What Resource-Limited Emergency Medicine Teaches About Risk

There is a particular moment, familiar to anyone who has worked in a resource-limited emergency department, that never quite happens in a well-resourced one. A patient in front of you, a question in your head, and the realisation that the investigation which would answer it — the scan, the assay, the specialist opinion — is not available tonight. Not delayed. Not 'long wait'. Not available. The question doesn't go away. The patient doesn't go away. What changes is that the answer now has to come from you.

I practise emergency medicine in a resource-limited state hospital, in a setting where that moment is not an exception but a working condition. This piece is about what that condition teaches — because I've come to think the lessons say something general about clinical risk, and something pointed about how we're currently building decision-support technology for medicine.

Investigations as outsourced reasoning

In a fully resourced department, it is possible — common, even — to use investigations as a substitute for reasoning rather than a test of it. Uncertain about the abdomen? Scan it. Troponin equivocal? Repeat it, and the one after. Each test is individually defensible. Cumulatively, something subtle happens: the clinician's probability estimate stops doing load-bearing work. The differential becomes a list of things to exclude with machinery rather than a ranked set of beliefs to be examined. You can practise for a long time this way, and mostly safely — the machinery is good.

Resource limitation removes the option. When the scanner is unavailable and the specialist is hours away, your pre-test probability is no longer an academic concept from an evidence-based medicine seminar. It is the decision. You discover quickly that 'I want to rule it out' and 'I believe this is unlikely but the consequence of missing it is unacceptable' and 'I genuinely don't know' are three different cognitive states — and that the well-resourced version of you had been allowed to blur them, because the same CT request covered all three.

The first lesson of scarcity, then, is diagnostic honesty. You learn what you actually think, because what you actually think is all you have. Examination findings get re-examined rather than re-ordered. The history gets taken again, properly, because the history is the highest-yield test still available. It is uncomfortable, and it is the best training in clinical probability I know of.

Risk calibration with real stakes

The second lesson is about thresholds. Every clinical decision is implicitly a threshold decision — at what probability of serious disease do I act, observe, or reassure? In a resource-rich system, thresholds can be set almost arbitrarily low, because the cost of checking is a form to sign. The system will absorb near-unlimited caution. Whether that produces better outcomes is genuinely debated — over-investigation has its own harms, incidental findings have careers of their own — but the individual clinician rarely feels the cost of their own caution.

Scarcity reprices everything. The observation bed you use is one another patient needs tonight. The transfer you arrange consumes an ambulance that serves a wide area, and the receiving centre's patience is a finite shared resource too. Caution is no longer free, which means it has to be allocated — spent where the threat justifies it, withheld where it doesn't. That is risk calibration in its real form: not the elimination of risk, which was never on offer anywhere, but its deliberate, accountable distribution across patients under a constraint.

What surprised me is how clarifying this is rather than how burdensome. When caution is free, every decision can hide inside 'just to be safe'. When it is priced, you must be able to say — to yourself, at minimum — why this patient gets the resource and that one doesn't. The reasoning is forced into the open, where it can be examined. Plenty of error survives in the open too. But a particular kind of unexamined, reflexive practice does not.

The discipline of acting under irreducible uncertainty

The third lesson is the hardest to convey to anyone who hasn't worked this way. In a well-resourced system, uncertainty feels temporary — a state you pass through on the way to the result that resolves it. Scarcity teaches that uncertainty is frequently irreducible on the timescale that matters: the decision must be made before the answer can exist. Send this patient on a long transfer on clinical suspicion alone, or watch them overnight with what's available and reassess at first light?

What develops, over time, is a discipline with three parts. You make the decision explicitly rather than letting it make itself through delay — deferral is also a decision, usually the worst-examined one. You build the reassessment into the plan as a hard structure rather than an intention: who looks again, at what, when, and what finding triggers escalation. And you make peace with the audit that matters — not 'was I right?', which on ambiguous cases is substantially luck, but 'was the decision sound given what was knowable at the time?'. Outcome and decision quality come apart under uncertainty. Settings that cannot afford the illusion of certainty teach you to judge yourself on the second, because the first was never fully yours.

That last distinction, between outcome and decision quality, may be the most exportable thing scarcity teaches. It is true everywhere, in every system. Resource limitation just makes it impossible to ignore.

What this has to do with decision-support

I said at the start that this points at how we build clinical technology, so let me make the connection explicit.

Most clinical AI and decision-support is built, implicitly, for the resource-rich workflow: it assumes the next test is available, treats investigation as the natural response to uncertainty, and measures itself on diagnostic accuracy against a resolved ground truth. The recommendation engine that says 'consider CT angiography' is not wrong, exactly. It is answering the easy half of the question. The hard half — the half that resource-limited clinicians answer every shift — is: what do you do when you can't, or shouldn't, or not yet? What is the best decision given this patient, these resources, this distance to definitive care, tonight?

That is a harder problem, and a more universal one than it looks. 'Resource-limited' is not a category of unfortunate elsewhere; it is every system at 4am, every department in a winter surge, every setting where the scanner queue is six hours and the decision is needed now. A well-resourced system under load becomes a resource-limited system, and its tools degrade badly because they were built on assumptions of abundance.

I don't know of a system that does this well. I suspect the teams best placed to build one are those who have spent time watching medicine practised where abundance was never the baseline. Scarcity, it turns out, is not just a hardship to be engineered away. It is a forcing function for exactly the kind of reasoning that the next generation of clinical tools claims to want to support — and mostly doesn't yet understand.

What this means

None of this romanticises scarcity. Resource limitation harms patients; the missing scanner is not a philosophy seminar, and I would not trade the lesson for the machine if offered the machine. The point is narrower and, I think, more durable. Constraint exposes the true structure of clinical decisions — probability, thresholds, the price of caution, the difference between a good outcome and a good decision — that abundance allows clinicians and systems to leave unexamined. Medicine everywhere runs on that structure. Some of us just get to see it without the cladding.