A week as a clinical reviewer

A project opens with a written brief, not a meeting. It sets out the clinical area, the kind of reasoning under test, and — importantly — what a strong submission looks like, with two worked examples.

The brief for a diagnostic reasoning project might ask for presentations where the obvious answer is wrong: the chest pain that is not cardiac, the headache that is not migraine. Not zebras for the sake of it. The cases a competent clinician would get right and a pattern-matcher would not.

Reading the brief properly takes twenty minutes and saves several hours later. The most common cause of returned work is a good case that answers a different question than the brief asked.

Writing a case takes longer than people expect — typically forty minutes for something usable. You are not recalling a patient. You are constructing a presentation that behaves like a real one: history that leads somewhere, examination findings that constrain the differential, investigations that a real department would actually run in that order.

The discipline is that you must be able to defend the correct answer. Every case needs a reasoning trail: why this diagnosis, why not the three nearest alternatives, and what would change your mind.

This is also the point where confidentiality matters. Cases are constructed, never lifted from records. Nothing that could identify a real person goes into a submission — a rule with no exceptions.

The second half of the work is reading what a model produced and judging it the way you would judge a registrar's assessment.

The failure mode is rarely a wrong diagnosis stated plainly. It is a fluent, well-structured answer that omits the safety-netting, or that reaches the right conclusion by reasoning that would be dangerous if applied to the next patient. Marking that as correct because the endpoint matched is the single most common mistake new reviewers make.

So the review asks three things. Is the conclusion right? Is the reasoning that produced it sound? And would acting on this answer be safe?

The most valuable thing you produce is not the score. It is the explanation.

"Incorrect" tells a research team nothing. "Correctly identifies the likely diagnosis but does not consider the time-critical alternative, and gives no advice on what should prompt re-presentation" tells them exactly what to fix. That paragraph is the deliverable, and it is the part that genuinely requires a clinician.

Six hours, spread across a week, mostly in the evening. Submissions are reviewed and returned with feedback, and the first one usually comes back with notes — that is expected, and it is how calibration happens.