The Future of Clinical Decision Support Is Silence
We’ve been ignoring the lesson for twenty years. Now the stakes are higher.
The Future of Clinical Decision Support is a new series from MDCalc exploring how AI, evidence, and physician judgment are reshaping medicine. These essays are intended to spark discussion about where technology is taking clinical decision support – and where physicians feel it should go.
Blaring “Mary Had a Little Lamb” every 4–19 minutes from 2am to 4am in a fluorescent-lit ER has only two outcomes: insanity, or alert fatigue. The tune is permanently burned into my ears; it’s the bed alarm song used in my ED when a patient at risk of falling tries to climb out of bed.
Alert fatigue doesn’t happen because physicians and nurses can’t pay attention. It happens because our tools are so figuratively (and sometimes literally) noisy.
I’m a clinical informaticist who’s worked on Epic configuration for a decade, and this isn’t a story about good guys and bad guys. It almost never is in healthcare. Most alerts exist for perfectly defensible reasons: someone wants to prevent a medication error, catch a dangerous interaction, or make sure an important diagnosis isn’t missed. Look at each intervention individually and it can seem irresponsible not to build it.
The problem appears in aggregate. When every reasonable concern becomes another interruption, clinicians have to distinguish the handful of alerts that matter from hundreds that don’t. Consider a drug-interaction warning I’ve seen dozens of times for two medications I prescribe together routinely. I understand the interaction. I’ve already accounted for it. I override the alert. When the same warning appears tomorrow, I override it again. Eventually, the system hasn’t taught me to prescribe more safely; it has taught me that clicking “override” is one of the steps necessary to prescribing this medication.
That conditioning matters. Drug-interaction alert override rates have been reported above 90%, which is often described as “alert fatigue.” But I think that phrase puts the emphasis in the wrong place. The problem isn’t that clinicians got tired of useful warnings. We trained healthcare workers, through repetition, that the warnings usually aren't useful. If "clicking through" is the rational response to the first ninety-nine alerts, we shouldn’t be surprised when the hundredth that that actually matters gets the same treatment.
That’s how a well-intentioned tool-design problem becomes a patient-safety problem. But here's where GenAI could make things so much worse: we can reproduce that Pavlovian click response at a scale that traditional clinical decision support never could.
The AI version of the same mistake
Generative AI clinical tools are walking straight toward this failure mode. Traditional clinical decision support generally requires someone to define a rule and build an alert. Generative AI can produce a recommendation, warning, differential, caveat, or explanation about basically anything. The question is no longer whether we can provide decision support at any given moment.
(We can.)
The harder question is whether we should.
Unfortunately, nearly every incentive pushes toward more. Consumer software taught an entire generation of product teams that engagement equals success: more sessions means more clicks and more time in product. Bring that mindset into AI-assisted decision support and you get more recommendations, more prompts, more opportunities to surface the product. Frequent use looks like value on a dashboard. It looks good in a pitch deck. Everyone is happy.
Except, potentially, the doctor trying to take care of the patient.
News flash, buddy: most decisions in medicine don’t need decision support. A well-trained physician managing an uncomplicated presentation usually has the experience to act without help. It's why we do residency. Not every patient with chest pain needs a reminder that pulmonary embolism is a disease entity. Not every infection needs an AI-generated differential diagnosis. Not every medication order needs a warning explaining the adverse effects of a medication I’ve prescribed five hundred times. The AI can be perfectly correct about all of those things and still be completely useless.
Worse, inserting a tool into the decisions clinicians already know how to make may have a cost. If we increasingly rely on prompts for judgments that we previously made confidently, we risk trading competence for convenience: the more routine judgment we outsource, the less practice we get exercising it—and the harder it may become to recognize when the tool is wrong.
The safety net shouldn’t become the trapeze.
MDCalc’s premise has always been the opposite: surface the right tool for the right moment, and stay invisible the rest of the time. Not every presentation needs a Wells' Score. The clinician’s judgment about when to reach for a tool is itself part of clinical reasoning. A calculator that muscles into a decision the physician already knows cold isn’t decision support; it’s noise with safety branding.
AI makes this distinction even more important because generative AI always has something to say.
That doesn’t mean it has something worth saying.
The question isn’t whether the AI is right
Consider a healthy 30-year-old with an ankle sprain. If I order ibuprofen, I don’t need AI interrupting me to explain that NSAIDs can cause kidney injury. But if that patient’s creatinine doubled this morning and I missed it? Yes! Interrupt me!
I don’t need a chatbot reminding me that redness in a patient with straightforward cellulitis can theoretically be caused by vasculitis. But if I’m about to discharge someone whose vital signs and labs suggest evolving sepsis? Please, for the love of God, speak up.
The same principle is obvious outside medicine. Imagine if your car announced every thirty seconds that you were driving within your lane, your speed was appropriate, and no collision had been detected. All perfectly accurate. You’d turn the thing off before you got to the freeway. But if you’re about to merge into a car in your blind spot, you absolutely want it to beep.
That’s the distinction clinical AI needs to learn. The goal isn’t to maximize how often the system is correct. It’s to maximize how often the system contributes something that matters. “Consider sepsis” adds nothing if I’ve already recognized sepsis and started treatment. “This patient’s lactate has doubled since you last reviewed the chart” might. “Consider pulmonary embolism” in every patient with chest pain is technically defensible. Recognizing that this patient has findings that make pulmonary embolism substantially more likely than my current plan suggests could change care.
So the most important question for AI-assisted clinical decision support isn’t simply, “Was the AI right?” It’s: Did the system identify something clinically important that the clinician hadn’t already accounted for—and was it important enough to interrupt them?
That’s a much harder standard than engagement, accuracy, or even acceptance rate. It also recognizes something we’ve spent twenty years pretending isn’t true: every interruption has a cost. Clinician attention is finite, and every unnecessary interruption spends some of that attention—and some of the system’s credibility. The more often you tell me something I already know, the less carefully I’m listening when you finally tell me something I don’t.
More intelligence, fewer interruptions
None of this means AI should watch fewer clinical decisions. Quite the opposite. I want AI silently examining medications, labs, vital signs, imaging, notes, trajectories, and the decisions we’re making. I want it capable of noticing relationships I missed at hour nine of a ten-hour shift.
I just don’t want it narrating everything it notices.
That distinction — between monitoring and interrupting — is where I think much of the future of clinical decision support lies. The best systems may perform far more analysis than today’s CDS while producing far fewer interruptions. They should recognize that I’ve already considered the drug interaction, already adjusted for the renal function, already evaluated the pulmonary embolism—and therefore have nothing useful to add.
Silence isn’t the absence of clinical decision support. Sometimes silence is evidence that the decision support actually understood what was happening.
We don’t need fewer safety nets. We need safety nets that pull taut when something is actually wrong instead of narrating every step on the way down. A system that silently evaluates thousands of signals and speaks only at the moment of meaningful disagreement is fundamentally different from one that comments on every decision.
The future of CDS isn’t less intelligence. It’s more intelligence producing fewer interruptions.
Where this goes next
As LLMs become embedded throughout clinical workflows, like ambient documentation, order entry, differential generation, and patient messaging, the amount of AI-generated clinical content physicians encounter is going to explode. If we apply the instincts of traditional CDS to that new capability, we’ll annotate, flag, recommend, and caveat our way straight into Alert Fatigue 2.0. And physicians will respond exactly as they did the first time: we’ll figure out which button makes the interruption disappear and develop the muscle memory to press it.
“Whatever you say, AI," we'll mutter under our breaths, rolling our eyes, and clicking [IGNORE] as fast as possible.
The tools to avoid that fate will be architecturally quiet by default: present everywhere, audible almost nowhere, and calibrated to speak when the value of interrupting exceeds the cost of one more interruption. That’s a much harder product to build than one that surfaces constantly and calls it engagement. It requires better models, better clinical context, better calibration, and, perhaps hardest of all, the institutional restraint to accept that sometimes the best evidence your product is working is that the physician never noticed it was there.
For twenty years, we’ve built systems that trained clinicians to ignore them. We shouldn’t be surprised by what happens if we make AI even louder.
The future of clinical decision support should be intelligent enough to earn our attention... by rarely asking for it.