The Wire: 12 November 2026

The ten stories

1. S-Detect and ChatGPT give early-career breast US readers little lift Academic Radiology · 10 Nov

   Retrospective study of 299 biopsy-proven breast lesions: the senior radiologist still led on sensitivity and accuracy; S-Detect topped specificity and matched the expert statistically (kappa 0.686 vs 0.713), while ChatGPT lagged (0.373). Neither assistant consistently improved junior readers.

2. Automated EHR tracking lifts colonoscopy follow-up updates 83% to 94% JAMA Network Open · 10 Nov

   Vanderbilt cohort of 35,802 adults: after adenomas were found on colonoscopy, a decision support tool that auto-populates the EHR health-maintenance tab raised correct surveillance tracking from 83% to 94%. Single centre, before-after, but a cheap fix for a known cancer-prevention gap.

3. Synthetic-only training matches real images for reading rapid tests npj Digital Medicine · 10 Nov

   The SynSight pipeline trained entirely on synthetic images, with no real-world library: 98% sensitivity and 99% specificity on HIV rapid tests and up to 99% accuracy on COVID-19 tests. A route to machine learning for low-resource settings where proprietary image libraries block entry.

4. Open-source AI scribe runs in-house for under $30 per doctor a month JMIR Medical Informatics · 10 Nov

   Berta, a modular open-source scribe hosted entirely on institutional infrastructure, reports running costs under US$30 per physician per month, 70-95% below commercial subscriptions. An implementation report, not an outcome trial, but a procurement-relevant data point on cost and sovereignty.

5. Ohio State and Microsoft sign five-year smart hospital build-out HIT Consultant · 10 Nov

   Five-year partnership: an Azure and IoT smart-hospital model live at two sites, 70-80% of storage and compute moving to Azure, cloud backup for the EHR, and Dragon Copilot ambient documentation at the bedside. The plumbing phase of hospital AI, at enterprise scale.

6. Generative AI went from zero to one in five registered trials by 2025 JAMA Network Open · 9 Nov

   Analysis of 1,212 interventional trials registered on ClinicalTrials.gov starting from 2022: planned AI use spread well beyond radiology and gastroenterology, and generative AI rose from no trials in 2022 to one in five in 2025. Methods reporting in registrations remains patchy.

7. Mental health chatbot users bond early but sustained use unproven JMIR Formative Research · 9 Nov

   Exploratory longitudinal study of 527 self-referred users of the LLM-based chatbot Clare: high baseline distress and an early subjective working alliance, but evidence on sustained engagement and outcomes remains limited. Design signals, not proof of benefit.

8. Glasgow ED randomises AI chest X-ray tool on and off, week by week Newton's Tree · 9 Nov

   SEARCH-ED: a CE-marked chest X-ray AI flagging signs of heart failure, COPD and lung cancer is switched on or off for the whole emergency department in randomised one-week blocks, giving NHS Scotland real-world evidence before a national adoption decision. A pragmatic trial design worth watching.

9. Barnsley awards ambient voice contract; NW London weighs its own HTN Health Tech News · 9 Nov

   Barnsley Hospital NHS FT awarded an ambient voice technology contract to Lyrebird Health running November 2026 to February 2028, while the North West London acute provider group opened market engagement for a 2027-2030 deal. NHS ambient scribes are now procurement, not pilots.

10. Automated AI coronary CT read agrees with experts mainly on normals The International Journal of Cardiovascular Imaging · 7 Nov

    Single-centre screening cohort of 172 people and 2,582 coronary segments: significant stenosis was rare (0.7%). AI-radiologist agreement was high for ruling stenosis out (negative agreement 0.992) but only fair for grading and plaque type. Supports supervised triage, not autonomous reads.