BRENNEN

Industries / Health / Healthcare

Care at home

Nursing, monitoring and hospital-level treatment given in the patient's own home rather than in a hospital or clinic.

7

Milestones achieved, from 1859 to 2020.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

In depth

The same ground is charted in more depth, as an industry of its own.

The road

AchievedWatch forOpen

Achieved

  1. 1859

    Nursing the poor at home

    William Rathbone, a Liverpool philanthropist, hired the nurse Mary Robinson to care for his wife at home during her final illness. After his wife died he asked Robinson to nurse the city's poor in their own homes, the start of district nursing in England, and worked with Florence Nightingale to spread the service. In 1887 Queen Victoria granted £70,000 from the Women's Jubilee Fund to found the charity now called the Queen's Institute of Community Nursing.

    qicn.org.uk

    Credit

    William Rathbone and Mary Robinson

  2. 1893

    Visiting nurses in New York

    Lillian Wald founded the Henry Street Settlement in New York in 1893 and with it launched the Visiting Nurse Service of New York, which sent nurses into patients' homes. In the 1918 influenza pandemic she directed the public health nursing response in New York City.

    europepmc.orgeuropepmc.org

    Credit

    Lillian Wald

  3. 1969

    Hospice care at home

    Cicely Saunders opened St Christopher's Hospice in London to its first patient on 13 July 1967, by the hospice's account the first to link expert pain and symptom control, compassionate care, teaching and clinical research. In 1969 its staff began what it calls the world's first home-based hospice service.

    stchristophers.org.uk

    Credit

    Cicely Saunders and the staff of St Christopher's Hospice

  4. 2005

    Hospital-level care at home

    B. Leff and colleagues treated older patients with pneumonia, heart failure, flare-ups of COPD (a chronic lung disease) or cellulitis (a skin infection) at home instead of on a ward, at three Medicare managed care systems and a Veterans Administration medical center; 455 patients took part. Compared with hospital care, stays were shorter (3.2 against 4.9 days) and mean cost was lower ($5,081 against $7,480), with some evidence of fewer complications. Patients offered the option took it at rates of 69% at two sites and 29% at the third.

    europepmc.org

    Credit

    B. Leff and colleagues

  5. 2012

    Telehealth's mixed result

    The Whole System Demonstrator trial in England gave home monitoring equipment (telehealth) to people with diabetes, COPD or heart failure; 3,230 people in 179 general practices took part. In the first year, 4.6% of the telehealth group died against 8.3% of controls, and fewer were admitted to hospital. A follow-up economic study put the cost at £92,000 per quality-adjusted life year (a year of life in full health) gained, with an 11% chance of being cost-effective at the usual £30,000 threshold.

    LessonHome monitoring can improve outcomes and still cost more than a health service will pay for the gain.

    europepmc.orgeuropepmc.org

    Credit

    A. Steventon, C. Henderson and the Whole System Demonstrator evaluation team

  6. 2019

    A watch checks the pulse

    In the Apple Heart Study, 419,297 people used a smartwatch app that looks for an irregular pulse. Over a median of 117 days, 2,161 (0.52%) were notified. Among the 450 who returned usable ECG patches (stick-on recorders of the heart's rhythm), 34% showed atrial fibrillation, an irregular heartbeat that raises the risk of stroke; when a notification came while the patch was worn, 84% matched atrial fibrillation on the patch.

    LessonA watch can screen hundreds of thousands of people at home, but confirming its alerts still needs follow-up: only 450 of the 2,161 people notified returned a usable patch.

    europepmc.org

    Credit

    M. V. Perez, K. W. Mahaffey and the Apple Heart Study investigators

  7. 2020

    Hospital at home approved

    On 25 November 2020 the US Centers for Medicare & Medicaid Services (CMS) launched Acute Hospital Care at Home, letting approved hospitals treat eligible patients, admitted from the emergency department or a hospital bed, in their homes. The rules include an in-person doctor's evaluation before care at home starts and two in-person visits a day by nurses or paramedics. Six health systems were approved on the first day. In 2026 Congress extended the program through 2030, and by the American Medical Association's count 366 programs at 139 health systems in 37 states have been approved.

    cms.govama-assn.org

    Credit

    CMS; the first six health systems; US Congress

Watch for

  1. Voice agents that call patients

    Hippocratic AI says its voice AI agents, which phone patients for tasks such as education after admission, medication adherence and scheduling, have had more than 180 million patient interactions, give correct clinical advice 99.90% of the time and caused no severe harm events, validated by more than 7,500 US-licensed clinicians. It says its new Nurse Co-Pilot, which a hospital nurse starts from the patient's record, can give each nurse back 1 to 4 hours a shift.

    Achieved whenan outside group publishes how often the agents miss a warning sign that should go to a human, from test calls it designed or audited itself.

    hippocraticai.comsmb.panews.com

    Claimed by

    Hippocratic AI, April 2026

  2. Hospital at home results

    Health systems report results from their hospital-at-home programs. Marshfield Clinic Health System reports more than 90% patient satisfaction, a 44% lower readmission rate and a 35% shorter average stay; Ochsner Health reports more than 1,000 bed-days saved in under a year.

    Achieved whenan independent study comparing hospital-at-home patients with similar patients treated on wards, across many of the approved programs, confirms fewer readmissions.

    ama-assn.org

    Claimed by

    the health systems, as reported by the American Medical Association, March 2026

Open

  1. 01

    A spoken red-flag test

    A set of at least 200 scripted phone calls in which an actor plays a patient at home who mentions, among routine requests, a warning sign that should go straight to a nurse, such as chest pain or new confusion. Run the calls against voice agents in test environments, with the maker's or a health system's written permission, and publish how often each agent hands the call to a human in time.

    The brief
    Credit · unclaimed
  2. 02

    OASIS forms from staged visits

    A public test of AI tools that fill in OASIS, the assessment form that US home health agencies complete under CMS rules. Actors play patients in a mock home, a nurse runs the visit, and each tool's filled form is compared item by item with the form completed by two clinicians certified in OASIS coding.

    The brief
    Credit · unclaimed

Open 01 · the brief

A spoken red-flag test

Credit · unclaimed

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

What to build

A set of at least 200 scripted phone calls in which an actor plays a patient at home who mentions, among routine requests, a warning sign that should go straight to a nurse, such as chest pain or new confusion. Run the calls against voice agents in test environments, with the maker's or a health system's written permission, and publish how often each agent hands the call to a human in time.

Why it is still open

Published tests are run by makers or in software sandboxes. Amazon's PatientAgentBench (July 2026) scores AI agents against simulated patients with a jury of language models, and its authors say a high score does not establish readiness for deployment. Hippocratic AI's own metric card for clinical escalation shows a baseline of 95.56% and no current figure.

alphaxiv.orghippocraticai.com

Built with what exists

scenario lists modeled on PatientAgentBench's; volunteer actors or a synthetic voice; a calling service such as Twilio to place and record calls; a speech-to-text model such as ElevenLabs Scribe v3 to transcribe them (see the Note "Fewer word errors"); two nurses to grade each call blind.

  1. First step

    Write 40 scripts, half with a red flag, and have two nurses agree which calls should be handed on before any agent is tested.

  2. Done looks like

    At least 200 calls against each of at least 3 agents, every call graded by two nurses, the share of red-flag calls handed to a human within the call reported with a 95% confidence interval, and the scripts and recordings published.

Open 02 · the brief

OASIS forms from staged visits

Credit · unclaimed

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

What to build

A public test of AI tools that fill in OASIS, the assessment form that US home health agencies complete under CMS rules. Actors play patients in a mock home, a nurse runs the visit, and each tool's filled form is compared item by item with the form completed by two clinicians certified in OASIS coding.

Why it is still open

Epic says its AI records visits and suggests OASIS documentation for field clinicians, but publishes no accuracy figures. A September 2026 scoping review of AI in home-care nursing found 16 studies, most of them early-stage, and a search of Europe PMC found no published test of AI filling in OASIS items.

epic.comeuropepmc.org

Built with what exists

the public OASIS-E2 instrument, in effect since 1 April 2026; a furnished mock home and volunteer actors; the AI documentation tools on the market, under trial or maker permission; a speech-to-text model such as ElevenLabs Scribe v3 (see the Note "Fewer word errors") and a language model such as Claude or Gemini as an open baseline.

cms.gov

  1. First step

    Download the OASIS-E2 start-of-care form, choose 25 items that can be judged from conversation and observation, and write 5 visit scripts that cover them.

  2. Done looks like

    On 30 staged visits, each tool's item-by-item agreement with the two certified coders is published beside the coders' agreement with each other, and a tool counts as matching people when its agreement is within 5 percentage points of theirs.