is not intended to guide personal medication decisions. It makes no claim of clinical effectiveness or permission for RAI products. Institutional evaluation should consider intended use, human review and product-specific documentation separately.
Sources
- FDA — FDA Adverse Event Monitoring System (AEMS) Public Dashboard ↗
Accessed: 2026-10-28 - Martin et al. — Validation of Artificial Intelligence to Support the Automatic Coding of Patient Adverse Drug Reaction Reports, Using Nationwide Pharmacovigilance Data (Drug Safety, 2022) ↗
Accessed: 2026-10-28 - European Medicines Agency — Signal management ↗
Accessed: 2026-10-28
This article is for education and information. Referencing an organization does not imply its endorsement of RAI Health Pro. This is not a claim of clinical effectiveness for a product or model.