The Wire: 5 November 2026

The ten stories

1. Cross-paradigm audit of large language model judges for clinical text evaluation npj Digital Medicine · 4 Nov

   Six open-weight LLM judges audited against expert panels on four German clinical datasets. Swapping in an AI judge shifted 57.2% of pointwise categories, swapped 41.7% of pairwise winners and missed 10.2% of clinical errors; recall 0.13 to 0.90. Position and identity bias in every judge.

2. AI methylation model personalizes childhood brain tumour prognosis Nature Genetics · 3 Nov

   MANTIS predicts 10-year survival from methylation data (2,786 patients). Retrospective re-analysis: in a model-selected ~60% of average-risk medulloblastoma, no detected survival disadvantage with lower-dose irradiation (P=0.683) - not proven safe dose reduction; prospective trials needed.

3. Generative AI and Maternal Health Outcomes in Pregnant and Postpartum Individuals: A Randomized Clinical Trial JAMA Network Open · 4 Nov

   400 participants from minoritized racial and ethnic groups, 12 months: 'Rosie' genAI maternal-health chatbot vs monthly children's books. Null primary: anxiety AMD 0.10 (95% CI -0.77 to 0.96), depression AMD -0.23 (-1.05 to 0.59); no difference in ED visits, postpartum care, contraception or sleep.

4. Silent steering of an LLM shifts mental health treatment advice JAMA Network Open · 3 Nov

   Proof of concept: inference-time activation steering of an open-weights LLM (DeepSeek v4 Flash) shifted depression advice between medication and self-care, invisibly to the user. 372 responses, one model; closed models untested. Responses scored by an LLM rater (ICC 0.81 vs a blinded human).

5. TEMPO pilot: a regulatory middle path for AI mental health devices npj Digital Medicine · 3 Nov

   Commentary: the FDA-CMS TEMPO pilot lets selected digital mental health devices commercialize in Medicare under enforcement discretion - a working QMS but no premarket authorization - while generating real-world data toward clearance. Up to 10 participants in the mental health track.

6. FDA final rule: Class II special controls for AI digital cervical cytology imaging Federal Register · 5 Nov

   Final rule classifies the digital cervical cytology slide imaging system with AI algorithm into Class II with special controls (21 CFR 864.3900) - the second named AI device type codified within a week, each creating a 510(k) predicate pathway.

7. RETBind unifies OCT and fundus imaging in one retinal AI model npj Digital Medicine · 3 Nov

   Patient-level paired pretraining on OCT plus ultra-widefield fundus images in one shared encoder; consistent gains over single-modality approaches across nine ocular classification and segmentation benchmarks, especially AUPR in imbalanced settings.

8. Athenahealth connects patient records to Anthropic's Claude via MCP Healthcare IT News · 3 Nov

   Patients can authorize Anthropic's Claude to query their athenahealth records - labs, medications, conditions, visit history, care plans - via an open-standard Model Context Protocol server, with revocable access, to help them prepare for conversations with their care teams. Announced 2 Nov.

9. GE HealthCare adds AI lung and aortic stenosis tools to Vscan Air GE HealthCare · 3 Nov

   Company announcement: Lung Guidance and AutoAS are FDA 510(k)-cleared and CE marked for Vscan Air SL. In validation, non-experts acquired diagnostic-quality lung exams at 95.2% success vs 96.8% for experts (descriptive). AutoAS flags suspected moderate-to-severe aortic stenosis.

10. Parents wary of AI-led behavioural therapy for their children JAMA Pediatrics · 2 Nov

    US national survey (572 parents, November 2025): 273 not comfortable at all with AI-led behavioural therapy for their child; only 129 somewhat or very comfortable. Comfort tracked with parents' own AI use. Nonprobability sample, so not nationally representative.