You can’t nudge your way out of a clinical judgment problem

Two and a half years ago, I wrote about a chair. A randomized trial at Parkland Memorial Hospital had shown that placing a chair within three feet of a hospital bed, facing the patient, changed physician behavior more reliably than any memo ever could. 63% of physicians sat down with their patients when the chair was already positioned, versus 8% who had to retrieve it from the closet. Patient satisfaction rose accordingly. I loved that study; I still do.

But a trial published in JAMA this spring has me thinking hard about what the chair can’t do.

For the past decade, healthcare has been increasingly focused on changing behavior by changing the environment around clinicians. Default order sets with pre-checked boxes. Best practice alerts. The chair. Behavioral economists call this approach “choice architecture.” The underlying logic, borrowed from Richard Thaler and Cass Sunstein, is sound and well documented: Don’t lecture the human, but instead, redesign the environment around the human. It’s inexpensive, scalable, and often works. But notice the quiet assumption embedded in the approach. The clinician is treated as a fixed quantity, someone whose surroundings can be engineered but whose judgment cannot be improved. We’ve become good at improving defaults. The more complicated question is whether we’ve invested as much in helping doctors become better decision-makers.

What if the problem isn’t knowledge?

The new JAMA trial explored that question directly. Researchers at the University of Pittsburgh randomly assigned 800 board-certified emergency physicians at non-trauma centers to either usual education or a theory-based video game called Night Shift. Physicians played the game on iPads for two hours upfront and 20 minutes each quarter thereafter. The game uses case-based narratives and time-pressured pattern recognition puzzles designed to recalibrate the intuitive judgments physicians use when deciding whether an injured older adult needs to transfer to a trauma center.

Undertriage of severely injured patients 65 and older fell from 57% in the control group to 49% in the intervention group, a model-adjusted difference of seven percentage points, with no increase in overtriage. That happened across 1,147 hospitals, with no EHR build, no new committee, and no consultants (I can say that). Roughly one in 13 severely injured older adults who would have stayed put under usual care instead got to a trauma center, where the literature says mortality drops by 10% to 25%.

Here’s the detail that should make every chief medical officer (CMO) pause: 94% of these physicians had completed Advanced Trauma Life Support, the criterion-standard for trauma care. They had the training. They knew the guidelines. And the undertriage rate was still north of 50%. This is not a knowledge problem, and we’ve known it for a generation. A 1999 meta-analysis in JAMA found that didactic continuing medical education, the conferences and lectures we still spend billions on, produces no significant change in physician behavior. Interactive education that lets clinicians practice skills does.

Why would a video game succeed where classroom work failed? Because the median physician in this trial saw two severely injured older adults per year. Two. At that rate, a 30-year career supplies roughly 60 reps of a high-stakes pattern recognition task. No athlete, pilot, or musician would accept those practice conditions, yet physicians are the only elite performers we expect to maintain rare-event skills purely through live performance. The game compressed a career’s worth of rare cases into an afternoon, with immediate feedback that real practice almost never provides. This shouldn’t surprise us either. A meta-analysis of simulation-based education with deliberate practice found it substantially superior to traditional clinical education for skill acquisition.

Nudges manage clinicians. Training invests in them.

The trial’s authors draw the contrast that serious games engage clinicians as partners in improving their own judgment, while nudges seek to shape behavior from the outside and do nothing for the decision maker’s actual competence.

I lived on the other side of this as a chief medical information officer. Most people who have tried to configure drug-drug interaction checking know that an interaction alert that might genuinely help a brand-new attending is pure noise to a hospitalist who is twenty years in. A warning that makes sense for a primary care physician is nonsensical for the cardiologist who co-prescribes those two drugs deliberately, every week, with eyes open. The system cannot tell the difference, because choice architecture often must treat every prescriber as the same interchangeable decision point.

Across 34 studies, override rates for these alerts run from 55% to 98%. And when nudges leave the carefully tended garden of academic studies, their effectiveness shrinks. An analysis of all 126 trials run by the two largest US nudge units, covering 23 million people, found real-world effects of 1.4 percentage points versus the 8.7 points reported in academic journals. Still useful, but six times smaller than we may believe.

There’s a deeper point here about what kind of professionals we think we employ. A nudge says: We’ve arranged your environment so that your unexamined reflexes produce the right answer. Training says: We’re investing in your judgment because you are the instrument, and the instrument is worth maintaining. One of these is how you treat a decision point. The other is how you treat a physician. The game’s effect also travels. It lives in the clinician’s head, follows them across every hospital where they pick up shifts, and requires no maintenance ticket. The chair only works in rooms that have one.

Honesty requires a caveat. The trial’s composite clinical outcome, 30-day mortality or readmission, did not improve. The study wasn’t powered for it, and the mortality case for trauma center transfer rests on decades of prior evidence, but anyone claiming this trial proved a survival benefit is ahead of the data. What it proved is that physician judgment, long treated as fixed, is trainable at national scale for the price of an iPad and three hours a year.

What this means for your portfolio

None of this is an argument against nudges. It’s an argument that we’ve been running a one-instrument orchestra. Choice architecture excels when the right answer is known, and the barrier is friction or forgetting: the chair, the default, the pre-checked box. Judgment training excels when the decision is rare, time-sensitive, and pattern-based, and when feedback is too sparse for experience to teach, such as trauma triage, pediatric sepsis in a community emergency department, and stroke recognition.

What can a CMO or chief health information officer do? Audit your alert inventory and ask, for each one: Is this compensating for a friction problem or a calibration problem? If an alert fires constantly and gets overridden constantly, you may be using furniture to solve a judgment problem. Next, pick one low-base-rate, high-stakes decision in your system and pilot judgment training for it, dosed like the therapeutic it is. In this trial, the effect was strongest within 30 days of gameplay and faded after; the quarterly boosters were doing real work. Training is not an event. It has a half-life.

Keep placing the chair. It works, and it costs nearly nothing. Just remember what it did and didn’t do. The chair changed where the doctor sat. It was never going to change what the doctor saw.