The thirty-second triage

A woman walks into a crowded emergency department on a winter evening with chest tightness and a vague, hard-to-place nausea, and within a minute of the front desk taking her details a number appears on the triage screen. The number is low. It places her in the queue for the less urgent track, the one where the wait stretches past two hours on a night like this. A nurse glances at the score, glances at her, and — because the department is full and the system has been right far more often than not — accepts the routing and moves to the next arrival. Nothing about this is reckless. It is, by every local measure, the department working as designed.

The score was produced by an acuity model that reads the intake fields, the vitals captured at the desk, and whatever of the patient's history the record surfaces, and returns an estimate of how sick she is relative to everyone else in the room. It is a genuinely useful tool. On most nights it sorts the truly urgent toward the front faster than a stretched human team could, and the published validation behind it is real. But on this night, with this patient, the number is low for a reason no one in the department can see — and that invisibility, not the number, is the thing worth stopping on.

The clinician inherits a call she cannot read

Consider what the nurse actually has in front of her. Not the model's lifetime track record, which she cannot use to second-guess this particular output. She has a number, and the number disagrees, slightly, with the unease she felt looking at the patient — the diaphoresis, the way the woman was sitting. What she does not have is the one thing that would let her resolve the disagreement: the basis for this score. Which fields drove it down. Whether the model leaned on an absent prior cardiac history that simply was not in this record. Whether the atypical presentation that worried the nurse is exactly the kind of case the model systematically underweights. None of that is on the screen. The screen shows a verdict and withholds the minutes.

So the nurse is left with a bad choice dressed as an easy one. She can defer to the score and move on, which is what the pace of the room pushes her toward and what the score's reputation invites. Or she can override it on a hunch she cannot articulate, accepting that if she is wrong she has slowed the room for nothing and exposed herself to a second-guessing she has no record to answer. The model has handed her a recommendation and kept the reasoning. It has made her accountable for a call whose basis she was never permitted to read. That is not decision support. It is risk transfer with a clean interface.

A score that cannot show its basis has not removed the clinician's judgment. It has overruled it without telling her why.

The patient is owed an account of her own case

Now move the clock forward. Suppose the low score was wrong — the nausea was cardiac, the two-hour wait mattered, and afterward someone asks how she came to sit in the slow queue. The honest answer the department can give today is thin: the model returned a low acuity estimate and the team, reasonably, followed it. But that answer cannot tell the patient or her family the only thing they need to know, which is what the system actually relied on when it decided she could wait. It cannot say which features pulled the number down, what the model could not see, or how close the call was to landing in the other queue. The account that is owed to her — an account of her case, not the model's average performance across a cohort she was never part of — does not exist, because nothing was built to capture it at the moment the score was produced.

This is the gap the situation exposes. A model can be excellent in aggregate and mute in the instance. The aggregate is what gets it cleared and deployed; the instance is where a real person is sorted. Accountability is owed to the patient on the table, and the patient on the table is one case, not a distribution. The only account that answers to her is an account of that case: what was weighed, what was missing, what it would have taken to land somewhere else. A system that can produce a validation curve but not the basis for the decision it just made has standing with a regulator and none at all with the person it routed.

What the department should require

The fix is not a more accurate model, though a better model is always welcome. It is that every consequential routing carry, alongside the number, a basis the clinician can inspect in the moment and the patient can contest afterward — the features that moved the score, the inputs the model leaned on, the inputs it lacked, and a mark when the case sits near the edge of the model's evidence. Captured at the instant of the decision, not narrated after the fact when the outcome is known and the story tends to flatter. A record made at the same moment as the decision — a Decision Receipt for the routing — is what turns a population-trained score into something a single patient can stand on.

None of this slows a good model down. A score with a real reason for placing this patient here loses nothing by showing it; the nurse integrates it with what she sees and either confirms the call or overrides it with grounds she can point to. The only model that suffers under this requirement is the one whose recommendation could not survive a careful clinician reading the reason behind it — and that is precisely the recommendation a patient most needs the standing to question. The thirty seconds the triage took is not the problem. The problem is that those thirty seconds left nothing behind that the woman in the slow queue could ever ask to see.

The scenario above is illustrative — a composite drawn to show a pattern, not an account of any real person, hospital, or event.

— Dispatches · Summit Cognitive · The Casebook