BRENNEN

Industries / Health / Healthcare / Senior living and assisted care

Medication

Medication in a care home is the chain of prescribing, supplying, storing, giving and reviewing each resident's medicines.

6

Milestones achieved, from 1991 to 2023.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1991

    A list of risky drugs

    Beers and colleagues at the University of California, Los Angeles, asked 13 US experts in two rounds of a Delphi survey, a way of reaching agreement through repeated questionnaires, which medicines are inappropriate for nursing home residents. They published 30 criteria covering sedatives, antidepressants, antipsychotics, blood pressure drugs, painkillers and more, designed to run on pharmacy records so that a computer could apply them.

    ebi.ac.uk

    Credit

    M. H. Beers and colleagues, UCLA Division of Geriatric Medicine

  2. 2008

    Checks for what is missing

    An 18-member panel from academic centres in Ireland and the United Kingdom agreed STOPP, 65 criteria for prescriptions that are likely to be inappropriate in older people, and START, 22 criteria for useful medicines that are often left out. The third version, agreed in 2023 by 11 experts from 8 European countries, has 133 STOPP and 57 START criteria.

    ebi.ac.ukebi.ac.uk

    Credit

    Gallagher, O'Mahony and colleagues

  3. 2009

    Errors for two in three

    The Care Homes Use of Medicines Study (CHUMS) followed 256 residents in 55 care homes in West Yorkshire, Cambridgeshire and central London, who took a mean of 8.0 medicines. It found that 69.5% had at least one prescribing, monitoring, dispensing or administration error, a mean of 1.9 errors per resident. Causes included doctors who did not know the residents, staff workload, interruptions during drug rounds and ordering systems that did not work well.

    pmc.ncbi.nlm.nih.gov

    Credit

    Barber, Alldred and colleagues, funded by the Department of Health's Patient Safety Research Programme

  4. 2013

    Dose pouches, more drugs

    In Sweden a pharmacy machine can pack each person's regular tablets into a sealed pouch for each time of day, a system called multi-dose drug dispensing. A register study of 30,922 people aged 65 or over in the Västra Götaland region who moved onto it between 2006 and 2010 found they took 5.4 different drugs on average three months before the switch and 7.5 three months after, with worse results on several measures of prescribing quality. By 2023, 96.1% of Swedish nursing home residents got their medicines this way, up from 84.2% in 2014, by a later national register study.

    Lessona system that makes medicines easier to hand out did not make the list inside the pouch any shorter, and here it grew.

    ebi.ac.ukebi.ac.uk

    Credit

    S. M. Wallerstedt and colleagues; H. Shepherd, K. Johnell and colleagues

  5. 2014

    Rules for care home medicines

    On 14 March 2014 NICE published its guideline Managing medicines in care homes (SC1). It sets out how homes should keep an accurate list of each resident's medicines, review them, order, store and give them, support residents who look after their own, and handle covert administration, which is giving medicine hidden in food or drink without the resident knowing. A review in December 2017 found no new evidence that changed the recommendations.

    nice.org.uknice.org.uk

    Credit

    NICE

  6. 2023

    Weekly pharmacist misses target

    The CHIPPS trial gave 882 residents of care homes in England, Scotland and Northern Ireland a pharmacist independent prescriber, a pharmacist qualified to prescribe on their own authority, who visited weekly for six months, and compared them with usual care. Falls, the main outcome, did not differ significantly (rate ratio 0.91, 95% confidence range 0.66 to 1.26). The Drug Burden Index, a score of exposure to sedating and anticholinergic medicines, fell significantly, and no safety concerns were found.

    LessonOnly 32.7% of the pharmacists' changes involved medicines linked to falls, so falls over six months were a weak test of their work, by the authors' account.

    ueaeprints.uea.ac.ukueaeprints.uea.ac.uk

    Credit

    David Wright, Richard Holland, Christine Bond and colleagues, published in the BMJ on 14 February 2023

Watch for

  1. Software lifts deprescribing

    MedSafer is software that reads a resident's medicine list and conditions and suggests which medicines could be stopped. In a stepped-wedge trial in five long-term care homes in New Brunswick, Canada, in which homes switched to the tool at staggered times, 36.4% of residents with a potentially inappropriate medicine had at least one stopped with the tool, against 12.7% without it. Falls were higher in the period with the tool (adjusted odds ratio 1.77). The software's co-developers, Emily McDonald and Todd Lee of McGill University, ran the trial and co-founded MedSafer Corp, which licenses it, and McDonald said it "should be the new standard of care".

    Achieved whena group with no stake in MedSafer repeats the result in other care homes and finds no rise in falls.

    jamanetwork.comhealthenews.mcgill.ca

    Claimed by

    McGill University and MedSafer Corp researchers, May and August 2025

Open

  1. 01

    Crush-or-not question test

    A public set of 300 questions of the kind care home staff ask when a resident has trouble swallowing, such as whether a named tablet may be crushed, split or mixed into food. Each answer comes from the product's official label and is checked by two pharmacists. Then score at least three AI assistants on the set.

    The brief
    Credit · unclaimed
  2. 02

    Shared test of chart checkers

    200 made-up care home residents' charts, each with age, conditions, kidney function, recent falls and a medicine list, written to contain known problems from STOPP/START version 3 and the American Geriatrics Society Beers criteria, labelled independently by two pharmacists and published with a scoring script, so that rule-based checkers and AI models can be scored on the same set.

    The brief
    Credit · unclaimed

Open 01 · the brief

Crush-or-not question test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of 300 questions of the kind care home staff ask when a resident has trouble swallowing, such as whether a named tablet may be crushed, split or mixed into food. Each answer comes from the product's official label and is checked by two pharmacists. Then score at least three AI assistants on the set.

Why it is still open

No test of AI answers to crushing questions was found. The nearest is a 2025 study by Henry and colleagues in which GPT-4o matched generic medicines to all their available formulations, such as modified-release or coated forms, only 49% of the time, and crushing questions turn on the formulation.

arxiv.org

Built with what exists

DailyMed, the US National Library of Medicine's label database, and its API; the UK electronic Medicines Compendium for UK labels; Inspect, the UK AI Security Institute's open-source evaluation tool; AI models through their APIs; a small model such as Liquid AI's LFM3, tuned on the answer key on one graphics chip (see the Note "One GPU to fine-tune").

  1. First step

    Pull 50 modified-release tablets from DailyMed and write one crushing question for each, with the label sentence that answers it.

  2. Done looks like

    300 questions, at least 100 of them about modified-release or coated products, published with the answer key and a label link for each; scores for at least three AI models, with the answers that would let staff crush a product that must not be crushed counted separately.

Open 02 · the brief

Shared test of chart checkers

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

200 made-up care home residents' charts, each with age, conditions, kidney function, recent falls and a medicine list, written to contain known problems from STOPP/START version 3 and the American Geriatrics Society Beers criteria, labelled independently by two pharmacists and published with a scoring script, so that rule-based checkers and AI models can be scored on the same set.

Why it is still open

Checkers are tested on private records. MedSafer was tested on residents' real charts in New Brunswick, and a December 2025 test of an AI medication safety review on 277 real NHS patients in Cheshire and Merseyside found it identified all issues and actions correctly for only 46.9% of patients; its authors released 45 vignettes of the failures, not a test set. A 2025 study by Socrates and colleagues at Yale ran an AI pipeline on the records of 100 older emergency patients against the Beers and STOPP criteria: it spotted when a criterion was relevant well (positive predictive value 0.83) but was right only 47% of the time when it recommended stopping a drug, and the records stay private. No public labelled set of care home charts was found.

arxiv.orgebi.ac.ukjamanetwork.com

Built with what exists

Synthea for the made-up patients; the criteria tables in the STOPP/START version 3 paper and the AGS Beers criteria; Python for the scoring script; Inspect to run AI models on the set.

  1. First step

    Generate 50 residents aged 80 or over with Synthea and hand-write STOPP findings into 10 of them to test the labelling form.

  2. Done looks like

    200 charts with at least 400 labelled findings, the agreement between the two pharmacists reported as a kappa statistic, and recall and precision for at least one rule-based checker and three AI models.