BRENNEN

Industries / Health / Healthcare / HealthTech

Monitoring and alarms

Monitoring and alarm software gathers readings from bedside monitors, wearables and home devices, decides which ones matter, and alerts a clinician, on hospital wards and at home.

7

Milestones achieved, from 1966 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1966

    A computer reads bedside monitors

    At the University of Southern California's shock research unit in Los Angeles, Herbert Shubin and Max Harry Weil connected a digital computer to bedside monitors so that it collected patients' vital signs automatically. A 2015 review of computing in intensive care credits them with "introducing the computer to the ICU in 1966"; Weil's own history says digital processors were first used in their unit in 1962, in shock units the two had begun at USC in 1958.

    pmc.ncbi.nlm.nih.govhrcak.srce.hr

    Credit

    Herbert Shubin and Max Harry Weil, University of Southern California

  2. 1980

    A shared test for alarms

    Software that sounds an alarm for a dangerous heart rhythm has to be tested on recordings whose rhythms are already known. In 1980 a team at MIT and Beth Israel Hospital in Boston first distributed the MIT-BIH Arrhythmia Database: 48 half-hour, two-channel heart recordings from 47 people, 23 picked at random from about 4,000 day-long recordings made at the hospital from 1975 to 1979 and 25 picked for rare but important rhythms. PhysioNet, which hosts it under an open license, calls it "the first generally available collection of standard test recordings for evaluating arrhythmia detectors", and a 2026 FDA submission for heart-rhythm software still tested on it, following the ANSI/AAMI EC57 testing standard.

    physionet.orgaccessdata.fda.gov

    Credit

    George Moody and Roger Mark, with colleagues at MIT and Beth Israel Hospital

  3. 2010

    Ward oximeters page the nurse

    Dartmouth-Hitchcock Medical Center in New Hampshire put patients recovering from surgery on continuous pulse oximetry, with a message to the nurse's wireless pager whenever a reading crossed an alarm limit. Comparing 10 months after the change with 11 months before, rescue events (emergency calls to save a deteriorating patient) fell from 3.4 to 1.2 per 1,000 discharges and transfers to intensive care from 5.6 to 2.9 per 1,000 patient days, while two comparison wards saw no change.

    Lessonthe gain does not always carry over: in a 2025 randomized trial of 400 surgical patients at two Copenhagen hospitals, wireless monitoring with alerts to staff phones did not significantly cut the time patients spent with severely abnormal vital signs, its main measure, although fewer had adverse events by day 30 (31.5% against 42.5%).

    europepmc.orgpmc.ncbi.nlm.nih.gov

    Credit

    A. H. Taenzer, J. B. Pyke, S. P. McGrath and G. T. Blike, Dartmouth-Hitchcock

  4. 2013

    Alarm fatigue named a hazard

    In April 2013 The Joint Commission, which accredits US hospitals, issued Sentinel Event Alert 50: its voluntary reports held 98 alarm-related events from January 2009 to June 2012, 80 of them deaths, and the most common contributing factor was alarm fatigue, staff numbed by constant alarms; it estimated that 85% to 99% of alarm signals need no clinical action. From 2014 a National Patient Safety Goal required accredited hospitals to name their most important alarms, and from 1 January 2016 to have policies on alarm settings and on who may change or switch them off. A count at the University of California, San Francisco, published in 2014, found 2,558,760 alarms in 31 days on 77 intensive care beds, or 187 audible alarms per bed per day, and 88.8% of 12,671 annotated heart-rhythm alarms were false.

    LessonThe Joint Commission had a goal on alarms in 2003 and 2004 and retired it in 2005, judging that enough attention had been paid, and the problem returned.

    digitalassets.jointcommission.orgdigitalassets.jointcommission.orgpmc.ncbi.nlm.nih.gov

    Credit

    The Joint Commission; Barbara Drew, Xiao Hu and colleagues, UCSF

  5. 2017

    One warning score for England

    An early warning score gives 0 to 3 points for how far each of six routine measurements is from normal (breathing rate, oxygen saturation, systolic blood pressure, pulse, consciousness or new confusion, and temperature), adds 2 points for extra oxygen, and ties the total to how fast staff must respond. The Royal College of Physicians produced NEWS in 2012 and NEWS2 on 19 December 2017; NHS England and NHS Improvement mandated NEWS2 for adults in acute and ambulance trusts with a Patient Safety Alert in April 2018, and by the count NICE quotes from NHS England's clinical lead, 139 of 140 acute trusts had moved or were moving to it. A 2020 NICE briefing reviewed five software systems that compute the score from typed or monitor-fed readings and send alerts, costing £30,000 to £90,000 to install and £0.35 to £0.70 per bed per day.

    Lessonthe code that adds up the score needs its own checks; NICE records an April 2015 field safety notice that VitalPAC versions 2.3 and 3.0 under-scored the early warning score.

    rcp.ac.uknice.org.uk

    Credit

    Royal College of Physicians; NHS England and NHS Improvement

  6. 2020

    A risk model pages nurses

    Kaiser Permanente Northern California's Advance Alert Monitor computes a deterioration risk score every hour from each ward patient's electronic record; specially trained nurses watch the alerts remotely and pass them to rapid response teams. Rolled out at 19 hospitals from August 2016 to February 2019 and later at all 21 of Kaiser's hospitals in the region, it was linked with lower mortality in the 30 days after an alert: 15.8% against 20.4% in hospitals not yet using it, an adjusted relative risk of 0.84, or about 520 fewer deaths a year by the authors' estimate.

    Lessonthe rollout was staggered, not randomized, and the authors write that they cannot rule out broad cultural change as the cause.

    pmc.ncbi.nlm.nih.gov

    Credit

    G. J. Escobar, V. X. Liu, P. Kipnis and colleagues, Kaiser Permanente Division of Research; funded in part by the Gordon and Betty Moore Foundation

  7. 2025

    Heart alerts never reviewed

    In mobile cardiac outpatient telemetry, a patient at home wears a heart monitor that sends events to a monitoring service, where technicians review them and report serious ones to the ordering doctor. Some ECG events received by Philips' Monitoring Service Application from July 2022 to July 2024 were never routed to a technician, because an analysis step in the software had been disabled; they included episodes that should have been escalated, such as atrial fibrillation, pauses and ventricular tachycardia. Philips sent customers a correction on 18 December 2024, and the FDA put it in its most serious recall class, citing 109 reported injuries and 2 deaths; the recall record lists 41,282 customers. Philips offered to reprocess the missed data but left it to clinicians to tell their patients.

    Lessona routing fault in monitoring software can run unnoticed for two years unless the service checks that every event it receives is reviewed.

    fda.govaccessdata.fda.gov

    Credit

    Philips; US Food and Drug Administration (FDA)

Watch for

  1. Epic's warning score cleared

    Epic's Deterioration Index Version 2 runs inside Epic's electronic record and scores each adult inpatient from 0 to 100 for the risk of an escalation of care within 24 hours (transfer to intensive care, a rapid response call, or a code: an emergency call for resuscitation) or death within 72 hours, from 195 inputs such as vital signs, laboratory results and nursing assessments. The FDA cleared it on 18 September 2026 as substantially equivalent to an earlier cleared tool, eCARTv5. In Epic's own retrospective test on about 10.7 million scored observations from three US sites, the area under the curve (a 0.5-to-1 measure of how well a score ranks people who go on to have an event above those who do not) was 0.83; at the high-risk threshold the score flagged 3% of observations and caught 28% of events, and 34% of its flags were followed by one. An outside before-and-after study at RWJBarnabas Health's 11 hospitals linked the earlier version, with alerts pushed to rapid response teams, to fewer in-hospital deaths among high-risk patients (18.6% against 23.1%).

    Achieved whena study by people outside Epic measures Version 2 in routine use and reports alerts per patient-day, events caught and missed, and patient outcomes against a comparison group.

    accessdata.fda.govhitconsultant.net

    Claimed by

    Epic Systems, September 2026

  2. AI filters implant heart alerts

    An insertable cardiac monitor is a small recorder placed under the skin that sends suspected abnormal heart rhythms to the clinic through a remote monitoring platform. Implicity designed its ILR ECG Analyzer Version 2 to cut these monitors' false alerts: it re-reads each episode and labels it normal or abnormal, for monitors from Abbott, Biotronik, Boston Scientific and Medtronic. The FDA cleared it on 2 July 2026 on Implicity's retrospective test of 2,659 episodes from 1,710 patients at 70 European and US centers: on monitors that already filter alerts with their maker's own AI, it kept 98.3% of true arrhythmias and correctly labeled 61.6% of the rest as normal; on monitors without that filter, 94.3% and 75.6%.

    Achieved whena clinic or research group outside Implicity reports, from routine use, how many false alerts the tool removed, how many true arrhythmias it labeled normal, and how much review time it saved.

    accessdata.fda.gov

    Claimed by

    Implicity, July 2026

Open

  1. 01

    A NEWS2 answer key

    A public set of at least 2,000 synthetic observation sets, each with the correct NEWS2 sub-scores, total and response level worked out from the Royal College of Physicians' charts. It covers every boundary value of the six measurements, both oxygen saturation scales (the second is for patients at risk of type II respiratory failure), extra oxygen, new confusion, a single measurement scoring 3, and missing, zero or impossible entries. Run it through every NEWS2 calculator that can be reached: open-source code, free web and phone calculators, test versions of hospital e-observation systems with their makers' permission, and general AI chat models. Publish each one's error rate after giving makers time to fix what is found.

    The brief
    Credit · unclaimed
  2. 02

    A synthetic ward alarm bench

    An open bench that generates at least 1,000 bed-days of simulated ward monitoring (heart rate, breathing rate, blood pressure, oxygen saturation and ECG), with deteriorations planted at known times and recorded noise and sensor faults mixed in, then replays the streams through alarm rule sets and reports, for each, alarms per bed per day and how many planted deteriorations it caught and how fast. Rule sets include fixed monitor-style limits, wider limits with delays, NEWS2 recomputed continuously, and open machine-learning false-alarm filters.

    The brief
    Credit · unclaimed

Open 01 · the brief

A NEWS2 answer key

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 at least 2,000 synthetic observation sets, each with the correct NEWS2 sub-scores, total and response level worked out from the Royal College of Physicians' charts. It covers every boundary value of the six measurements, both oxygen saturation scales (the second is for patients at risk of type II respiratory failure), extra oxygen, new confusion, a single measurement scoring 3, and missing, zero or impossible entries. Run it through every NEWS2 calculator that can be reached: open-source code, free web and phone calculators, test versions of hospital e-observation systems with their makers' permission, and general AI chat models. Publish each one's error rate after giving makers time to fix what is found.

Why it is still open

Software has scored early warning wrongly before, and no public test exists. NICE records an April 2015 field safety notice that VitalPAC versions 2.3 and 3.0 under-scored the early warning score, and a 2025 Sheffield study reports that its hospital's system gives 3 points to any measurement entered as zero, so the authors checked every zero by hand. NHS England lists its FHIR standard for sharing NEWS2 scores as deprecated, the alpha FHIR guide behind it lists no test cases, and searches of Europe PMC and GitHub found no public test set for NEWS2 calculators, only a sample chart renderer.

nice.org.ukeprints.whiterose.ac.ukstandards.nhs.uknhsconnect.github.io

Built with what exists

the RCP's NEWS2 charts and its 2021 additional implementation guidance; a short reference scorer in Python, checked by hand against the charts; Browser Use 1.0, a web agent, to type each case into web calculators and read back the result (see the Note "Deleting 80% to start over"); Gemini Decision 1, which scores up to 1,000 fixed-answer questions in one request, to put the cases to a chat model in two requests (see the Note "A thousand questions, one pass"). Synthetic cases only; no patient data.

rcp.ac.uk

  1. First step

    Write the 200 boundary cases for the six measurements and the two oxygen scales, and run them through five free web calculators.

  2. Done looks like

    2,000 cases whose key two clinicians have checked and agreed on; at least 15 calculators tested, including at least 3 hospital e-observation systems; for each, the share of wrong totals and wrong response levels, broken down by rule (oxygen scale, extra oxygen, confusion, zero and missing values); cases, results and code published under an open license after each maker has had 30 days to respond.

Open 02 · the brief

A synthetic ward alarm bench

Credit · unclaimed

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

What to build

An open bench that generates at least 1,000 bed-days of simulated ward monitoring (heart rate, breathing rate, blood pressure, oxygen saturation and ECG), with deteriorations planted at known times and recorded noise and sensor faults mixed in, then replays the streams through alarm rule sets and reports, for each, alarms per bed per day and how many planted deteriorations it caught and how fast. Rule sets include fixed monitor-style limits, wider limits with delays, NEWS2 recomputed continuously, and open machine-learning false-alarm filters.

Why it is still open

Public alarm data are short clips around single alarms. The PhysioNet 2015 Challenge set holds 1,250 five-minute intensive care alarm records labeled true or false, and VTaC (2024) holds 5,037 six-minute ventricular tachycardia alarms, 71.4% of them false. Both can score a filter on single alarms but cannot show alarm load per bed-day or a missed deterioration, because they hold no full days and no outcomes, and full-day counts from real wards, such as UCSF's, are not public. The nearest synthetic work, Veritas-RPM (arXiv, April 2026), generates 530 synthetic epochs to test false-alert suppression in home monitoring, not ward streams.

physionet.orgphysionet.orgarxiv.org

Built with what exists

the Pulse Physiology Engine, an open simulator (Apache 2.0) forked from the US Defense Department-funded BioGears program, which models conditions including pneumonia, ARDS, airway obstruction and tension pneumothorax; recordings of baseline wander, muscle artifact and electrode motion noise from PhysioNet's Noise Stress Test Database (Open Data Commons Attribution license); the PhysioNet 2015 set (Open Data Commons Attribution) and VTaC (CC BY-SA 4.0) to train and check filters; Python. All patients are simulated.

pulse.kitware.compulse.kitware.comphysionet.org

  1. First step

    Simulate 20 bed-days in Pulse, 10 of them with a planted pneumonia or airway obstruction, add noise from the stress test database, and count alarms under fixed default limits.

  2. Done looks like

    1,000 bed-days with at least 200 planted deteriorations of at least four kinds; for at least five rule sets, alarms per bed per day, the share of deteriorations alarmed within 15 minutes and the median time to first alarm, with 95% confidence intervals; two clinicians confirm that a sample of 50 planted deteriorations looks clinically plausible; code and data released under open licenses.