"What is the dose in renal impairment?" looks like a lookup and is actually a clinical assessment wearing a lookup's clothes, and AI answers inherit the disguise: a single confident adjustment, delivered without stating which renal measure it used, which patient it assumed or which product it meant, is the characteristic shape of the renal answer that harms. This article does not provide dose values, deliberately, because values belong to the exact SmPC and current guidance for the exact product and patient; what it provides is the variable map, the things a safe renal answer must have resolved, and the rule that governs the whole territory: do not calculate from incomplete data, and do not accept a calculation that did.
The variables a renal answer must declare
The measure itself: eGFR and creatinine clearance are different quantities, product recommendations are written against specific measures, and using them interchangeably without checking which the SmPC specifies is the territory's foundational error; a trustworthy answer states which measure it used and why. Trajectory: stable impairment and changing function are different clinical situations, and an adjustment computed on last month's stable number is wrong for this week's acute decline; the answer must know which patient it is dosing. Body size and composition: low body weight, frailty and extremes of muscle mass distort creatinine-based estimates in known directions, which is precisely where estimate-based confidence is least earned. Dialysis: its presence, type and timing change everything and belong to specialist and renal-pharmacy territory, not to synthesis. Indication: the same medicine can carry different renal recommendations for different indications, so the answer must hold the indication, not just the molecule. Loading versus maintenance: the distinction that changes what gets adjusted, and that generic answers routinely blur. And the product itself: renal recommendations live in the exact SmPC, formulation-specific, per the product-level discipline at /blog/product-specific-prescribing-active-ingredient-not-enough. The classes where these variables do their most visible work, direct oral anticoagulants, nitrofurantoin, metformin, gabapentinoids and selected antimicrobials, are exactly the classes populating incident reports, which is not a coincidence: they are where the variables interact and where the lookup illusion is most expensive.
Making AI declare its working
The prompt discipline converts the variable map into behaviour: state the measure and value with its date, the trajectory, weight where relevant, the indication, the exact product, and the co-medicines, then require the answer to say which renal variable it used, which source's threshold framing it applied, and what information is missing. The response test is symmetrical: an answer that names its measure, cites the SmPC's renal section and lists what it does not know has behaved safely whatever its content; an answer that returns an adjustment without declaring its working has failed regardless of whether the number happens to match, because undeclared working is unverifiable working, and the 90-second screen's source and patient stages exist for exactly this. Calculators sit inside the same discipline: the estimate's inputs and assumptions visible, the result carried to the SmPC's framing rather than treated as the answer, and the do-not-calculate rule governing throughout, missing weight, undated creatinine, unknown trajectory, then the task is obtaining data, not computing around its absence.
The escalation half
Renal prescribing is also where escalation design earns its keep, and the triggers are worth holding as bright lines: acute kidney injury and rapidly changing function; dialysis in any form; severe impairment meeting high-risk medicines, anticoagulants and narrow-therapeutic-index territory above all; paediatric renal dosing; and any case where the variables above cannot all be resolved. The routes are the ordinary good ones, renal pharmacy, medicines information, specialist advice, per the framework's team-working competency, and the sick-day layer completes the picture: medicines with temporary-withholding logic in intercurrent illness belong in the monitoring plan's intercurrent section, /blog/ai-medication-monitoring-plans-anp, with the patient taught the rules in plain language. The territory's one-sentence summary for the stack: renal questions are assessments, not lookups, the SmPC and current guidance hold the values, and AI's legitimate job is assembling the variables and sources while declaring, loudly, what it was not given.
Frequently asked questions
Why do different references frame renal thresholds differently?
Because they were written against different measures, populations and purposes, the register problem again; the resolution is the product's own SmPC framing for product decisions, with discrepancies escalated rather than averaged.
Are renal calculators safe for frail older patients?
They are estimates whose known distortions concentrate exactly there, which is why the result is an input to judgement plus the SmPC, and why low-weight frail patients sit permanently in the extra-care category.
What should be documented for a renal adjustment?
The measure and value used, its date, the source applied, the adjustment made, the monitoring and review plan, and any advice sought, the working shown, which is both governance and the next clinician's safety.