The Wire: 2 November 2026

The ten stories

1. Accuracy and impact of an AI safety system for missed specialist referrals in primary care npj Digital Medicine · 30 Oct

   Quasi-experimental, seven Singapore primary care clinics, ~25,000 notes scanned weekly: sevenfold detection increase. 428 flagged; 256 total detections (225 clinician-verified, 31 patient-reported). Blinded one-week comparison: AI PPV 61.8% vs human 51.8%. One week, not generalisable.

2. Understanding clinical reasoning variability in medical large language models: a mechanistic interpretability study npj Digital Medicine · 31 Oct

   GPT-5, MedGemma-27B and OpenBioLLM-70B; 3,035 perturbations of physician-validated oncology cases. Staging accuracy shifted over 50% from prompt format alone; definitive staging despite insufficient information. Takeaway: architecture-specific validation.

3. Predicting persistent fever in patients with cancer receiving antibiotics: a machine learning approach npj Digital Medicine · 31 Oct

   Retrospective. Internal 987 stays/794 patients (Essen); external 7,409/6,256 (Brigham and Women's). TabPFN AUROC 0.82 internal, 0.78 external; external overprediction suggests recalibration. Adjunct only; prospective safety untested.

4. GE HealthCare debuts Allia Celeo, a ceiling-mounted interventional system designed for smooth procedural flow and AI-powered precision imaging GE HealthCare · 1 Nov

   Company introduction. CE marked; US 510(k) pending; not for US sale. No patient-benefit evidence.

5. Interpretability-driven feature engineering enables compact and accurate multi-outcome risk stratification at hospitalization for chronic hepatitis B npj Digital Medicine · 31 Oct

   36,024 hospitalisation records/22,001 patients. SHAP-guided compact models (80.8% fewer features) classifying non-progression, cirrhosis and HCC recorded at admission - not future progression. External AUCs 0.829/0.865/0.827. No proven clinical benefit.

6. Standard and decoupled knowledge distillation for lightweight radiographic knee osteoarthritis grading and triage Scientific Reports · 31 Oct

   8,260 images/4,130 patients, patient-disjoint internal splits. Three-run mean binary accuracy 84.04% baseline vs 84.82% with distillation. No external cohort. Methods result.

7. Interpretable graph learning of spatial TME patterns in digital pathology for gastric cancer prognosis npj Digital Medicine · 30 Oct

   Prognostic association within TNM stages (validation n=1,871 plus TCGA-STAD n=213), not treatment benefit.

8. AnySleep: a channel-agnostic deep learning system for high-resolution sleep staging in multi-center cohorts npj Digital Medicine · 30 Oct

   Over 20,000 nights across 28 datasets; 13 trained/validated, 15 held out for testing. Fine-timescale staging evaluated indirectly. Model public. No demonstrated clinical outcome improvement.

9. Deep Learning-Assisted Diagnosis of Middle Ear Diseases in Otoendoscopy Across Clinician Experience Levels: Retrospective Reader Study JMIR Medical Informatics · 30 Oct

   13 clinicians, 100 images: junior accuracy 54.4% to 86.6%; expert-junior gap 24.3 to 3.1 points. Disease-enriched, single-centre set; unaided reading always first, so learning effects cannot be separated. Retrospective, not an outcome trial.

10. Cleerly Announces FDA Clearance of Latest Version of Cleerly LABS Cleerly (company announcement) · 30 Oct

    FDA clearance announced; automatic November rollout. CONFIRM2 training data 6,550 people/46 countries, ~50% female. Broader applicability is the company's claim, not comparative outcome evidence.