S. Menon · Radiology

‹ The Wire archive

The Wire · 16 Nov 2026

The Wire: 16 November 2026

Ten developments in healthcare AI, summarised from published sources and read by an AI voice. News, not medical advice.

Listen

The ten stories

  1. Persona prompts swing LLM clinical decisions on identical facts npj Health Systems · 14 Nov

    Mount Sinai team, 20 open-weight models and 5 million decisions over 5,000 emergency and discharge texts: baseline yes-rate 42.8%, and simply assigning a decision-style persona swung answers 36.9-46.4% (9.5 points) with facts unchanged. Model size and medical fine-tuning gave no protection.

  2. Compact stroke AI reads CT and MRI, and says where it is unsure npj Digital Medicine · 14 Nov

    DFE-Net, built to represent spatial uncertainty: multicentre cohorts of 6,650 CT patients (8,873 images) and 67 MRI patients reached AUROC 0.974 (CT) and 0.978 (MRI), Dice 0.70 lesion localisation on CT, from a 0.2-million-parameter model. Retrospective; prospective validation still needed.

  3. Physicians rarely notice when a trusted AI starts failing npj Digital Medicine · 13 Nov

    Randomised experiment, 48 physicians classifying heart sounds with AI help as its reliability fell in steps from 90% to 0%: only 12.5% spontaneously noticed, and even asked directly just 37.5% did. Cognitive forcing did not rescue oversight. The human safety-net assumption takes a hit.

  4. Doctors would act on 68% of AI abstracts, false ones included Journal of Medical Internet Research · 13 Nov

    Vignette study, 20 emergency and trauma physicians in 4 German hospitals rating GPT-5-written abstracts, half with planted false claims: intent to act on 68%; true vs false told at 78% sensitivity and only 43% specificity, and false abstracts read as credible as true ones. Small sample, big signal.

  5. DeepSeek matches GPT on records but folds to prompt attacks Journal of Medical Internet Research · 13 Nov

    Multicentre test, 50 real records from 10 hospitals, 14 specialties, 7 models: DeepSeek-R1 and GPT-4o-reasoning tied on diagnostic accuracy (64%). But under adversarial prompting every model produced medical disinformation - DeepSeek-R1 in 98% of attacks, the best refuser still in 70%.

  6. FDA clears Bracco-Avicenna AI for brain aneurysms on CTA Bracco (press release) · 13 Nov

    510(k) clearance for UPTUNITY | Aneurysm: flags suspected saccular aneurysms over 3 mm on head CT angiography as a concurrent reading aid, with quantitative follow-up comparison for known aneurysms. First of Bracco's UPTUNITY portfolio with Avicenna.AI; US launch and RSNA showcase to come.

  7. Heidi signs Samsung Medical Center as first Korean partner The Asia Business Daily · 13 Nov

    MOU signed 12 Nov: Heidi's ambient documentation will integrate with Samsung Medical Center's Darwin system, with telemedicine and AI-agent plans. Vendor-reported scale: 3 million+ consultations a week in 110 languages, 190 countries. Australia's scribe lands in Korea while the NHS buys Lyrebird.

  8. Reinforcement learning lifts AI chest X-ray report quality Nature Communications · 12 Nov

    UNIRG-CXR, an 8-billion-parameter vision-language model trained with reinforcement learning against a clinical-error reward, beat the previous best ReXrank score by over 50% on two test sets, with fewer clinically significant errors and steady performance on unseen hospitals and demographics.

  9. Autonomous AI discharges a third of skin-cancer referrals safely Chelsea and Westminster Hospital NHS FT · 12 Nov

    Real-world pathway, 8,391 urgent suspected skin cancer referrals at two hospitals: AI autonomously discharged 31% and 25%, routine follow-up fell 27% to 12%, biopsies 27% vs 43% face-to-face, saving 2,851 clinician hours (about 62% capacity). Congress-presented, not yet peer-reviewed; 86% consented.

  10. 1,600 FDA-cleared AI devices, almost none prove patient benefit Journal of Medical Internet Research · 11 Nov

    Analysis: 1,614 AI devices greenlit by the FDA as of October 2026, yet almost none carry evidence they improve patient outcomes - an earlier PLoS One review found only 3 of 1,357 devices assessed any patient outcome. Clearance regulates safety and performance claims, not clinical benefit.