What is a safety signal?
A safety signal is a pattern in adverse event data that suggests a device may be associated with a new or increasing risk. In practice it takes a few recognizable shapes: a sudden spike in reporting, a gradual trend that builds over months, a device reporting far more than comparable peers, or a failure mode that has never appeared before.
A signal is a hypothesis, not a verdict. It says "this is worth a human look," nothing more. The job of detection is to generate good hypotheses and suppress bad ones.
Why raw counts fail
The instinct is to count reports and watch the number. The problem is that the number has no denominator. A device that generated forty reports last quarter and eighty this quarter has not necessarily become less safe. It may simply be selling twice as much, or have caught a wave of publicity that prompted reporting, or be the subject of a recall that drove a filing surge.
Counts also duplicate. One real event can generate an initial report, several supplements, and a final report. Without deduplication, routine paperwork looks like a cluster. These distortions are covered in detail in our guide to the MAUDE database. Every serious detection method exists to get around the missing denominator.
The detection methods that work
No single method is sufficient, so practitioners run several and weigh the results. The common families:
- Spike detection. Flags a sudden jump in reporting volume against a recent baseline, using counts-appropriate statistics so a low-volume device is not held to the same yardstick as a high-volume one.
- Trend detection. Catches a slow, sustained rise that no single week would trip, using rank-based tests that resist noise.
- Change-point detection. Identifies the moment a device's reporting behavior shifts from one regime to another.
- New-problem detection. Surfaces failure modes that do not match a device's historical pattern, which is where the earliest signals often hide.
- Peer-outlier detection. Compares a device against others in its class so a product reporting disproportionately stands out, even without an exposure denominator.
- Disproportionality analysis. A standard pharmacovigilance technique that flags device-event pairs occurring more often than the rest of the database would predict.
Thresholds should adapt to a device's maturity: a newly approved product with a handful of reports needs different sensitivity than an established one generating hundreds. A fixed threshold either drowns you in false alarms on big devices or misses real signals on new ones.
Statistical signals vs clinical signals
A statistical signal is a number crossing a line. A clinical signal is a plausible mechanism by which a device causes harm. The two are not the same, and conflating them is a common failure. Good detection narrows thousands of reports down to a short list of statistical signals; a clinician or engineer then decides which ones describe a real problem. The statistics earn their keep by making that human review tractable, not by replacing it.
From detection to action
A detected signal is the start of a workflow, not the end. The next steps are reading the underlying narratives, checking whether the pattern holds across related product codes, comparing against peers, and deciding whether it warrants escalation. The value of automated detection is that it delivers a ranked, deduplicated short list so analysts spend their time on judgment instead of triage.
For a candid look at where conventional surveillance breaks down, read why most signal detection is broken. To see signal data on real cardiac devices, browse the free MAUDE lookup or read about post-market surveillance more broadly.