Week 7: Triage: Order by Need

1Run it first come, first served

Nobody argues for arrival order here. The chest pain goes first, and if you asked the wrist — who arrived an hour earlier and will now wait a further half hour — they would agree.

That unanimity is remarkable and worth interrogating, because it is the only place in this course where you will find it.

2What triage actually is

A priority queue with a fixed number of classes. Within a class, arrival order. Between classes, strict priority.

The Australasian Triage Scale uses five, and the thing that makes it a real instrument rather than a vibe is that each category comes with a target time to treatment:

Other countries run near-identical instruments under different names — the Manchester Triage System, the Emergency Severity Index — and the convergence is itself informative. Several health systems, working separately, arrived at roughly five classes with roughly these targets.

The target is what turns a wait into a failure. A category 2 patient at twelve minutes is a department having a bad afternoon. A category 1 patient at twelve minutes is the thing the entire apparatus exists to prevent. Notice that this is a completely different way of reporting a queue from anything in Act I: not a mean, not a percentile, but a rate of breaching a promise that was made per person.

The scene above collapses five categories into three, so that the room stays readable. Everything that follows is true of five.

Two properties do the real work.

It is preemptive at the front. A category 1 arriving at minute 59 of your wait goes ahead of you, and there is no cap on how often that can happen. Your wait is not a queue position; it is a queue position that can be invalidated at any moment by an ambulance.

It sorts on information FCFS throws away. Arrival order is a fact about the queue. Urgency is a fact about the person. That is the whole of the gain, and it is week 6’s lesson again: the rules that beat first-come-first-served all do it by using information first-come-first-served ignores.

3Where the idea came from

Worth two minutes, because the origin explains the shape.

The word is French — trier, to sort — and the practice is usually traced to Dominique Jean Larrey, Napoleon’s surgeon-in-chief, who organised battlefield care around a principle that was genuinely radical at the time: treat by severity of wound, and disregard rank.

Sit with that for a second. The rule triage replaced was not first-come-first- served. It was sort by social position. The innovation was not “use need” so much as “stop using something else”, and it had to be imposed by an institution against the interests of the people it demoted.

The modern mass-casualty version, START and its relatives, keeps the structure and adds the part nobody likes to discuss. Four tags: immediate, delayed, minor — and a fourth for the patients who will not survive with the resources available, who are therefore not treated first even though they are the most ill.

Sickest-first and most-lives-saved are different rules, and in a disaster they disagree.

In a normally resourced department they mostly agree, which is why the distinction stays invisible and why everybody can support triage without ever being asked which of the two they support. Week 12 is the lecture where that question finally has to be answered out loud.

4The same rule, outside a hospital

Once you can see the pattern — fixed classes, a target per class, strict priority between, arrival order within — you find it running in places with no clinical connection at all.

Ambulance dispatch. A call handler, working from a script, assigns a category before anybody has seen the patient, and the category determines the response-time standard the service is measured against. It is triage performed over a phone line by somebody who cannot see the patient, on information from a caller who is frightened. Everything in Priya’s half of this lecture applies to it twice over.

Incident severity. Every operations team on earth runs P1 to P4 with a response target attached to each, which is the Australasian Triage Scale with servers instead of patients. The failure modes are identical, including the famous one: everything gets filed as a P1 by whoever raised it.

Manufacturing and repair. A production line down is a category 1. A degraded machine is a category 3. The targets are in the maintenance contract.

Content moderation and safety queues. Reports are classified on arrival and routed by severity, by a classifier that is often literally a model, at a scale where nobody can audit an individual decision.

The medical version is the one with the best-developed ethics and the worst-tolerated failures, which is why it is worth studying even if you will never work in a hospital. Everybody else is running the same machine with less scrutiny.

5Priya's half of the lecture

Everything above treats the category as given.

It is not given. It is produced — in about ninety seconds, by a tired human with incomplete information — and the system’s performance is bounded by the quality of that judgement, not by the elegance of the queue discipline sitting on top of it.

Three failure modes, all documented.

Under-triage of atypical presentation. Category descriptions encode a prototype. Patients who do not present like the prototype get sorted down. Women having heart attacks are the canonical study.

Under-reading of pain in people accustomed to not being believed. The measurement instrument is a conversation, and conversations carry every bias the participants brought with them.

The bottom category’s wait is unbounded. Strict priority in a busy department means the least urgent patients can be functionally starved — the same mathematics as last week’s trolley, with much worse consequences. Departments handle it with ageing rules, exactly as schedulers do, and the rules are largely informal.

There is a fourth, and it belongs to the institution rather than the nurse. The classifier is an input to the performance report. When a department is measured on the share of category 2 patients seen within ten minutes, the category becomes something the department has an interest in. Nobody has to be dishonest for this to bend: a borderline case carded down is a borderline case that cannot breach, and there is always a defensible clinical story for either choice.

6What a bad classifier actually costs

Two consequences follow, and they generalise past hospitals.

Errors in one direction are not errors in the other. Carding a sprained wrist as a category 2 costs somebody thirty minutes. Carding a heart attack as a category 3 can cost a life. A classifier reported as “92% accurate” has told you almost nothing, because that single number has averaged over two failures that are not remotely comparable. Ask for the confusion matrix, always, and ask which cell is the one that kills.

A sorting rule is only ever as good as the thing it sorts on — which means that every time you improve the rule, you should ask whether the same effort spent on the measurement would buy more. A department choosing between a cleverer queue discipline and a second triage nurse is usually looking at a false choice in which the second option wins.

7The disagreement, stated plainly

Adaeze’s framing is that the ordering rule is where the distributive work happens.

Priya’s is that the rule is only ever as good as the measurement feeding it, and that a beautifully designed priority queue fed by a biased classifier is a machine for laundering that bias into something that looks like clinical objectivity.

The scene refuses to settle it, and the two readings of its readout are why.

Read down the arrival-order row: 53% of category 1 patients wait more than half an hour. Under a 92%-accurate triage, 3.0% do. Triage is not a marginal improvement — it is the difference between half of them and one in thirty. Adaeze’s framing wins that comparison and it is not close.

Read across the accuracy rows and you get Priya’s. Every one of that 3.0% is a patient the department looked at, assessed, gave a number to, and sent to a chair. Arrival order never claimed to have done any of that. A queue that promises nothing cannot mis-sort you; the promise is what creates the failure mode, and the promise is still worth making.

We do not resolve this, and you should not expect the course to resolve it for you. It is the live question in the field, and the final design brief is marked partly on whether you noticed that your own design has a classifier in it somewhere.

Next week: a rule that sorts not on need, but on what you are willing to pay.