The Wire: 8 November 2026

The ten stories

1. GPT-5 reads clinic notes to explain unescalated lipid therapy JAMA Cardiology · 7 Nov

   28,113 patients with ASCVD, half above LDL-C goal: a GPT-5 pathway reading clinic notes found treat-to-target strategies in 28.1% of notes and no escalation in 88.7%, mostly adverse effects, patient preference, near-goal values or deferred retesting. A documented target tripled escalation.

2. Gait foundation model reads whole-body health from 3D movement Nature Medicine · 6 Nov

   3,328 adults, one depth camera, five motor tasks: learned movement embeddings beat engineered gait features for age (r=0.69), BMI (r=0.91) and visceral fat (r=0.83), significantly predicted 1,887 of 3,210 phenotypes across 18 body systems, and added signal beyond age, BMI, fat and height alone.

3. AI scribes speed notes most for non-English-concordant visits JAMA Network Open · 6 Nov

   209,009 encounters across 286 clinicians: AI scribe use tracked with same-day note closure gains of 10.2 percentage points (95% CI 7.4-12.9) when clinician and patient shared a non-English language, versus 1.7 points (0.9-2.6) for English visits; interpreter visits inconclusive.

4. Nurse LLM dependence grows with use; training tracks with less BMC Nursing · 6 Nov

   459 nurses in six Chinese hospitals: functional dependence on LLMs scored higher (3.94/5) than existential dependence (3.02/5); heavier use tracked with higher scores, related training with lower. Authors caution the scores set no validated severity levels or prevalence of problematic dependence.

5. AI safety net catches all remarkable brain CTs in validation npj Digital Medicine · 5 Nov

   External validation on 2,001 emergency and trauma non-contrast brain CTs: commercial detectors alone left 31 of 324 remarkable scans wrongly deprioritized; a BrainCT-CLIP generalist plus specialist detectors caught all (100% sensitivity/NPV), deprioritizing 18% (95% CI 17-20). Retrospective.

6. Updated FDA-cleared AI devices rarely report clinical benefit JAMA Network Open · 5 Nov

   56 updates to FDA 510(k)-cleared AI devices: 75% changed how the device is used while only 16% improved the algorithm; data-source reporting was absent in 42.9% and 85.7% reported no clinical effectiveness. A cleared update is not evidence of patient benefit.

7. FDA classifies mild cognitive impairment prognostic software Federal Register · 5 Nov

   Final rule (91 FR 69690) puts prognostic assessment software of mild cognitive impairment into Class II with special controls under 21 CFR 882, effective 5 Nov - one of three 2024 De Novos codified the same day, each creating a 510(k) predicate pathway for follow-on devices.

8. Foundation model reads coronary angiograms, predicts events Nature Cardiovascular Research · 5 Nov

   News and Views on DeepCORO-CLIP: a video-text foundation model generating angiography reports that align more closely with experts than routine single-reader reports and predict 1-year cardiovascular events - non-randomized two-center study; commentary, not the trial itself.

9. Map splits digital heart care into deploy vs invest-first Nature Medicine · 5 Nov

   82,999 US census tracts mapped: in high-burden communities with no or shrinking cardiology supply, 18% of tracts (19.9 million residents) are digitally ready for deployment now while 82% (74.1 million) need infrastructure or adoption investment first - often in tracts next door to ready ones.

10. ChatGPT lifts Nigerian frontline workers' patient answers Journal of Medical Internet Research · 5 Nov

    36 Lagos frontline workers answered 15 real patient questions with and without ChatGPT-3.5 (1,080 responses): clinician raters scored GPT-aided answers higher on every metric (P<.001), most for completeness and empathy. Observational; rated answers, not patient outcomes; vendor-affiliated team.