The alert worked. Then nothing happened.

Most health systems have a clinical decision support project that everyone agreed was successful and that … changed absolutely nothing. My version was a pediatric blood pressure alert.

A quick tutorial, for which continuing education credits will regrettably not be offered: Pediatric hypertension is inherently more difficult to detect than adult hypertension because there is no single number to remember. Before age 13, the threshold varies by age, sex, and height percentile, requiring clinicians to translate each reading via a reference table. No clinician does that reliably at the bedside during a well-child visit, nor should anyone have to. This is precisely the kind of arithmetic a computer should perform silently in the background, and ours did.

By every measure agreed upon in advance, the alert worked. It fired when it should. Clinicians acknowledged it. Orders were placed. Yet nothing downstream changed for most patients.

For a long time, I assumed the build had failed because of alert fatigue, a poorly chosen trigger, or a missing order set. I now think it failed because we handed a clinician an ambiguity without specifying what happened next. By default, the next step belonged to the person we had just interrupted, who was, conveniently, already overworked. The alert identified something that might be wrong and then walked away, presumably satisfied with its contribution.

A new study in Pediatrics made me reconsider that build because it includes a finding that should not be possible under any theory of clinician behavior ever used to justify an implementation plan.

When the easiest clinical action isn’t the one clinicians choose

Researchers James Nugent and David Kaelber queried the TriNetX network (encompassing 87 million patients across 51 healthcare organizations) for children aged 1 to 17 with a new diagnosis of primary hypertension. They asked whether these children received the workup recommended by the 2017 American Academy of Pediatrics guideline during the six months after diagnosis.

The answer was mostly no. Fewer than one in five received a urinalysis (19.9%) or a basic chemistry panel (17.7%). Ambulatory blood pressure monitoring (ABPM), which the guideline considers necessary to confirm the diagnosis, reached 11.9% by 2024. That figure had doubled during the decade, which provides a fairly bleak indication of where it started.

The surprising finding was that the most frequently ordered test was the echocardiogram, at 21.1%.

A urinalysis (UA) needs a specimen cup and costs roughly $20. An echocardiogram requires a referral, often an authorization, an appointment, an echo technologist, and a pediatric cardiologist to interpret the study. Under the guideline, echocardiography is not indicated until drug therapy is considered, typically after six to 12 months of lifestyle intervention. The inexpensive, fast, universally indicated test lost to the expensive, slow, gated, later-indicated one.

Every implementation framework I have ever used predicts the opposite. Health systems build order sets because they assume friction suppresses ordering. If friction were the binding constraint, urinalysis should have won by a mile. The commentary accompanying the study calls the result “perplexing,” which is accurate. It then moves on, which is unfortunate, because the perplexing part may be the most useful finding in the paper.

Why clear next steps matter more than reducing clicks 

The tests make more sense when sorted according to a different variable. Not cost, effort, or even guideline priority, but what happens when the result returns.

An echocardiogram produces a relatively clear disposition. If the result is normal, the family receives reassuring news and the concern about structural heart disease is largely closed. If something appears abnormal, the child goes to cardiology and the next several decisions move to another clinician. Either way, the path is fairly obvious when the in-basket message arrives.

Now consider other tests performed on the same child. The UA shows trace protein, or perhaps the chemistry panel shows creatinine at the upper end of normal. No disposition is waiting. Instead, there is repeat testing, a judgment about whether the finding is meaningful, an uncertain threshold for nephrology consultation, and a conversation with a parent before the clinical significance is fully understood. The result does not route neatly to the next step. It takes up residence in the physician’s head.

ABPM is the extreme case. A positive study does not complete the episode of care or hand the child to someone else. It confirms the clinician is caring for a child with hypertension and begins a longitudinal relationship with no obvious handoff built into it. The test answers the diagnostic question while creating several operational ones.

The tests completed most often were the ones whose results tended to resolve. The tests completed less often were the ones whose results opened another question that someone had to continue holding. That distinction is not captured in a workflow map showing the number of clicks required to place each order, but it may explain more about ordering behavior than the clicks do.

I have never met a physician, and I don’t believe a physician exists, who consciously skips a urinalysis because an abnormal result might create more work. The mechanism is subtler and almost certainly subconscious. Some orders close a visit, while others open a loop, and open loops are easier to defer. Each individual deferral may be defensible. Blood pressure can be rechecked at the next visit, and labs can be ordered then. There may be no need to alarm the family today. Repeated tens of thousands of times, those reasonable decisions appear in a national database as 19.9%.

Nobody chose 19.9%.

The authors did not know the specialty of the ordering clinician, so some echocardiograms were almost certainly ordered by subspecialists after referral. Echocardiography is also not a foolish test in this population; the literature cited by the authors shows left ventricular hypertrophy in 33% to 41% of young people at initial diagnosis. The guideline’s objection is principally about sequence, not value. Those caveats may partly explain the echocardiography rate. They do not explain the ambulatory monitoring rate or solve the larger problem.

The real problem starts before the diagnosis

Only 0.5% of children in the study cohort carried a hypertension diagnosis code. The expected prevalence is 2% to 5%, suggesting that most hypertensive children were never identified at all.

Elevated pediatric blood pressure is the least-resolved result in the entire sequence. It is not a diagnosis or a disposition, and it does not point toward one immediate action. It is an ambiguity delivered during a visit with no spare capacity, and the appropriate response is to open yet another loop. The clinician must verify the measurement, review prior readings, determine when it should be repeated, explain the concern without creating unnecessary alarm, and make certain that the child does not disappear into the generous space between “follow up later” and “someone is actually responsible.”

We are measuring adherence to the evaluation of a diagnosis that is usually never made. When the diagnosis is made, it is usually not confirmed by the test the guideline requires. The denominator has already fallen apart before the quality measure begins.

Somewhere, a dashboard is probably reporting that the alert fired successfully.

Build a worker, not another reminder

The first response should be to centralize the workup because the volume is too low to distribute. The study identified roughly 50,000 incident cases across 51 organizations over seven years, or approximately 140 cases per organization per year. Among pediatricians in a large health system, most will diagnose pediatric hypertension only about once a year. Nobody develops muscle memory at that frequency, and nobody should be expected to.

Healthcare already accepts this logic for anticoagulation clinics, oncology navigation, and heart failure bridge programs. Pediatric hypertension has a similar profile. It’s rare enough that distributed expertise is unrealistic, yet protocolized enough that much of the work does not require a physician. A pooled team could own the workup, including device logistics, insurance obstacles, scheduling, family education, and follow-up. This would not remove the primary care clinician from the relationship. It would remove the fantasy that expertise emerges from annual repetition.

The second response should be to make technology a worker rather than a reminder. Traditional decision support interrupts a person and asks that person to do something. It is essentially a computerized colleague who points at an unfinished task and then leaves. A more useful system would identify elevated readings, determine which tests had already been completed, prepare the appropriate orders, initiate logistics for ambulatory monitoring, and support scheduling.

Most importantly, it would carry the disposition logic for the results. Trace protein would arrive with a defined next step. A borderline creatinine would be accompanied by a protocol rather than a question mark. An abnormal ambulatory study would initiate a care pathway rather than generate another message for a physician to consider between patient visits. Genuine clinical judgment would remain essential, but it would be reserved for cases requiring judgment rather than wasted on reconstructing a routine process from memory.

In my experience, the most effective technology resolves work rather than merely routing it. Much of the current conversation about AI in the EHR concerns drafting text, which is useful and considerably less interesting than giving software responsibility for moving clinical work toward completion. A prettier note is not a care model.

The governance questions every CDS program should answer

No clinical decision support build should ship until the team has documented what happens after every plausible result and who is responsible for making it happen. “The ordering clinician decides” may be a legitimate answer, but it should be a deliberate decision, not the default outcome of everyone leaving the governance meeting five minutes early.

The design question is not merely whether an alert causes an order to be placed. It is whether the system can carry the clinical process to a safe and unambiguous next state. Many consequential failures occur after the order is placed, after the result returns, and after the original visit has disappeared from everyone’s immediate field of view.

Measure the clinical funnel, not the alert

Guideline adherence rates establish that something is wrong but reveal little about where the system failed. Health systems should report the clinical funnel instead. Among children with an elevated reading, how many were recognized? How many received a diagnosis? How many had the recommended tests ordered and completed? How many results produced a documented clinical action? How many open loops remained unresolved?

Most health systems already possess this data. The difficult part is agreeing that a clinical process should be measured through completion rather than when the EHR records a click.

Leadership teams may already suspect where the local funnel falls off a cliff. My guess is that failure is not where the alert fires or even where the order is placed.