5 AI Medical Advances Backed by Published Evidence

Claims about AI in medicine range from sober trial results to marketing. This article keeps to five areas where we could read the published evidence ourselves, in a journal abstract, a regulator's page or a developer's own release. For each we say what was studied, how large it was, what was found, and what it does not show.

1. Did AI help find more breast cancers in routine screening?

What was studied. PRAIM is an observational, multicentre, real-world study of AI-supported double reading in organised mammography screening for women aged 50 to 69 at 12 German sites. It compared AI-supported reading with standard double reading. Radiologists chose voluntarily whether to use the AI system.

Size and result. From July 2021 to February 2023, 463,094 women were screened, 260,739 of them with AI support, by 119 radiologists. The detection rate was 6.7 per 1,000 with AI against 5.7 per 1,000 without, which is 17.6% higher (95% confidence interval +5.7% to +30.8%). The recall rate was 37.4 per 1,000 against 38.3, which the authors describe as non-inferior. The positive predictive value of recall was 17.9% with AI and 14.9% without. The paper is in Nature Medicine, volume 31, March 2025.

What it does not show. Because radiologists chose whether to use the tool, the two groups may differ in ways other than the AI. A higher detection rate does not show fewer deaths, and it can include cancers that would never have caused harm. The record we read lists an AI developer, Vara MX Healthcare, among the authors' affiliations and shows no conflict-of-interest statement, so check the paper's own disclosures. For a randomised trial in the same field, see the MASAI result in our human stories article.

2. Does AI help doctors miss fewer polyps during colonoscopy?

What was studied. A systematic review and meta-analysis in the World Journal of Gastroenterology (2025, volume 31, issue 21) pooled randomised trials of computer-aided detection, in which software highlights suspicious areas on the live colonoscopy image. The review is registered in PROSPERO and searched PubMed, Embase and CENTRAL through August 2, 2024.

Size and result. It included five trials with 1,624 patients. The adenoma miss rate was 15.9% with the software against 35.9% with routine colonoscopy (odds ratio 0.34, 95% CI 0.26 to 0.45). Overall adenoma detection was not significantly different (42.3% against 41.3%, odds ratio 1.03, 95% CI 0.79 to 1.34), though detection on the first pass was higher (43.2% against 36.5%). The authors rated the certainty of the evidence as moderate.

What it does not show. The benefit was in small lesions of 5 mm or less and in sessile serrated lesions. For advanced adenomas the difference was not significant. The trials used a tandem design, in which each patient is examined twice, which the authors note does not reflect routine practice. Only one trial was double-blind, and the authors say the long-term effect on colorectal cancer prevention is unconfirmed. The paper's own numbers are slightly inconsistent: patient counts in one table sum to 1,633, not 1,624.

3. Has an AI-discovered drug shown benefit in patients yet?

What was studied. GENESIS-IPF is a phase IIa, double-blind, placebo-controlled trial of rentosertib, a drug for idiopathic pulmonary fibrosis (a scarring lung disease) that its developer, Insilico Medicine, says was discovered and designed with an AI approach. Results were published in Nature Medicine on June 3, 2025. We read the company's release, not the paper.

Size and result. It enrolled 71 patients at 22 sites in China, who received placebo or one of three dose regimens for 12 weeks. The release says the trial met its primary endpoint of safety and tolerability, with similar rates of treatment-emergent adverse events across groups. For forced vital capacity (a measure of lung function, reported as a secondary endpoint), the 60 mg once-daily group changed by +98.4 mL against -20.3 mL for placebo.

What it does not show. Twelve weeks and 71 patients cannot show lasting benefit. The company's lead investigator is quoted saying the sample per group was relatively limited and needs validation in larger cohorts. The release gives no lung-function figures for the lower-dose arms, and it describes only that Insilico has begun discussions with regulators about larger studies. Being AI-discovered says how the molecule was found. It does not make the trial result stronger.

4. Do AI note-taking tools in clinics reduce burnout?

What was studied. A randomised crossover trial in the Journal of the American Medical Informatics Association (volume 33, issue 5, May 1, 2026) compared two ambient AI scribe products, which listen to a consultation and draft the clinical note, at one tertiary academic medical center. It was open-label, so clinicians knew which tool they used.

Size and result. 160 outpatient clinicians volunteered and were randomised into two groups of 80, and surveys from 136 respondents were analysed. Both tools improved workflow satisfaction and reduced personal and work burnout scores on the Copenhagen Burnout Index, but the differences between the tools were not meaningful on burnout. Satisfaction improved by 2.51 points on a 1 to 7 scale for one tool and 1.91 for the other (mean difference 0.60, 95% CI 0.32 to 0.90). Time in notes differed by 3.19 minutes a day between the products (95% CI -4.87 to -1.50).

What it does not show. The design compares two products, and the abstract we read does not describe a comparison with no scribe at all, so it cannot say how much of the burnout change is due to the technology. Participants were volunteers, and the authors link the unchanged efficiency measures to participant selection and the study period's timing. Time spent on notes after hours showed no meaningful difference. Because these tools process patient conversations, see our guide to questions about AI data retention.

5. What does the FDA record about AI-enabled devices?

What it is. The FDA publishes a list of AI-enabled medical devices that are authorised for marketing in the United States, intended to help innovators, providers and patients see where AI is used. The page says listed devices have met the FDA's applicable premarket requirements, which include a focused review of safety and effectiveness. It says the list will be updated periodically, and the latest decision date visible in the portion we read was June 29, 2026.

What it does not show. The FDA states plainly: "The list is not a comprehensive resource of AI-enabled medical devices." Devices were mainly identified by AI-related terms in their authorisation summaries, so some may be missing. The page we read gave no total count, and the agency says it will explore ways to identify devices that use foundation models, including large language models. An authorisation is a regulatory decision. It is not a randomised trial showing better outcomes for patients, which is why the first four sections matter.

How can I read a medical AI claim for myself?

What are the limitations of this article?

We read abstracts, a repository record, a regulator's page and one company release. We did not read the full text of the Nature Medicine, JAMIA or World Journal of Gastroenterology papers, so details outside those abstracts are not covered. Five items are a small selection, not a survey of AI in medicine, and we chose them because their evidence was readable, which favours areas with more published research. Candidates that appeared only in search-engine summaries were left out. Rankings of benefit would need a published test, which we have not run. See our methodology and corrections pages.