Modules 44 through 48 sit on top of everything built so far, rather than replacing any of it — the differential from Lesson 3, the bias audit from Lesson 4, the calibration habits from Lesson 8 and 10 all still run. What changes is one input: instead of asking "what does the evidence say," this lesson asks "what does the evidence say for someone like this patient specifically," which is a genuinely different question whenever the patient sits outside the population a guideline was built from.
This is also where the course's honest tension is sharpest. A guideline is, almost by definition, evidence aggregated across many people who are not the one in front of you. Personalization isn't a license to ignore the guideline — it's a structured way of asking whether this specific patient is close enough to the trial population for the guideline's conclusion to transfer, and what to do when the honest answer is "not entirely."
"The guideline says X" and "X is right for this patient" are not the same sentence, and conflating them is one of the most common ways good evidence gets applied badly. This lesson is about keeping the two sentences separate on purpose.
This is the distribution-shift problem, applied to clinical guidelines instead of models — a rule trained on one population doesn't automatically generalize to an individual case that sits outside that population's distribution.
Module 44 — The Applicability Check
Before a guideline is applied, Module 44 asks a specific question: on what axes does this patient differ from the population the guideline's evidence was drawn from, and does any of those differences plausibly change the risk-benefit calculation the guideline assumed? This is a narrower question than "is this patient different" — nearly every patient is different from a trial population in some way that doesn't matter. The module is built to isolate the differences that do.
"This patient is 84 and the trial capped enrollment at 75" is worth stating explicitly and then actually reasoning about — not as an excuse to abandon the guideline, but as the specific fact that determines whether it still applies.
This is checking whether a data point falls inside a model's training distribution before trusting its prediction — an out-of-distribution flag, not a rejection of the model.
Module 45 & 47 — Weighing Conflicting Evidence, Weighing Patient Preference
Two more modules extend the applicability check into the two directions personalization actually pulls. Module 45 handles the case where two guidelines, or a guideline and a more recent study, genuinely conflict — and specifies naming the conflict explicitly rather than quietly picking whichever one supports the reasoner's prior instinct. Module 47 handles the harder case: the evidence is clear, but the patient's own stated preference points a different direction, and the module's job is to make that a visible trade-off rather than a silent override in either direction.
Overriding a clear patient preference and silently deferring to it without discussion are both failure modes — the module's job is to force the trade-off into an actual conversation instead of letting either default win by omission.
This is a multi-objective optimization with two named objectives — evidence-predicted outcome and stated patient preference — surfaced explicitly rather than collapsed into a single silent utility function.
Worked Example: A Guideline That Doesn't Quite Fit
A standard anticoagulation guideline for atrial fibrillation, evidence drawn from trials that largely excluded patients over 80 with a prior major bleed.
This patient is 86 with a GI bleed eighteen months ago. That's precisely the excluded population — the guideline's risk-benefit balance was never actually measured for someone like this, which doesn't mean anticoagulation is wrong, but does mean the guideline's confidence interval doesn't transfer cleanly.
A newer, smaller observational study in exactly this excluded population suggests a lower-dose regimen preserves most of the stroke-prevention benefit with meaningfully less bleeding risk — weaker evidence than the original guideline, but evidence that actually includes this patient's profile where the guideline's evidence doesn't.
The patient, once the trade-off is explained, says the memory of the prior bleed makes her more afraid of a repeat bleed than of a stroke — a preference the evidence alone can't settle, and shouldn't be allowed to silently override.
The final decision — lower-dose anticoagulation, explicitly documented as a deviation from the standard guideline dose, with the applicability gap and the patient's stated preference both recorded as the reasons why — is exactly what personalization is supposed to produce: not a rejection of evidence, but a visible, defensible account of why this patient's evidence-based plan differs from the population-level default.
Module 48 — Documenting the Deviation
Module 48 is the closing habit: whenever a plan deviates from a standard guideline for personalization reasons, the deviation itself needs to be documented as explicitly as the plan — not buried in a note, but stated as its own line, specifying which axis of applicability drove the deviation. This matters for continuity of care as much as for the reasoning itself: the next clinician who sees this patient needs to know the standard dose was a deliberate, reasoned departure, not an error to be silently corrected back.
An undocumented deviation looks identical to an error to whoever reads the chart next — the one-sentence log is what turns a reasoned personalization into something that survives a handoff instead of getting quietly reversed.
This is a changelog entry for a decision that deviates from a default configuration — the point isn't just making the deviation, it's making the deviation legible to the next person who reads the code.
Homework for Lesson 12
- Take a guideline-driven decision you're familiar with. Run Module 44's applicability check: name the population the evidence came from, and one way a real patient of yours differed from it.
- If evidence and patient preference ever genuinely pulled in different directions in a case you know, write Module 47's trade-off explicitly — the cost of following preference, stated in the evidence's own terms — rather than resolving it silently in either direction.
- Write a one-sentence Module 48 deviation log for that case, the kind a future clinician could read cold. Bring it forward — Lesson 13 asks what turns this kind of judgment, repeated across many cases, into something recognizable as expertise rather than just competence.
This lesson draws directly on Module 44 — Guideline Applicability Check, Module 45 — Conflicting Evidence Resolution, Module 47 — Preference-Evidence Trade-Off, and Module 48 — Deviation Documentation, all from the VibeRounds Prompt Directory. If you're coming from the evidence side, the companion Evidence-Based Medicine for Techies course — see especially its treatment of external validity — pairs well with Module 44's applicability question here. Neither course is a clinical decision tool; see the VibeRounds disclosure statement for full terms.