Signal Detection: From Disproportionality to a Label Change

A safety signal is information suggesting a new potentially causal association, or a new aspect of a known one, that warrants further investigation.

The definition matters. A signal is not a conclusion, and treating it as one in either direction, dismissing it or acting on it as fact, is the failure mode at both ends.

The statistical screen

Most systematic detection starts with disproportionality analysis on a spontaneous reporting database: FAERS, EudraVigilance, VigiBase.

The question it asks is narrow. For a given drug and event, is this pair reported more often than expected relative to everything else in the database? Common measures are the proportional reporting ratio (PRR), the reporting odds ratio (ROR), and Bayesian approaches such as the information component or empirical Bayes geometric mean.

What disproportionality is not:

• Not incidence. Spontaneous databases have no denominator. You cannot compute a rate from them.
• Not causality. Disproportionate reporting is a reporting pattern.
• Not immune to bias. Notoriety bias, where publicity drives reporting, and the Weber effect, where reporting peaks in early marketing years, both generate disproportionality with no underlying change in risk.

It is a screen for prioritisation. That is all, and it is genuinely useful as that.

Signals do not only come from statistics

• Individual case review, especially well-documented cases with positive dechallenge and rechallenge.
• Literature monitoring, which is a standing obligation.
• Clinical trial data from ongoing studies.
• Regulatory communications about the class from other authorities.
• Non-interventional studies and registries.

A single well-documented case of a rare, serious event with no other explanation can be a stronger signal than a large disproportionality score on a common event.

Validation, then evaluation

Validation asks whether there is enough evidence to justify further analysis: is the association plausible, is it already known and in the label, are the cases real and distinct.

Evaluation is the assessment proper: all available data, biological plausibility, temporal relationship, dose relationship, alternative explanations, and what it means for benefit-risk.

The outcome is a decision, and the possible decisions include doing nothing with a documented rationale. Each is recorded, and the record appears in the next periodic report.

Where it leads

A confirmed signal that changes the benefit-risk picture leads to a core data sheet change, then local label variations in every market, then implementation tracking. That chain is the reason signal management is a regulatory function and not only a pharmacovigilance one.

For serious findings the chain also runs the other way: expedited reporting obligations, IND safety reports for products still in development, and direct healthcare professional communications.