Reading FDA Adverse Event Signals: From FAERS to AEMS

A dashboard count of adverse event reports answers a narrow question: how many reports matching a search are present in the system. It does not tell you how many people used the product, whether the product caused an event, or whether one product is safer than another. Disproportionality analysis extracts useful early warnings from this imperfect stream, but its output must be followed by clinical and epidemiologic investigation.

The system changed, but the interpretation did not#

FAERS, the FDA Adverse Event Reporting System, became familiar as the public database for drug and therapeutic-biologic safety reports. On March 11, 2026, FDA launched the broader Adverse Event Monitoring System, AEMS, to unify reporting across regulated product categories and migrate historical data from legacy systems, and you will therefore meet both names in articles, files, and dashboard documentation.

The transition changes access and infrastructure, not the basic evidentiary status of spontaneous reports; the AEMS page repeats the central limitations: a report does not establish causation, information may be incomplete or unverified, duplicates exist, and reporting data cannot establish occurrence rates.

When older studies say "FAERS analysis," they refer to the drug-reporting corpus and workflow available at that time. When current public pages say "AEMS," they describe the newer consolidated platform. The historical label should not be mistaken for a different type of causal evidence.

What enters the database#

Patients, family members, health professionals, manufacturers, and others can report a suspected adverse event or medication error. Manufacturers have regulatory duties to forward qualifying reports they receive, while direct public and professional reporting is generally voluntary.

An individual case safety report may include the product, event terms, and timing. It may include patient characteristics, outcome, and reporter type. It may include other medicines and a narrative. Completeness varies widely. The presence of a product as "suspect" reflects the reporter's suspicion, not FDA confirmation, and a report may describe an event caused by the underlying illness, another medicine, a drug interaction, or coincidence.

The same case may arrive through more than one route, such as a patient and a manufacturer. Follow-up information can also produce linked versions, and analysts use case identifiers and deduplication rules, but public counts should not be assumed to represent unique people without checking how records were handled.

Why raw counts are not rates#

An incidence rate requires a numerator and a denominator over time. The reporting system has a numerator-like count, although incomplete and duplicated, but it lacks a reliable denominator showing how many people used each product, at what dose, for how long, and with what clinical characteristics.

Ten thousand reports for a medicine taken by tens of millions of people cannot be compared directly with one thousand reports for a rarely used medicine. Reporting probability also differs. New products receive attention, serious events are more likely to be reported, and familiar events may be ignored, so a count can rise after a warning, news story, lawsuit, or social-media discussion even when the biological event rate is unchanged. That stimulated reporting is one reason a time-series spike is not self-interpreting, and underreporting is why the absence of many reports does not prove an event is absent.

What disproportionality measures#

Disproportionality methods ask whether a particular product-event pair occupies a larger share of reports than expected from the rest of the database. A reporting odds ratio, for example, compares the odds that reports mentioning one product contain the event with the corresponding odds among reports for other products. Proportional reporting ratios use a related comparison. Empirical Bayesian approaches shrink unstable estimates, especially when counts are small.

These calculations are useful for ranking combinations that deserve review. They are not ordinary clinical risk ratios. Both numerator and comparison come from a selected reporting system, not from complete populations of treated and untreated people.

The reference set also matters. If an event is commonly reported for many products in the same class, the contrast may shrink even when the event is real. If competing events dominate one product's reports, another event's share can appear lower. Product indication, age, co-medications, media attention, and reporting practices can all shape the statistic.

A signal is a question with structure#

A useful signal is not merely "many reports." It is a product-event pattern that is new, unusual, clinically serious, biologically plausible, or supported by informative cases. Disproportionality can identify that pattern; then you ask:

• Did the event begin after the product was started?
• Did it improve after withdrawal or recur after restart?
• Are there plausible alternative causes?
• Do cases share a distinctive clinical pattern?
• Is the event already part of the treated disease?
• Does pharmacology or mechanism support the relationship?
• Is the pattern present in trials, registries, claims, electronic records, or other countries' systems?

The answers may strengthen, weaken, or refine the signal. A rare, characteristic event with a convincing time course can be informative even with few reports, while a large collection of vague, highly confounded reports may remain difficult to interpret.

What FDA may do after detection#

FDA posts quarterly lists of new safety information or potential signals of serious risks identified through surveillance, but appearance on such a list does not mean the agency has concluded that the product caused the risk. It means evaluation is occurring or new information has been identified.

Further review can lead to no action, enhanced monitoring, or requests for analyses or studies. It can lead to changes to prescribing information, a safety communication, risk-management measures, or other regulatory action. The outcome depends on the total evidence and the product's benefits as well as its risks.

You therefore have to look beyond the initial list. Later communications, label revisions, and completed safety evaluations reveal what the investigation concluded. A signal's regulatory history is more informative than the dashboard statistic alone.

How to audit a dashboard claim#

When a paper or social-media post cites FAERS or AEMS, ask seven questions:

1. Is the result a raw count, a reporting proportion, or a disproportionality statistic?
2. Were duplicate and follow-up reports addressed?
3. How were product names, event terms, dates, and suspect roles defined?
4. Was a sensible comparator chosen, including products used for similar conditions?
5. Did authors account for indication, co-medication, publicity, and time on market?
6. Are they claiming incidence, relative risk, or causation that the data cannot provide?
7. Was the signal checked against case narratives or an external data source?

A polished forest plot or large reporting odds ratio does not answer these design questions. Small counts can create unstable estimates, and multiple searches can surface chance signals unless methods and hypotheses are declared clearly.

Why spontaneous reporting still matters#

Randomized trials are usually too small and too short to detect every rare, delayed, or context-specific harm. They may exclude people with complex illness or many co-medications. Once a product is used broadly, spontaneous reports provide wide surveillance and can surface unusual clinical patterns quickly.

Their value comes from sensitivity and timeliness, not from precise risk estimation. They function like a smoke detector: an alarm justifies checking for fire, but the alarm does not measure the size or prove the source. Other data systems supply denominators and comparators for that next stage.