A safety alert that gets dismissed nine times out of ten is no longer a safety control. It is a speed bump clinicians have learned to drive over, and the result is that the one warning that should have stopped an order gets the same reflex click as the dozen that should never have fired.
This is alert fatigue. Doctors stop reading drug-interaction warnings not because they are careless, but because most of the warnings carry no value for the patient in front of them. The prescription-safety problem in most systems is not too few alerts. It is too many low-value ones, and the cost of that excess is paid by the rare alert that mattered.
Override is the default, not the exception
A 2024 systematic review and meta-analysis by Felisberto et al. in Health Informatics Journal (PubMed 38899788) found the overall prevalence of drug-drug interaction alert override by physicians was 90 percent (95% CI, 85 to 95 percent). Roughly nine alerts dismissed for every ten shown.
The number matters because it tells you override is not a rare edge case. It is the modal response. When a system fires an interaction warning, the prescriber's typical move is to dismiss it and continue.
A control with that dismissal rate has stopped controlling anything. The clinician is no longer reading alerts. They are clearing a queue.
More alerts make prescribing less safe
The cause is a chain, not a single failing. It starts with volume. Many systems alert on every theoretical drug pairing in the literature, regardless of dose, route, or whether the combination is relevant for this specific patient. High volume plus low patient-relevance produces overload. The prescriber cannot weigh each warning on its merits in the seconds available during an order, so they adopt a faster rule. Dismiss.
Once dismissal becomes the habit, the system loses the ability to separate a routine pairing from a dangerous one. The severe interaction that warranted a stop arrives inside the same stream as a dozen low-value ones and gets the same click. The meta-analysis above measured the override rate across all such alerts. The practical worry is that the habit of dismissal does not pause for the warnings that deserve attention.
So adding more alerts does not make prescribing safer. Past a point it makes prescribing less safe. Each extra low-value warning reinforces the dismissal habit and lowers the odds that the important one is read.
Relevance is the safety mechanism
This is our design stance at Meddies, stated as intent rather than a measured outcome. We have not produced an override-reduction figure, and we do not claim one.
The screening runs at the point of prescribing, inside the hospital EMR, while the prescription is being written. Timing is part of the design. A warning that arrives after the order is placed comes too late to change the decision cheaply.
What the screening aims to surface is the interaction that changes this patient's decision, not every theoretical pairing in the reference set. A warning earns the doctor's attention when it depends on the patient's current medications, the dose, and the clinical context. The mere existence of a documented pairing somewhere in the literature is not enough. Fewer warnings, each one load-bearing, is the goal.
Each alert is paired with its source. The doctor sees the basis for the warning and can verify it in seconds rather than dismissing it on reflex. A sourced alert is checkable. An unsourced one can only be trusted or ignored, and under time pressure it gets ignored. Showing the evidence is what lets a clinician treat a warning as a claim to evaluate rather than noise to clear.
The bet underneath all of this is simple. A smaller set of patient-specific, sourced alerts is more likely to be read than a larger set of generic ones, and an alert that is read is the only kind that can prevent harm.
What we are claiming, and what we are not
We frame relevance-first screening as our design intent. We are describing how the system is built to behave and the reasoning behind it, not reporting a clinical trial result. The 90 percent override figure comes from the published literature and describes the general problem these systems face. It is not a measurement of Meddies.
The conclusion is narrow, and we think it is correct. Fewer, sourced, patient-specific alerts beat more alerts. The failure mode of prescription safety screening is not silence. It is noise that teaches the doctor to stop reading.
So the work ahead is to earn each warning a read. Relevance is how that read gets earned, and it is where a screening system either becomes a control again or stays a speed bump.