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.