MMOOZZGG.com

Research

Evidence for AI in medicine: 12 studies that counted patients, not accuracy

2 min read23 sources

Evidence for AI in medicine, on this page, means only work where the effect of AI was measured on patients or doctors in real practice: randomized, prospective and large observational studies. Company claims are on separate case study pages. Each study has a sample, a metric, a result and a limit.

On this page · 4 sections

Randomized trials

MASAI, Sweden, 106,000 women, The Lancet 2026: +29% cancers found, -44% reading workload, -12% interval cancers; limit: one country, one screening program. EAGLE, Mayo Clinic, 22,641 patients, Nature Medicine 2021: +32% diagnoses of low ejection fraction; limit: two US states. Ryzhikh coloproctology center, Russia, 1,128 patients, 4 clinics, 2024: adenoma detection rate 47.2% against 41.3%; limit: a short-term measure, not mortality.

Prospective and large observational studies

Johns Hopkins University, 590,000 patients, 5 hospitals, Nature Medicine 2022: sepsis mortality 20% lower; limit: a university news release that does not describe the design as randomized. A hospital in Toronto, 13,649 patients, CMAJ 2024: unexpected deaths 26% lower; limit: one ward as the intervention group. San Diego, more than 6,000 hospital admissions, npj Digital Medicine 2024: sepsis mortality 17% lower. Penda Health, 39,849 visits, 2025: -16% diagnostic errors; limit: errors judged by experts, one network, study paid for by the model developer. NHS Grampian, 10,889 women, 2024: +12% cancers; limit: workload was modeled. Stockholm, one-year follow-up after ScreenTrustCAD, 2024, as reported by the developer company: +15% cancers, reading more than 36% faster; limit: comparison of the periods 2018–2019 and 2023–2024.

Surveys and adoption

Mass General Brigham and Emory, 1,430 clinicians, JAMA Network Open 2025: burnout 21.2 points lower in one of the two systems after 84 days; limit: response rate of 11–30%. Kansas, 133 clinicians, 99 answered the survey, JAMIA Open 2025: 73% of respondents spend less time on documentation after a shift; limit: self-assessment. NYU Langone, 336,000 patients, Nature 2023: 80% of readmissions predicted, about 5% more accurate than the standard model.

What this means

There is evidence in screening, in hospitals and in documentation; in first-line diagnosis from complaints there is less so far. The limit of almost every study is one country or one network: the result must be checked on your own population. For service AI, the level of proof is simpler: an A/B test on two weeks of requests.

Questions and answers

Why is the study with 85.5% diagnostic accuracy not here?

It is a test on archived cases, not on patients; the authors themselves write that it needs validation before clinical use. It is listed among the case studies.

Sources and fact-checking

Checked by the editors on 29 September 2026.

  1. Lund University, 30 January 2026: Lund University, Sweden (lunduniversity.lu.se)
  2. Alizila, Alibaba Group, 6 December 2023: DAMO Academy, China (alizila.com)
  3. University of Aberdeen, 21 March 2024: NHS Grampian and the University of Aberdeen, Scotland (abdn.ac.uk)
  4. Lunit, 2 December 2024: Capio S:t Göran Hospital, Stockholm (lunit.io)
  5. PR Newswire, C the Signs, 28 April 2025: 1,084 NHS general practices, England (prnewswire.co.uk)
  6. Clinical Trials Arena, 4 February 2025: NHS and NIHR, United Kingdom (clinicaltrialsarena.com)
  7. Digital Platforms catalog, case no. 2768: Almazov National Medical Research Center and a large Russian bank (platforms.su)
  8. Digital Platforms catalog, case no. 2775: Tyumen Region Ministry of Health (platforms.su)
  9. Mayo Clinic News Network, 6 May 2021: Mayo Clinic, USA (newsnetwork.mayoclinic.org)
  10. University of Oxford, Oxford Cardiovascular Science, 30 May 2024: University of Oxford, United Kingdom (cardioscience.ox.ac.uk)
  11. Healthcare IT News, 29 September 2021: Apollo Hospitals, India (healthcareitnews.com)
  12. Digital Platforms catalog, case no. 2779, 2024: Ryzhikh National Medical Research Center of Coloproctology (platforms.su)
  13. Digital Platforms catalog, case no. 3702, 2 September 2026: a telemedicine service, dentistry (platforms.su)
  14. ComNews, 17 April 2026: La Peregrina dental clinic network (comnews.ru)
  15. Scientific Reports, 18 November 2025: a private dental clinic, Spain (nature.com)
  16. Johns Hopkins University, Malone Center, 22 July 2022: Johns Hopkins University, USA (malonecenter.jhu.edu)
  17. Unity Health Toronto, 16 September 2024: St. Michael’s Hospital, Toronto (unityhealth.to)
  18. UC San Diego Health, 23 January 2024: UC San Diego Health, USA (health.ucsd.edu)
  19. OpenAI, 22 July 2025: Penda Health clinic network, Kenya (openai.com)
  20. Emory University, 22 August 2025: Mass General Brigham and Emory Healthcare, USA (news.emory.edu)
  21. JAMIA Open, 21 February 2025: University of Kansas Medical Center, USA (academic.oup.com)
  22. NYU Langone Health, 7 June 2023: NYU Langone Health, USA (nyulangone.org)
  23. Microsoft AI, 30 June 2025: Microsoft AI, USA (microsoft.ai)

Found a mistake? Tell us

Publications

Mike Mozg

Agentic AI use cases at AT&T: nearly 500 processes now run on agents

AT&T no longer uses AI only to help staff work faster: it is rebuilding whole business processes around agents. SiliconANGLE reported on 14 November 2026 that nearly 500 agentic AI use cases at AT&T are already in production. Here is how the company rebuilds them, where a human still checks the agent, and what other large companies can take from it.

News4 min

Cases on this topic: 30

  • Market case · World

    Lund University, Sweden 29% more breast cancers found and 44% less radiologist workload, 106,000 women

    The randomized MASAI trial of 106,000 women showed that AI mammography screening finds 29% more cancers with no rise in false positives, and the workload of radiologists falls by 44%. The final results appeared in The Lancet in January 2026.

    Oncology

    30 January 2026 · source: lunduniversity.lu.se

    Read the case
  • Market case · World

    DAMO Academy, China 5 cancers and 26 other pancreatic lesions missed at the first diagnosis were found among 20,530 patients

    The model for AI pancreatic cancer detection on an ordinary CT scan without contrast was tested on 6,239 patients from 10 centers. Its sensitivity in finding cancer was 34.1% higher than the average radiologist. In a retrospective test on 20,530 patients in a row, the AI found 5 cancers and 26 other lesions missed at the first diagnosis, with 99.9% specificity: about one false alarm per 1,000 studies.

    Oncology

    20 November 2023 · source: nature.com

    Read the case
  • Market case · World

    NHS Grampian and the University of Aberdeen, Scotland +12% breast cancers found in 10,889 women, up to 30% less workload by estimate

    The first formal prospective evaluation of this AI tool in UK breast screening: in a study of 10,889 women, AI breast cancer screening helped find 12% more cancers than usual practice, mostly invasive and high grade. Modeling showed up to 30% less workload for radiologists.

    Oncology

    21 March 2024 · source: abdn.ac.uk

    Read the case
  • Market case · World

    Capio S:t Göran Hospital, Stockholm +15% cancers detected, -11% recalls, image reading 36% faster

    A Stockholm hospital replaced one of the two radiologists who read each image in its screening program with AI. According to the developer, after the rollout cancer detection rose from 4.8 to 5.5 per 1,000 women examined, the recall rate fell from 2.8% to 2.5%, and reading time became more than 36% shorter.

    Oncology

    2 December 2024 · source: lunit.io

    Read the case
  • Market case · World

    1,084 NHS general practices, England 20% more confirmed cancers and over 61,000 urgent referrals avoided

    The developer reports that GPs in 1,084 practices used its cancer risk assessment platform: more than 235,000 risk assessments and 13,585 cancers diagnosed between January 2020 and September 2024. The share of referrals that ended in a diagnosis was 20% above the national average, and more than 61,000 unnecessary urgent referrals were avoided.

    Oncology

    28 April 2025 · source: prnewswire.co.uk

    Read the case
  • Market case · World

    NHS and NIHR, United Kingdom almost 70,000 women and 30 sites in an announced £11 million AI screening trial

    On 4 February 2025, the UK government announced the EDITH trial of AI breast screening. Almost 70,000 women at 30 sites are to take part, with £11 million of funding from the National Institute for Health and Care Research. The goal: to test whether screening mammography can use one specialist instead of two.

    Oncology

    4 February 2025 · source: clinicaltrialsarena.com

    Read the case
  • Market case · Russia

    Almazov National Medical Research Centre 89% accuracy of the complication forecast, SYNTAX Score in seconds instead of 20–40 minutes

    The AI in Medicine Center at the Almazov Centre built the “Risks” model, which predicts complications for inpatients in real time, and the “Heart” model, which calculates the SYNTAX Score from a coronary angiogram. According to the catalog, the accuracy of the complication forecast is 89% and the accuracy of the index calculation is 94.74%.

    Cardiology

    source: platforms.su

    Read the case
  • Market case · Russia

    Tyumen Region Ministry of Health more than 735,000 patients checked, about 300,000 found at very high risk

    Since 2020, AI cardiovascular risk prediction has worked in 19 medical institutions of the region. It reads the records of registered patients in the background and notifies the doctor about those who need attention. By the end of 2024 it had checked more than 735,000 patients, and about 300,000 of them had a very high risk.

    Cardiology

    source: platforms.su

    Read the case
  • Market case · World

    Mayo Clinic, USA +32% diagnoses of low ejection fraction, 22,641 patients, 348 doctors

    The EAGLE randomized trial in 45 institutions in Minnesota and Wisconsin: AI analysis of an ordinary ECG helped primary care doctors diagnose low ejection fraction 32% more often than usual practice. 22,641 patients and 348 doctors took part.

    Cardiology

    6 May 2021 · source: newsnetwork.mayoclinic.org

    Read the case
  • Market case · World

    University of Oxford, United Kingdom 10-year heart attack risk from a routine heart CT scan, treatment plan changed for up to 45% of patients

    AI heart attack risk prediction reads inflammation in the coronary arteries on a standard heart CT scan and predicts heart events years ahead. In a study of more than 40,000 patients from 8 hospitals, two thirds of serious events happened in people without significant narrowing of the arteries. In a pilot at 4 NHS hospitals, the risk score changed the treatment plan for up to 45% of patients.

    Cardiology

    30 May 2024 · source: cardioscience.ox.ac.uk

    Read the case
  • Market case · World

    Apollo Hospitals, India AI cardiovascular risk score built on data from more than 400,000 patients over 10 years

    Apollo Hospitals, one of the largest hospital networks in India, built an AI cardiovascular risk score on data from more than 400,000 patients collected over a decade. The network says the tool is more accurate than the Framingham score for Indian patients. It also uses lifestyle data: diet, smoking, physical activity and stress.

    Cardiology

    29 September 2021 · source: healthcareitnews.com

    Read the case
  • Market case · Russia

    Ryzhikh National Medical Research Centre of Coloproctology adenoma detection rate of 47.2% against 41.3% in a randomized trial of 1,128 patients

    The first Russian decision support system for colonoscopy highlights subtle polyps on the screen in real time. A multicenter randomized trial in 4 clinics with 1,128 patients showed an adenoma detection rate of 47.2% with AI against 41.3% without AI.

    Endoscopy

    2024 · source: platforms.su

    Read the case
  • Market case · Russia

    A major Russian telemedicine service, dentistry areas suspicious for caries marked on a photo in 5–10 seconds

    The service added a free AI caries detection check for tooth photos: a user uploads images in the mobile version of the website and gets the areas suspicious for caries marked in 5–10 seconds. According to the company, the method also finds hidden caries in the form of a white spot. Published on 19 August 2026.

    Dentistry

    2 September 2026 · source: platforms.su

    Read the case
  • Market case · Russia

    La Peregrina dental clinic network a pilot of an AI module that automates the first reading of CBCT scans

    A dental clinic network and a developer are piloting dental CBCT AI: a module that analyzes cone-beam computed tomography (CBCT), a 3D X-ray method. By the developer’s description, the module analyzes the full scan volume, finds caries, periodontitis (gum disease) and impacted teeth, marks traces of earlier treatment, and prepares a base for a draft treatment plan.

    Dentistry

    17 April 2026 · source: comnews.ru

    Read the case
  • Market case · World

    VideaHealth, a US dental AI company per the company, +80% of children’s cavities correctly found, the first FDA-cleared algorithm for children

    The company reported an 80% improvement in the correct detection of cavities in children with its AI algorithm, which it calls the first pediatric dental AI cleared by the FDA. The algorithm detects cavities in children from age 3, tooth wear and chipping also from age 3, and tartar from age 12.

    Dentistry

    9 February 2024 · source: dentistrytoday.com

    Read the case
  • Market case · World

    Dental Care Alliance, United States AI X-ray analysis rolled out across the network: more than 400 practices, more than 900 dentists, 24 states

    One of the large dental support organizations in the United States announced a rollout of AI X-ray analysis across its whole network: more than 400 practices, more than 900 dentists, 24 states. According to the developer, the FDA-cleared algorithm finds, outlines and measures disease to the millimeter and shows it to the patient.

    Dentistry

    23 July 2024 · source: overjet.com

    Read the case
  • Market case · World

    Private dental clinic, Spain on average 9.93 fewer clinic visits with remote AI monitoring of orthodontic treatment

    A pilot controlled study of 39 children aged 7–12 with aligners: the group with an AI app for remote monitoring visited the clinic 21.65 times on average versus 31.58 in the control group, a difference of 9.93 visits. Treatment duration and the number of emergency visits did not differ significantly.

    Dentistry

    18 November 2025 · source: nature.com

    Read the case
  • Market case · Russia

    Yaroslavl Region Ministry of Health more than 145,000 ECGs in 2023, 848 patients with acute coronary syndrome sent to emergency care

    A portable cardiograph with built-in AI works in the ambulances of Yaroslavl and Rybinsk and in more than 130 rural first-aid stations. The ECG is processed on site, the data goes to the regional cardiology center, and a doctor confirms the conclusion within minutes. In 2023 it processed more than 145,000 ECGs, twice as many as in 2022.

    Emergency and inpatient care

    source: platforms.su

    Read the case
  • Market case · World

    Johns Hopkins University, USA 20% lower sepsis mortality, 590,000 patients, 5 hospitals

    A sepsis early warning system developed at the university was tested on 590,000 patients in five hospitals. Sepsis mortality fell by 20%, and severe sepsis was detected about 6 hours earlier. The results were published in Nature Medicine on 21 July 2022.

    Emergency and inpatient care

    22 July 2022 · source: malonecenter.jhu.edu

    Read the case
  • Market case · World

    St. Michael’s Hospital, Toronto -26% unexpected deaths in the general internal medicine unit, 13,649 patients

    The hospital early warning system AI analyzes more than 100 parameters from the medical history and current state and warns doctors about a decline. After it was introduced in the general internal medicine unit, unexpected deaths fell by 26%. Deaths outside palliative care during the system’s 19 months of work were 1.6%, against 2.1% before. The study was published in CMAJ on 16 September 2024.

    Emergency and inpatient care

    16 September 2024 · source: unityhealth.to

    Read the case
  • Market case · World

    UC San Diego Health, USA 17% fewer sepsis deaths in two emergency departments

    The sepsis prediction AI tracks more than 150 patient parameters in the emergency department without a break, to spot sepsis before the obvious signs appear. In two departments of a university hospital, across more than 6,000 admissions, sepsis deaths fell by 17%. Published in npj Digital Medicine on 23 January 2024.

    Emergency and inpatient care

    23 January 2024 · source: health.ucsd.edu

    Read the case
  • Market case · World

    South Korean university hospitals cardiac arrest forecast 24 hours ahead on more than 50,000 hospital beds

    According to the developer, its cardiac arrest prediction AI forecasts an arrest within 24 hours and ran on more than 50,000 hospital beds in South Korea in October 2025, including more than 20 tertiary hospitals. The developer reported a profit for the third quarter of 2025 and explained it by the growth of its main products and by efficient management.

    Emergency and inpatient care

    13 November 2025 · source: koreaherald.com

    Read the case
  • Market case · Russia

    Moscow City Health Department and 19 regions 14.4 million visits with a diagnosis hint in Moscow polyclinics

    The clinical decision support system reads the patient’s complaints during the visit and offers the doctor three likely diagnoses out of 265 groups. By May 2024 it had processed 14.4 million cases in Moscow polyclinics, and by March 2025, 19 regions had connected to it.

    Primary care

    source: platforms.su

    Read the case
  • Market case · Russia

    Central Polyclinic of Vladimir 12 health indicators in 10 seconds instead of 15–20 minutes

    In September 2025 the polyclinic installed a diagnostic assistant with a camera. In 10 seconds it reads 12 indicators from the patient’s face, including pulse, blood pressure, body mass index and the risk of type 2 diabetes. According to the catalog, a standard first check by a nurse takes 15–20 minutes.

    Primary care

    source: platforms.su

    Read the case
  • Market case · Russia

    Health Committee of the Volgograd region first checkup stage in 15–20 minutes instead of a week

    Mobile ECG screening kits with AI analysis and a diagnosis hint now cover all 33 districts of the region: 68 kits, linked to the regional medical system. The first stage of a preventive checkup fell from a week to 15–20 minutes, and the time to examine a patient was cut in half.

    Primary care

    source: platforms.su

    Read the case
  • Market case · World

    Penda Health clinic network, Kenya 16% fewer diagnostic and 13% fewer treatment errors across 39,849 visits, per the network and OpenAI

    In 15 clinics in Nairobi, an AI clinical decision support tool sits inside the medical record and checks the doctor’s actions in real time with a green, yellow or red signal. In a study the network ran with OpenAI across 39,849 visits, doctors with the AI consultant made 16% fewer diagnostic errors, 13% fewer treatment errors and 32% fewer errors in taking the patient history than doctors without it.

    Primary care

    22 July 2025 · source: openai.com

    Read the case
  • Market case · World

    Mass General Brigham and Emory Healthcare, USA burnout prevalence down 21.2 percentage points in 84 days at Mass General Brigham, 1,430 clinicians in two systems

    Two health systems tested an AI ambient scribe, software that listens to a visit and writes the medical documentation. In Boston, where 873 clinicians took part in the pilot, burnout prevalence fell by 21.2 percentage points in 84 days. In Atlanta, where 557 clinicians took part, documentation-related well-being rose by 30.7 points in 60 days.

    Documentation

    22 August 2025 · source: news.emory.edu

    Read the case
  • Market case · World

    NYU Langone Health, USA a language model trained on 4.1 billion words of records predicts 80% of readmissions

    A university hospital trained a language model on the records of 336,000 patients from 2011–2020: 4.1 billion words of clinical notes and discharge summaries. According to the hospital, the model predicts 80% of readmissions within 30 days, about 5% better than a standard model, and it runs in the hospitals of the system.

    Documentation

    7 June 2023 · source: nyulangone.org

    Read the case
  • Market case · World

    University of Kansas Medical Center, USA 73% of surveyed clinicians spent less after-hours time on documents, 67% reported a lower risk of burnout

    The center offered an AI scribe for doctors, a platform that records visits and drafts notes, to 181 clinicians; 133 started to use it, and 99 of them answered the survey. After a median of 92 days of use, 73% of respondents reported less documentation time outside working hours, 67% a lower risk of burnout and 64% higher job satisfaction.

    Documentation

    21 February 2025 · source: academic.oup.com

    Read the case
  • Market case · World

    HCA Healthcare, USA about 60,000 shift handoffs a day, up to 10 million hours a year on documents, an AI summary rated 86% accurate by nurses

    A chain of 190 hospitals is testing, in 5 of them, generative AI that builds a nurse shift handoff summary from the medical record. According to the company, about 60,000 handoffs happen in the chain every day, and documents and conversations around them take up to 10 million hours a year. The plan is to give the tool to all 99,000 nurses.

    Documentation

    29 July 2025 · source: blog.google

    Read the case
All cases: 62
Contact us