Summary
In 2025 and 2026 three areas of clinical AI produced evidence that a non-specialist can read and weigh: AI-supported mammography reading, an AI sepsis early-warning system, and ambient AI scribes. This digest sets the findings side by side, notes what each can and cannot show, and suggests what a clinic should do with them. This document is general information about published evidence. It is not medical, legal or regulatory advice.
- A randomized trial of nearly 106,000 women reported 12% fewer interval cancers with AI-supported mammography reading.
- An AI sepsis alert received US FDA 510(k) clearance in May 2026, and its developers report 18% lower in-hospital death when clinicians acted on alerts in time. An independent test of a different, widely used sepsis model found it missed 67% of sepsis cases.
- A randomized trial of 238 physicians found a 9.5% time saving per note for one of two scribe tools and no statistically significant saving for the other.
- In all three, a person stayed in the loop, and in all three the authors or sources state limits.
The practical message is not that clinical AI works or fails. It is that results are specific to a tool, a setting and a way of working, and that the quality of the evidence differs greatly between products.
Scope and method
This is an evidence digest, not original research. We read the published trial reports where we could and, where we could not, the university releases and news reports that summarize them. We flag each case. We report figures as the sources give them and do not recalculate them. Selection is ours: we chose three evidence sets with published outcome data in routine settings, and this digest does not cover every clinical AI study.
Evidence at a glance
| Area | Design | Headline result as reported | Main limit |
|---|---|---|---|
| Mammography screening (MASAI) | Randomized trial in Sweden’s national program, nearly 106,000 women | Interval cancers 1.55 vs 1.76 per 1,000 (12% lower); screen-reading workload 44% lower | One program, one AI system, already used double reading |
| Sepsis early warning | 2022 study of more than 764,000 encounters at five US hospitals, described by CIDRAP and Healthcare Dive | 18% lower in-hospital death when clinicians acted on alerts in time | Developer-reported; compares timely with untimely confirmation |
| Ambient AI scribes | Randomized trial, 238 physicians, 14 specialties | One tool saved 41 seconds per note (9.5%); both about 7% better burnout scores | One center, about two months, occasional note errors |
The table repeats figures from the sections below. Compare the designs before you compare the numbers. A randomized comparison with current practice is a firmer basis than a comparison of patients whose alerts were and were not acted on in time, and a vendor-reported figure needs outside confirmation before it carries the same weight.
Finding 1: AI support cut interval cancers in a national screening program
Lund University describes MASAI as a randomized trial within Sweden’s national mammography program. The February 2025 report found 29% more cancers detected with AI support (338 against 262), including 24% more early-stage invasive cancers, with a 44% fall in radiologists’ screen-reading workload.
The January 2026 follow-up in The Lancet, as summarized by Textbook of Digital Health, reported interval cancers of 1.55 per 1,000 women with AI support against 1.76 in the control arm. Invasive interval cancers were 16% fewer (75 against 89), and cancers of a non-luminal-A subtype were 27% fewer (43 against 59). Interval cancers matter because finding more cancers at screening can include slow-growing ones, whereas a fall in cancers between screenings suggests fewer were missed.
The AI triaged exams, with lower-risk exams read by one radiologist and higher-risk exams by two, and highlighted suspicious areas (Lund University). A radiologist still read every exam. We read the paper through these summaries, not the full text.
Finding 2: An AI sepsis alert has clearance and promising but developer-reported outcomes
Healthcare Dive reported on May 12, 2026 that an AI sepsis early-warning system received FDA 510(k) clearance. CIDRAP describes the supporting 2022 study of more than 764,000 patient encounters at five US hospitals, and reports that sepsis patients were 18% less likely to die in hospital when clinicians acted on alerts in time. It also reports that each hour of delayed sepsis treatment reduces survival by 8%.
The counterweight is an independent check of a different product. The University of Michigan validation across 38,455 hospitalizations reported 33% sensitivity, 12% positive predictive value and an area under the curve of 0.63, below the 0.76 to 0.83 originally reported. The model did not detect sepsis in 67% of patients with sepsis yet alerted on 18% of all patients. We found no head-to-head comparison of the two products.
Finding 3: AI scribes saved some time, unevenly, and made errors
UCLA Health reports that its randomized trial, published in NEJM AI, ran from November 2024 to January 2025. One tool cut time per note from 4:30 to 3:49, a 41-second or 9.5% saving. The other tool showed no statistically significant reduction against the control group. Both showed roughly 7% better burnout scores than controls.
UCLA also reports occasional clinically significant inaccuracies in AI-generated notes, such as omissions and pronoun errors, and one mild patient safety event. The authors call for longer, multi-institution studies.
Finding 4: Patient data duties apply whatever the tool
The studies above concern clinical outcomes. Any tool that touches patient information also carries data-protection duties. In the US, a vendor handling protected health information is generally a business associate, and 45 CFR 164.502(e) requires satisfactory assurance in a written contract. In the UK, data concerning health is special category data. The sources we read do not say how common compliance gaps are, so we make no claim about that.
Limits of this evidence
- Three evidence sets are not the field. This digest does not show how AI performs across clinical medicine.
- Secondary sources. Several figures come from university releases and news summaries rather than the full papers.
- Different questions. Screening accuracy, alert benefit and documentation time are not comparable outcomes, so we do not rank the three areas.
- Local settings. Sweden’s national program, five US hospitals and one academic center may not resemble your clinic.
- Time. Evidence moves. Check for newer studies before you rely on any figure here.
Several open questions remain in the sources. We did not find independent replication of the sepsis alert result in other health systems, published false-alert rates for that product, mortality outcomes for the mammography trial, or results on how the screening benefit varies across age groups or breast density. These are the gaps to watch in future studies.
What to do with this
Our own suggestion, not a finding from the studies:
- Ask for the design, not the claim. Request the study, who ran it, the comparison group and the outcome measured.
- Ask for error rates. Sensitivity, positive predictive value and alert volume, in writing.
- Keep a named person responsible. In every area above a clinician or radiologist reviewed or confirmed the output.
- Pilot small. One channel or team, a written pass mark, weekly review of errors, and a review date.
- Settle data duties first. Get the required agreements and controls confirmed before patient data moves.
Do not read this digest as a recommendation of any product. It reports what was studied, and a purchase decision needs evidence for the specific product and setting in front of you.
Our step-by-step evaluation guide turns this into a checklist. For help scoping an AI project, see our custom AI agent work or book a call.
Sources
- Lund University, AI-supported breast cancer screening: new results suggest even higher accuracy, February 4, 2025
- Lund University, AI support in breast cancer screening: fewer missed cancer cases, 2026
- Textbook of Digital Health, AI-supported mammography cuts interval cancers by 12% in first randomised trial of 100,000 women, January 29, 2026
- Lang et al., MASAI interval-cancer results, The Lancet, 2026, doi:10.1016/S0140-6736(25)02464-X (figures as summarized by the Lund University and Textbook of Digital Health pages above; we did not read the full paper)
- Healthcare Dive, Bayesian Health gets FDA nod for AI sepsis detection tool, May 12, 2026
- CIDRAP, FDA clears first AI-based early warning system for sepsis
- Wong et al., External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients, JAMA Internal Medicine, 2021
- UCLA Health, UCLA study finds AI scribes may reduce documentation time and improve physician well-being, November 26, 2025
- Lukac et al., Ambient AI Scribes in Clinical Practice: A Randomized Trial, NEJM AI, 2025 (figures as reported by UCLA Health; we did not read the full paper)
- Legal Information Institute, 45 CFR 164.502, Uses and disclosures of protected health information: general rules
- ICO, What is special category data? (updated 9 April 2024)
Frequently asked questions
Is this original research?
No. It is a digest of published studies, news reports on them and regulator pages, all listed in the Sources section. Zunkiree Labs did not run a survey or collect data for it.
Which clinical AI result is the strongest?
In this digest, the two randomized trials, the MASAI mammography trial and the UCLA ambient scribe trial, use the strongest designs. Each tested specific tools in specific settings and its authors or sources state limits.
Do the results mean AI is safe to use in clinical care?
No. The digest reports what the studies found and where they stop. Notes contained errors in the scribe trial, and an independent check found one widely used sepsis model missed most cases. Safety depends on the tool, the setting and the oversight.
How should a clinic use a digest like this?
Use it to ask better questions of suppliers: what evidence, from whom, measured how, and with what error and false-alarm rates. Then run a small supervised pilot.
What you get:
- A side-by-side table of three clinical AI evidence sets
- Plain-English summaries of the MASAI, sepsis alert and scribe findings
- The limits of each result as stated by the sources
- A short action list for clinics assessing AI tools
- Full source list with publishers and dates