Any safety system that is ignored 96 percent of the time is not an effective safety system. It is only a record that the alert fired, and the clinician did not act.
Override rates on drug safety alerts consistently run between 49 percent and 96 percent, depending on the alert and the setting. This fact has been repeated and has strong data. The usual reading of those numbers is that physicians have become desensitized or are not vigilant. The more accurate reading is that most interaction alerts in commercial systems carry low or negligible clinical significance, so for every alert that matters, twenty or more meaningless ones fire alongside it. I often must argue that dismissing the entire lot is not inherently a lapse in vigilance. It is a reasonable adaptation to a system that has cried wolf until the listener stopped listening.
The clinical consequence is not just the time lost to clicking. It is the rare alert that would have mattered, arriving in the same visual channel as the twenty that did not, and being cleared with them.
Most organizations that ask me to look at their decision support present with the same complaints. Override rates are high, clinicians are vocal, quality metrics are not moving, and safety events have informatics root causes. The decision support library has usually grown for years, one addition at a time, each one added for a regulation, a quality measure, a payer, or a committee, and almost none ever removed. No specialty in medicine would accept a formulary managed that way.
Real physician voices, twice a week
Free, and one click to unsubscribe.
So, I manage it the way a formulary is managed. I manage it as a clinical decision. Five retirement criteria do most of the work, and every one of them can be applied by a committee that already meets.
- Retire any alert whose override rate exceeds 90 percent and cannot be redesigned below it. Dismissed nine times out of ten, it is not providing safety.
- Retire any alert whose underlying guideline has been superseded and not replaced within the review window. Hint: That requires a review window that rarely exists.
- Retire any alert that fires outside the workflow and cannot be moved into it.
- Retire any alert that reports findings without strong recommendations.
- Retire any alert that no one will defend by name. Put it on a screen in a governance meeting and ask who owns it. Silence means it has no clinical sponsor.
The last criterion is the one that changes a clinical decision support library, and it is the one most committees have no procedure for. An alert no one will claim in a meeting is still firing on every one of your colleagues, several times a shift.
What replaces the retired rule is not nothing. Decision support done as clinical work follows the same discipline as any other clinical problem. The history: What is the library currently doing? The examination: Which alerts fire, how often, with what override rate, in which contexts? The differential: Which failure is operative, and is it more than one? The plan: What gets deactivated, redesigned, or built? The follow-up: Re-measure the override rates, the near misses, the clinical outcomes, and what the clinicians say. A rebuild worth doing looks like this: Reduce the sepsis bundle override rate from 96 percent to 65 percent over six months, with composite scoring and an integrated order set, governed by the sepsis committee, owned by a named physician with named analyst support, reviewed every two weeks, with discontinuation criteria written down in advance.
Notice what that plan costs. The library already exists. The engine is already licensed. The alerts are already firing. What is missing is a chartered clinical governance committee and a physician with the authority to turn things off. The price of admission is the smallest in the formulary.
Two questions are worth asking out loud at your next medical staff or informatics meeting, and the answers already exist in your electronic health record (EHR) system. What is our alert override rate, by alert and in aggregate, and who reports it? Which alert did we last retire, and when?
An organization that cannot name the last thing it turned off has a library that only grows. Your attention is the resource it consumes, and attention is the one input in your clinic that should never be purchased.
Thomas Powell is a physician executive, a career chief medical information officer (CMIO), and the founder of CMIO, LLC, which provides interim physician informatics leadership and advisory work to health systems. He has held the CMIO seat in academic, community, pediatric, and integrated delivery systems, and in most of them he was the first.
At the University of Arkansas for Medical Sciences (UAMS), he was CMIO and associate vice chancellor and led the move to Epic across the hospital and 43 clinical groups. At Aspirus Health, he was vice president and CMIO of an 18-hospital system with 1,300 employed physicians and advanced practice clinicians, and at Miami Children’s Hospital, now Nicklaus Children’s, he took 38 pediatric subspecialties from paper to Cerner. He served the State of Arkansas as medical and executive officer of the Office of Health Information Technology and the state health information exchange.
He trained at Duke University as a National Library of Medicine fellow in clinical informatics under Ed Hammond, and at UAMS he created Arkansas’s first clinical informatics master’s and doctoral program and helped build the state’s only clinical informatics board fellowship. He writes and speaks on applied clinical informatics, the governance of clinical AI, electronic health record usability, and the physician executive role.
He is the author of The Diagnostician: American Healthcare Is Sick, Misdiagnosed, and Getting Worse; It’s Time to Call the Doctor, forthcoming November 3, 2026, which applies the clinical diagnostic method to the health care organization itself. He shares updates on X.
