Nearly 99% of drug-allergy alerts get overridden. Most were never a real match

Every hospital's electronic health record is wired with clinical decision support: automated pop-ups that warn a prescriber when an order looks unsafe. One of the commonest is the drug-allergy alert, which fires when a prescribed drug appears to conflict with an allergy in the patient's record. A retrospective study in JAMIA Open, published on 27 January, examined a full year of those alerts at one large US medical centre and found that 98.9% of them were overridden — and that the great majority were never a true match to a documented allergy in the first place [s1].

A year of pop-ups

The researchers analysed every drug-allergy alert fired to prescribers in 2023 at a large academic medical centre in the Southeast [s1]. There were 101,492 of them, triggered for 9,111 unique patients — an average of 11.1 alerts per patient — and 98.9% were overridden [s1].

The more revealing number is what set the alerts off. Only 9.7% of the alerts were a definite match between the prescribed drug and the documented allergen [s1]. The remaining 90.3% fired for a different drug that merely shared a class or allergen group with something the patient had once reacted to [s1]. And 70% of all the alerts were triggered for patients who had already been given the very drug that triggered the alert [s1]. Of that group, 74% had received the drug after their allergy was documented — 52% of all alerts — and 79% went on to receive the same drug again after 2023 — 56% of all alerts [s1].

Read plainly, this describes a system warning clinicians away from drugs their own records show the patient has already tolerated. When the researchers compared override behaviour, alerts for patients who had previously received the drug and were a definite match were slightly more likely to be overridden than no-match alerts (odds ratio 1.18, 95% CI 1.03–1.33, P = .013), with override rates of 98.9% (71,357 alerts) versus 98.8% (30,135) [s1]. The differences are marginal because the override rate is close to 100% across the board — which is the whole problem.

Why "override everything" is dangerous

An alert that is dismissed 99 times out of 100 is not a safety net; it is noise that a clinician learns to click past. The risk is that the hundredth alert — the one that actually matters — is dismissed with the same reflex. This is alert fatigue, and the study's own framing is that the fix is not to switch the alerts off but to make them smarter: incorporating "tolerance assertions" — evidence from the record that a patient has safely taken a drug — into the alerting logic so that false positives are suppressed "without compromising safety" [s1]. That is precisely the kind of rule-refinement, and increasingly machine-learning-based filtering, that vendors are now selling as the next generation of decision support.

The catch is that the field cannot yet agree on how to tell whether such a change works. A systematic review of systematic reviews in the Journal of the American Medical Informatics Association, published in May, set out to find how alert fatigue is actually defined and measured [s2]. Across 22 included reviews, studies reported between one and 11 different alert metrics, and only a single article offered an operational definition of alert fatigue at all [s2]. The most common metrics were simply the quantity of alerts, the override rate and the acceptance rate [s2].

Those metrics are easy to game. A 90% override rate can be driven down by suppressing a category of low-value alerts, or driven up by an interface that makes overriding harder — neither of which tells you whether a clinician is paying more attention to the alert that counts. The review's recommendation is to define alert fatigue as "a significant, sustained decrease in appropriate alert response rates from an established baseline" — a harder thing to measure, and a much more honest one [s2].

What it means for a reader

Most patients never see these pop-ups, but they are shaped by them. An over-firing allergy alert can lead a clinician to reach for a second-choice antibiotic the patient does not actually need to avoid; a well-tuned one can catch a genuine cross-reactivity before it reaches the pharmacy. The evidence here is that the current generation of alerts is heavily weighted toward the first failure mode, firing overwhelmingly on class-level rather than exact matches, and on drugs the patient has already tolerated [s1].

It also sets a bar for the AI tools now being marketed to fix this. A smarter alerting engine is only an improvement if it is judged against sustained, appropriate clinician response — not against a lower raw override count [s2]. That is the same evidence gap that runs through the rest of hospital decision support, from sepsis-prediction alarms measured on discrimination rather than on how nurses respond to the FDA's own guidance on which decision-support software it will regulate. It is also the harder companion to the work on predicting which patients will not take their medicines at all and on safely stopping drugs older patients no longer need: getting the right warning to the right prescriber is only useful if the warning is still believed.

What to watch

Whether health systems that add tolerance logic or machine-learning filters to their alerts report the harder outcome — a sustained rise in appropriate responses to the alerts that remain — rather than a headline drop in override rates. The first is evidence of a safer system; the second can be achieved by making the noise quieter without making the signal louder.

Sources

1. [s1] Identifying false-positive drug allergy alerts based on drug tolerance assertions: a retrospective study. JAMIA Open, 27 January 2026. https://doi.org/10.1093/jamiaopen/ooag010
2. [s2] Alert fatigue measurement in clinical decision support: a systematic review. Journal of the American Medical Informatics Association, 18 May 2026. https://doi.org/10.1093/jamia/ocag064