Clinical Reasoning AR · Case study

Third-year medical students know the theory but freeze at the bedside. I designed an adaptive tutor that makes them discover the clinical reasoning themselves, on a holographic patient, so the knowledge actually transfers. It’s built to ride on top of an XR clinical simulation platform.

The learning method: guided-discovery induction. The learner is never handed the diagnosis. They gather the findings, induce the rule that links them (raised JVP + oedema → volume overload), name the illness script, then get tested on a presentation they were never taught. Reasoning the learner generates, not recognises, is what transfers to a real patient.

The induction method here was derived and extended from an earlier text-based prototype, Aux, an AI learning tool for medical students that teaches clinical reasoning through guided discovery rather than direct explanation. It shows students carefully chosen contrasting cases and probes them to induce the underlying pattern, with the goal of building durable illness scripts. This case study takes that method and extends it onto a holographic patient.

View the Aux prototype ↗

The problem & the opportunity

Third-year students can recite the mechanism of heart failure and still freeze at the bedside: their knowledge is inert, filed by discipline instead of by clinical cue. XR simulations render impressive patients, but a scripted scenario runs the same for everyone; it can’t adapt to how a particular student reasons. The platform owns immersion; transfer needs a tutor that responds to each learner’s thinking.

A byproduct: the reasoning becomes measurable

Because the student reasons in the open (free-text, step by step), the thinking is no longer hidden, and two things fall out of that. An instructor who can’t sit beside 120 learners gets a per-student read on who’s guessing, who’s pattern-matching, and who can transfer, with per-node mastery (observe → localize → discriminate → confirm) and the exact step where reasoning broke. And the method validates itself without a bolted-on exam: the signal is a transfer check, a never-taught presentation of the same illness script: did the reasoning fire on an unseen case, or only the one they practised? Both are genuine benefits, but they fall out of the method rather than drive it.

Instructor side: per-node illness-script mastery bars, a count of correct-first-try answers and reasoning breaks, and a full reasoning trace with one step flagged as a break. — **Left, the reasoning layer:** the student floats a plausible-but-wrong diagnosis (a PE); the tutor never says “wrong”, it surfaces *her own* findings as a contrast and lets the model fail on her reading, so she discards it herself. Reacting to the specific wrong turn *this* learner took is the one thing a scripted scenario can’t do. **Right, the byproduct:** because that reasoning is explicit, an instructor sees per-node mastery and the exact step where reasoning broke.

	Legs	Neck veins	Onset
Clot in lung (PE)	one calf swollen	may rise, acutely	sudden
Fluid overload	both legs, symmetric	raised	over days

Positioning: a layer, not a competitor

The sharpest decision was not to rebuild the platform. Rendering holographic patients is a solved, capital-heavy problem the XR-simulation platforms already own, and competing there burns a war chest to reach parity. So the tutor sits on top of one: it consumes the findings the patient already produces and adds the one thing a script can’t: an adaptive read on each learner’s reasoning. That turns a would-be competitor into a distribution partner and shrinks the build to “the reasoning layer.”

Scoping the MVP

Instead of “a tutor for all of medicine,” I scoped to one scenario, end to end: acute dyspnea → decompensated heart failure, chosen because its findings are visible at the bedside, exactly where attention-direction in AR earns its keep. The demo walks that full reasoning loop with the live instructor panel, so the mechanism and the outcome claim are tangible. It’s the argument for the MVP, not the shipped MVP.

Pedagogy decisions, made as product decisions

Discovery over telling: the student induces the pattern from findings they gathered: recall, not recognition, is what transfers.
Free-text, not multiple choice: menus give the answer away and inflate the metrics. Costlier to grade, honest about mastery.
A fading scaffold: heavy guidance first, then the training wheels come off and it only nudges when the learner stalls; support removed as they improve is what builds independence.

The reframe. Every one of these choices moves the product’s promise from “a vivid experience” to “a student who can reason on a patient they’ve never seen.” That’s a claim a program director can put in front of an accreditor, and it repositions the whole platform from content delivery to learning outcomes.

Guardrails against a hallucinating tutor

In clinical education a confidently wrong answer is worse than none, so the design keeps the model out of the position to invent one. The AI never owns the clinical truth: the findings and correct illness scripts come from the validated scenario; the tutor runs the dialogue and grades reasoning against that fixed answer key, not free judgment about what’s true. When its confidence in grading a free-text step is low, it abstains and flags the exchange to the instructor, so a model slip becomes a reviewable flag, not a fact taught to a student. Validated content owns the medicine; the tutor owns the conversation.

What this is, and what’s next

This case study is an interactive demo: a representative acute-dyspnea session with the reasoning loop and a live instructor panel; it makes the method tangible, but it isn’t a shipped AR product. Next: build the layer against real holographic-patient findings, run a controlled pilot measuring transfer against slide-study, and add a second scenario to test whether the method generalizes beyond one disease.