Don't Just Do Something
When I decided to become a doctor, I made a pledge. And the pledge was that I would put my patient's interest in front of my own interest. — Jerome Hoffman, MD
A teenage girl arrives in the ER on a backboard after a minor car crash with some neck pain, so she's put in a rigid collar by the paramedics. In the ER, however, she meets all five NEXUS low-risk criteria, which says she doesn't need imaging. Then her father arrives: loud, angry, and quick to mention he's a lawyer. He wants a CT of her spine because of her neck pain.
The emergency physician taking care of the teenager was Dr. Jerome Hoffman, lead investigator of NEXUS, the 34,000-patient study that validated those five criteria. He tried explaining the evidence and the risks of the scan. When that didn't work, he told the father the truth about incentives: The scan would end the argument, protect him from a lawsuit, take seconds, bill higher, and no one would ever blame him for a thyroid cancer twenty years later. Everything pointed toward the scan. Then he told the father why he still wouldn't order it: the pledge he made when he became a doctor.
The father, who had been shouting minutes earlier, went silent. They decided not to do a CT.
This story is better hearing Jerry himself tell it, so click below to hear the storytelling on an episode of This American Life. (It's a 5 minute piece if you've got the time.)
The decision rule gave Hoffman evidence that the risk was low enough to stop. Everything else he had to bring himself: the ten minutes, the words, the patience to manage a patient complaint. Let me tell you: it's much easier to just click "Order CT" in the EHR.
The cascade effect, circa 1986
40 years ago, James Mold and Howard Stein named this pattern in the New England Journal of Medicine: the cascade effect, a process that, once started, proceeds stepwise toward a seemingly inevitable conclusion.
Their index case was a 59-year-old admitted for a hernia repair whose cardiology consult snowballed through a stress test, the CCU, and an angiogram. It showed his coronary disease had slightly improved since his last one nine years earlier. All of that... for what?
Bad data and underestimated downstream risk play a part, but Mold and Stein blamed anxiety above all: the patient's, the doctor's, and a whole system's inability to live with a small risk. Now, the problem obviously is preventing bad cascades, but we all know that not every cascade is problematic — for example, we've got cascades in our bodies that stop us from bleeding (if you know, you know). The danger is the wrong trigger for starting a cascade (anxiety; uncertainty), because once a cascade starts it's almost impossible to stop in medicine.
Here's what Mold and Stein's thought could prevent bad cascades:
Get the story right. A good history can tell you whether you need a test at all. Call the family. Get the old records. They're worth their weight in gold.
Know your pretest odds. A positive result means little if the disease was unlikely to begin with.
Have a question before you order a test. If the answer won't change what you do, don't even ask it.
Accept some risk, and share it. Bad outcomes can't be engineered away, and patients will often take a small risk once someone explains it honestly.
Decision support gave us tools for #2 on that list. The other three got less attention, but cascades can start from any of them.
Do-gooders
My friend, gastroenterologist Dr. Spencer Dorn recently wrote about cascades and reminded me of the NJEM piece. He'd traced his 86-year-old patient's chronic abdominal pain to her opioids. Then her PCP ordered an abdominal MRI, which found an unrelated pancreatic cyst. Now she's being referred for an endoscopic ultrasound for a finding that certainly has nothing to do with her pain.
Spencer could very easily rant about the PCP. But I appreciate how actively he rejects snark. Cascades, he argues, are mostly driven by do-gooders: clinicians who want to help, patients who reasonably want answers, and a workflow where ordering a test takes seconds and explaining why you shouldn't takes ten minutes. His old attending's antidote: "Don't just do something. Stand there."
Spencer is more worried about patients increasingly getting preventive full-body scans, which medical societies don't recommend for otherwise low-risk, asymptomatic patients, and turning to chatbots that, in one recent study he cites, leaned toward costlier, test-intensive care. With all that data, he argues, knowing when to stop may be the hardest skill of all. I like to give this another name: clinical restraint.
The AI cascade
Nobody in the stories Hoffman, Mold'n'Stein, and Dorn tell is greedy or careless. They're trying to help, under uncertainty, with someone anxious in the room. I worry about what happens when we add AI to that.
AI models are do-gooders too. Perhaps the best do-gooders we've ever seen. They're built to be helpful, and a missed diagnosis is an obvious failure. An unnecessary workup is easier to overlook. Ask one about a symptom and the helpful-seeming answer is the comprehensive one: here's the differential, here are the tests that could rule things out, consider asking your doctor.
None of that is wrong, exactly. But a lot of the time, I see these chats as very reasonable-sounding first dominoes in a really long cascade effect.
I went back thru the drivers that Drs. Mold and Stein talk about, and they're alive and well, but I think they've just got new fuel:
Driver | 1986 | 2010s | AI-Era |
Anxiety | The doctor worries about missing something. A patient admitted for a hernia repair ends up in the CCU getting an angiogram. | Missing a diagnosis can mean a lawsuit or a failed quality measure. Ordering an unnecessary test rarely gets the same scrutiny. | If we penalize a model for missing a diagnosis but ignore unnecessary workups, we’re teaching it to keep testing. |
The urge to do something | Ordering a test feels better than standing there. A protocol gives everyone something to do. | Order sets and EHR defaults that make the full workup one click. Fee-for-service paying for volume. | Helpfulness. A list of things to do looks like a helpful answer. “You can wait” can look like the model hasn’t done its job. |
Patient expectations | The patient wants an answer. The doctor wants to give them one, and starts sharing their worry. | The patient arrives with a Google search and a request. Saying no may end up in your online reviews. | An anxious user asks whether they need more tests. A model trying to please them can give them more reasons to worry. |
Time | Getting the history takes longer than ordering the test. | You have fifteen minutes. You can order the scan now or spend most of the visit explaining why you shouldn’t. | The model can suggest ten tests in seconds. Someone else has to spend the time explaining why 9 of'em aren’t needed. |
But a clinician starts a cascade one patient at a time. A model's default can run across millions of conversations in a single day. AI can now sit at every link in the chain: suggesting the scan, reading the report, flagging the incidentaloma, recommending the follow-up. The cascade can gather momentum before a clinician ever sees the patient and has the chance to intervene. I see this all the time in the ER as well — poll any of my patients with chest pain and they're all sure they're having a heart attack. When I tell them I'm pretty sure they're not (and then do the appropriate tests to back it up), the wave of relief I see on their faces is real, and instant. The anxiety goes away, and they go home.
Hoffman, standing in front of that father, was that same circuit breaker. There's no guarantee an AI cascade will pass in front of anyone like him.
Nobody has clean odds
"Know your pretest odds" is obviously a super-medical term, but I think it also assumes we actually do know them. Maybe we don't.
In a 2021 survey of 553 primary care clinicians, Daniel Morgan and colleagues found that the median pretest probability estimate was too high in all four scenarios they tested. In the breast-screening scenario, the median estimate after a positive mammogram was 50%. The evidence-based range for that scenario was 3–9%. In the stress-test scenario, clinicians put cardiac ischemia at 70%. The evidence-based range was 2–11%.
Anxiety isn't our only problem. We also overestimate the odds in ways that feed cascades.
We could hand the odds to the machine. But a model trained on routine clinical records learns from how medicine has been practiced. And that record contains the cascade: the patients we decided to test and the incidentalomas we decided to chase.
That can set up a loop. More scanning finds more "disease," including disease that would never have hurt anyone. If a model treats diagnoses in a heavily tested population as the odds for everyone, it can overestimate risk and recommend more testing. If those results feed the next model's training data without correcting for how they were collected, the bias can persist. The problem extends beyond one patient's care: the cascade can persist in the data and compound with each model generation.
This is why good research matters — the NEXUS investigators went beyond mining charts. They prospectively recorded the criteria before imaging in more than 34,000 patients and checked them against the imaging results. Those patients had already been selected for imaging, so even NEXUS didn't measure the odds for everyone. It tested a specific stopping rule in a defined population. That took deliberate, expensive work. (Can you imagine 34,000 patient charts... in the 1990s?) The exhaust of routine care doesn't supply that validation on its own, and a model trained on it can inherit our blind spots.
There may be no neutral place to stand. The doctor's gut is biased, and a model trained on our records can inherit the doctor's gut, written down a few million times. Deciding when to stop can't be fully outsourced to a calculation, human or machine. At some point it's a choice about what risk we're willing to accept, and on whose behalf. That's what Hoffman's pledge was.
Knowing when to stop
When is it safe to stop? Dr. Dorn's patient had a clear explanation for her pain. But every clinician also carries the case where "functional" was the wrong answer. The mesenteric ischemia called gastritis. The colon cancer chalked up to IBS. Premature closure is the cascade's evil twin, and it harms people too.
We need a plan for stopping. Here's what I tend to use:
Make a positive diagnosis. Functional abdominal pain and IBS have real, actual criteria, not just vibes. When the history and appropriate evaluation support the diagnosis, name it. We shouldn't need another negative test to treat it as a diagnosis.
Look for the red flags that should say "keep going": weight loss, bleeding, anemia, symptoms that wake you up, a change in the pattern. That turns "we stopped" into "we're watching for these."
Use time as a test. A follow-up visit is one of the best diagnostics we have, and a diagnosis that holds up over weeks can earn trust a single negative scan can't.
Reopen the workup when something changes, symptoms don't follow the expected course, or new information changes the odds. If someone gets nervous, find out why.
AI should be able to explain the same plan to an anxious patient: here's why we're stopping, and here's what would make us look again. (It's probably great at this, we just need to bake it in.)
What AI decision support owes us
Some of our most useful decision rules (NEXUS, Canadian C-Spine, PECARN, PERC, the Ottawa rules) help identify patients who can avoid testing. Applied to the right patients, they give clinicians permission to stop. AI decision support should inherit that tradition, and it won't by accident.
When we grade these models, catching the diagnosis should be part of the evaluation. We should also count the tests, referrals, and follow-ups the model set in motion, and what those cost patients. It was great to see this starting to make its way into the literature, with papers from Microsoft AI including cost in its outcomes and NEJM AI looking at cost comparisons from AI and humans as well. (It's not obviously just the cost that matters but it's a way to quantify the amount of care.)
A tool that quantifies the risk of missing something should also show the risk of looking, whether that's radiation, false positives, or the workup that follows an incidentaloma. Hoffman did that math out loud for an angry lawyer father. Models should do it by default.
Make the case for not testing as easy to hand over as an order: plain language, ready to share with the patient and drop into the note.
A model should be able to say "no test needed."
I'm not sure we're there yet.
Recommending further evaluation every time may feel cautious, but it can start cascades across millions of conversations. And of course some cascades should get initiated, and run to completion tout de suite: treatment for septic shock, the STEMI activation, stroke alerts. Less isn't always more. We need to get better at deciding when to act and when to stop.
AI lacks a Hippocratic Oath, let alone a Hoffman pledge
Jerry Hoffman held the line because he had both the evidence and the conviction to keep his pledge: I don't want to do a CT scan that will harm my patient. A model has only what we build into it and what we measure it against.
Build it to be helpful the way anxious systems are helpful, and it will set off cascades faster, and at greater scale, than any of us could alone. Build it to know when to stop, and it could be the circuit breaker most patients never get that they also deserve.
Those design decisions are being made this year. The people making them should have to understand what it takes to talk down an angry father at 2am, when it's so much easier to just grab your mouse, click a button, and whisk the patient for a test they don't need.