The 90% override figure is real. The easy story built around it is not.
A 2024 systematic review and meta-analysis included 16 studies of responses to drug-drug interaction alerts. Fifteen were included in the quantitative analyses, and 11 contributed to the pooled estimate of physician overrides: 90% (95% CI, 85–95%). That result deserves investigation. It is not proof that nine of every ten alerts were clinically correct, that every override was unsafe, or that the same rate applies to every hospital.
The distinction matters. An override count tells us what happened at the alert. It does not tell us whether the alert deserved to interrupt the prescription.
The pooled estimate hides wide variation
The meta-analysis reported an I² of 100% for the override estimate, indicating extreme variation across the included studies. The pooled 90% figure is therefore a summary of very different systems and settings, not a universal baseline for any one implementation.
The review also did not test Meddies. It did not compare a patient-specific alert strategy with a generic one, and it did not establish that reducing alert volume improves patient outcomes. Those are separate questions.
So the defensible reading is narrow: overrides were common across the published studies, and any prescription-safety system should investigate why. Turning that finding into “doctors ignore safety” blames the user before examining the alert.
An override is a signal, not a verdict
One click cannot tell us whether the prescriber missed a dangerous interaction or dismissed a warning that did not fit the patient. To separate those cases, an evaluation needs the alert’s severity, the patient context available when it fired, the reason for overriding it, and an independent judgment about whether the action was appropriate.
Alert volume and clinical relevance are plausible parts of the mechanism. A system that surfaces many low-value warnings asks clinicians to spend attention repeatedly. But the 90% estimate alone does not prove that volume caused the overrides. It also cannot tell us which filters, thresholds, or presentation choices would improve the result.
This is where rhetoric usually runs ahead of evidence. “Fewer alerts are safer” may be a sensible design hypothesis. It is not a clinical outcome until it has been tested.
What Meddies can claim today
The current Meddies implementation includes a drug-interaction tool that accepts a medication list, resolves the drug names, and returns pairwise interactions with severity labels and descriptions. Its clinical instructions ask the system to check all medication pairs rather than only the pair named in the question, then run a separate safety review before presenting a recommendation.
That is current product behavior in code. It is not evidence that Meddies automatically screens every prescription at the moment of ordering, reduces alert volume, lowers override rates, or prevents medication harm. We have not published those measurements.
Our design direction is to use the medication list and available patient context to decide which findings deserve attention, then show the basis for the result. The boundary remains important: a design direction is a claim about what we are building, not proof that the design works in clinical use.
The test is still ahead
To evaluate a relevance-first alert strategy, we would need a defined comparison. The same prescribing cases should run through a baseline configuration and the proposed filters. The evaluation should report alerts per order, overrides by severity, reasons for overriding, independently reviewed appropriateness, and the errors or delays introduced by either configuration.
Only then could we say whether fewer alerts improved the signal rather than merely hiding warnings.
For now, the 90% estimate gives us a problem worth measuring, not a slogan. The work is to make each interruption justify itself and then test whether clinicians agree.