Manuel Rios, MD
Reltronic, Inc.
Clinical decision support alerts are overridden between 46 and 96 percent of the time. That figure is usually presented as evidence of a design problem, and it is one. But it is first a safety problem, because the alert that would have mattered arrives through the same channel as the ninety that did not, and it is dismissed by the same reflex.
The number is not a satisfaction score
A systematic review of alert fatigue measurement published in 2026 found average override rates across studies ranging from 46.2 to 96.2 percent. Severity does not protect against this. One study within that literature found that 88.2 percent of alerts classified as very severe drug-drug interactions were overridden. The alerts a system flags as most urgent are dismissed at close to the same rate as the rest.
It is tempting to read those numbers as a workforce problem, as evidence that clinicians are not paying sufficient attention. The literature does not support that reading. When overridden alerts are independently reviewed for appropriateness, a large share of the overrides turn out to be clinically correct. Reported appropriateness ranges from 63.4 to 100 percent for drug-allergy alerts and from 27 to 87.5 percent for renal alerts. In many categories, the clinician who dismissed the alert was right to dismiss it.
That reframes the problem entirely. Override behavior is not inattention. It is a rational response to a channel with a low base rate of useful signal. A clinician who has learned that most of what arrives in a given channel does not require action has learned something true about that channel, and has adapted correctly.
Desensitization is dose-dependent
The mechanism has been characterized. Ancker and colleagues studied 112 ambulatory primary care clinicians over a three and a half year period and found that clinicians became measurably less likely to accept an alert as the number of alerts they received rose. The effect was strongest for repeated alerts on the same patient.
The implication is uncomfortable for anyone building these systems. Alert volume is not a neutral quantity that a well-designed interface can absorb. Volume is itself the mechanism of harm. Every low-value alert added to a system degrades the response to every alert already in it, including the ones that were working. A system that adds coverage without regard to specificity is not neutral. It is actively spending down a finite resource, and the resource is clinical attention.
This is why the framing matters. Filed as a usability complaint, alert fatigue is a matter of preference, to be traded off against feature completeness and addressed at some later point in the roadmap. Filed correctly, as a patient safety problem, it becomes a design constraint that binds from the beginning. Overrides have been associated with medication errors and serious adverse events, including death, when clinically important information was inadvertently passed over.
What follows for the design of these systems
If clinical attention is treated as a fixed and exhaustible resource, several design positions follow that are otherwise easy to argue against.
- Specificity is the binding constraint, not coverage. A system that surfaces fewer findings at higher positive predictive value is not a less capable system. It is a system that has made the correct trade.
- Volume warrants an explicit budget. If every additional alert degrades response to the existing ones, then adding a new rule should require displacing an existing one, or demonstrating that the new rule clears a threshold the current set does not.
- Within-patient repetition deserves particular scrutiny, since it is the pattern most strongly associated with desensitization in the published work.
- Interruption should be the exception rather than the default. Information that a clinician can consult when it is relevant occupies a different cognitive category from information that stops work to be dismissed.
- The reasoning should travel with the finding. A statement that can be checked against its supporting observations in a few seconds is a different object from a conclusion that must be either trusted or dismissed. Only the first kind can be evaluated rather than pattern-matched.
The measurement problem underneath
Most institutional reporting on decision support counts alerts fired and rules deployed. Neither quantity describes whether the system is helping. The systematic review noted that alert fatigue measurement is not consistently defined across the literature, which means institutions are frequently comparing figures that are not comparable.
Two measures are more informative than volume. The first is the acceptance rate over time, which shows whether the system is being adapted to or attended to. The second is positive predictive value by alert category, which shows where a system is earning attention and where it is spending it. A vendor that can report the second figure has confronted the problem. A vendor that reports coverage instead has not.
The question worth asking
Institutions evaluating any clinical decision support system, including systems that describe themselves as using artificial intelligence, are entitled to ask one question before any other. Not what the system can detect, but what proportion of what it surfaces a clinician acts on, and how that proportion has moved over the period it has been running.
A system whose acceptance rate is falling is not being adopted. It is being tuned out, and the interval between those two states is where the risk sits.
References
- Alert fatigue measurement in clinical decision support: a systematic review. PubMed 42148822.
- Appropriateness of Alerts and Physicians’ Responses With a Medication-Related Clinical Decision Support System: Retrospective Observational Study. JMIR Medical Informatics, 2022.
- Ancker JS, Edwards A, Nosal S, Hauser D, Mauer E, Kaushal R. Effects of workload, work complexity, and repeated alerts on alert fatigue in a clinical decision support system. BMC Medical Informatics and Decision Making, 2017; 17:36.
Disclosure: the author is affiliated with Reltronic, Inc. This paper makes no claim about any Reltronic product and describes no proprietary method. Every figure cited is drawn from the published literature listed above.
Leave a Reply