Deprescribing looks like prescribing run backwards and is not, because the governing questions change: not "what does the guideline add?" but "what is this medicine still for, what would stopping cost, and what matters now to the person taking it?", asked for a resident whose frailty, life expectancy and priorities the single-condition evidence base never enrolled. AI enters this territory usefully as an organiser and dangerously as an answerer, and the failure mode has a name worth coining: the mechanical stop list, an AI-generated cull of a fourteen-medicine regimen, plausible line by line and blind to goals, withdrawal physiology and the human being, implemented as if organisation were decision. This article is the non-mechanical version.
The seven-way sort
Start where the polypharmacy method starts, /blog/ai-multimorbidity-polypharmacy-anp, with the frailty-specific categories: immediate harm risk, the combinations and continuations that jump every queue; no current indication, the legacy layer that care-home regimens accumulate across transfers; therapeutic duplication, including the brand-generic pairs fragmented records hide; high anticholinergic and sedative burden, the frailty-specific category where cumulative load drives falls, confusion and decline, and where burden thinking beats medicine-by-medicine thinking; long time to benefit, the preventive medicines whose payoff horizon should be compared honestly with the person's own; monitoring burden, the tests and appointments the regimen demands of someone for whom each is a cost; and must-not-stop-abruptly, the flag that separates candidates from methods, because a medicine can be right to stop and dangerous to stop carelessly. An AI pass that sorts fourteen medicines into these seven has done real work; what it has produced is an agenda, and the article's central rule follows: scores and lists are prompts, not decisions, and every downstream step belongs to people.
Goals, capacity and the conversation that decides
The decision layer deprescribing actually runs on. Goals of care first: what matters now, comfort, function, staying out of hospital, being awake for visits, because the same medicine is right under one goal and wrong under another, and no burden score knows which. Capacity and voice: the resident's own view wherever capacity permits, and where it does not, the best-interests process with family and carers properly inside it, the legal and human structure that makes a medication change legitimate rather than administrative. Carers and staff as evidence: the people who see swallowing difficulty, post-dose drowsiness and refusal patterns hold data no record contains, and a review that has not asked them has reviewed the list, not the person. And honesty about prognosis where goals require it, at the depth the person and family want, because time-to-benefit reasoning is only decent when the time half is handled with care.
Method: one change, watched, then the next
The implementation discipline that separates deprescribing from disruption. Sequence deliberately: highest-harm and no-indication categories first, one or two changes at a time, because simultaneous changes make benefit and harm unattributable, and frail physiology amplifies every transition. Taper where the flag says taper: withdrawal, rebound and recurrence are the categories the must-not-stop sort exists for, and the exact approach for the exact product comes from the SmPC and guidance, not from a synthesis's memory. Watch on purpose: a documented review plan after each change, what improvement is expected, what recurrence would look like, who is watching, when it is checked, which is the monitoring framework pointed backwards, /blog/ai-medication-monitoring-plans-anp. Record the reasoning: goal, category, decision, plan, the documentation that makes the review defensible and the next clinician safe. And keep the door open: deprescribing is trial-shaped, restarting is not failure, and telling the resident and family so is part of the method. AI's honest place in all of it is the same as everywhere in this cluster: it organises the evidence and drafts the structure, the clinician owns priorities, sequencing and follow-up, and the resident owns the goals.
Frequently asked questions
How often should care-home regimens be fully reviewed?
On the local and national schedule as the floor, and event-driven above it: transfers, falls, acute illness and goal changes each reopen the sort, because regimens drift precisely at transitions.
Where do deprescribing tools and burden scores fit?
As structured prompts inside the sort, useful for surfacing candidates and quantifying load, governed by the same rule as the AI pass: prompts, not decisions.
Who should be in the room for the review?
The prescriber, the resident or their voice, someone who gives daily care, and pharmacy where the regimen's complexity warrants, which in fourteen-medicine territory is usually.
How does this method interact with structured medication review programmes?
It is their clinical core with the AI organising pass added: the national and local review structures supply the schedule and governance, the seven-way sort and goals conversation supply the content, and the documentation serves both.
What about medicines prescribed by specialists?
They enter the sort like everything else and exit through communication: candidates identified in a specialist's domain go back as a conversation with the specialist, shared care respected, which the record of goals and burden makes easy to open.