The Wire: 30 October 2026

The ten stories

1. Impression-R1 for radiological impression generation with multicenter benchmarking and prospective evaluations Nature Communications · 28 Oct

   A reasoning model for radiology impressions, tested on 11,884 cases from six institutions. In a randomised within-reader study of 360 interpretations, AI assistance improved report quality (difference 0.28, 95% CI 0.10-0.46) and cut completion time 27%; harmlessness did not differ significantly.

2. A clinical multimodal vision language foundation model with fine-grain explainability Nature Medicine · 29 Oct

   EVLF-FM, a vision-language foundation model trained on 1.3 million images across 10 modalities and six specialties, averaged 0.895 internal diagnostic accuracy; external testing used 13,443 samples across 17 datasets. Prospective clinical studies are required before deployment, the authors say.

3. Impact of Unrestricted Artificial Intelligence Scribe Deployment in a Single Health System on Clinician Workload and Productivity: A Quality Improvement Study Annals of Internal Medicine · 27 Oct

   Enrolling in the Heidi scribe cut note-active time 1.21 minutes per note across 1.64 million encounters, with no gain in closure or work RVUs. Companion evidence: npj Digital Medicine's 2,901-clinician adoption study and Leeds' TORTUS emergency department pilot saving 13 minutes per consultation.

4. An agentic AI system enhances clinical detection of immune-related adverse events in a multi-phase validation study Nature Cancer · 29 Oct

   An agentic LLM system extracted six immune-related adverse event types from oncology notes (macro F1 0.92 for current events). In a randomised crossover study of trial staff, assistance cut annotation time by 2.71 minutes and raised inter-annotator agreement (alpha 0.22-0.51 to 0.82-0.85).

5. Doximity Launches DoseCheck to Catch Dangerous AI Drug Dosing Errors Before They Reach Doctors Doximity (via Business Wire) · 29 Oct

   Doximity tested nine widely used AI models on more than 8,000 dosing questions from real physician queries: at least 2.3% of dosing claims were wrong and 43.8% of detected errors serious or potentially lethal. Its free DoseCheck layer verifies claims against FDA labelling and guidelines.

6. Deep Reasoning AI and Innolitics Achieve FDA 510(k) Clearance for NeuroDR, AI Software for Neurovascular CT Angiography Deep Reasoning AI / Innolitics (via Business Wire) · 27 Oct

   FDA granted 510(k) clearance (K262134) to NeuroDR, which removes bone, soft tissue and veins from head-and-neck CT angiography and returns 3D MIP, volume-rendered and curved planar reformats to PACS. Developed with 3DR Labs, which ran the performance studies.

7. Deep Breathe Receives FDA Clearance for the First AI Dedicated to Detecting Absent Lung Sliding Deep Breathe (company announcement via EINPresswire) · 28 Oct

   FDA cleared Deep Breathe's AI that flags absent lung sliding on handheld ultrasound, a warning sign of pneumothorax. The Ontario company calls it the first clearance for AI dedicated to this finding and its second in 12 months.

8. OpenEvidence and NYU Langone Health Partner to Bring Patient-Aware Intelligence to 12 Million Visits in New York City NYU Langone Health · 29 Oct

   OpenEvidence, which calls itself the most-used AI among US doctors, is being embedded in NYU Langone's Epic workflow, giving clinicians literature-grounded answers with patient context across a system with 12.4 million outpatient and 421,000 emergency visits a year.

9. Google AI for Healthcare Professionals Certificate launched Google (company blog) · 29 Oct

   Google's self-paced certificate teaches clinicians to use AI for administrative tasks, patient communication and research organisation. It takes under five hours and carries up to 5.0 AMA PRA Category 1 credits; NYU Grossman, UAB, Touro and Purdue's nursing school are offering it.

10. Development and validation of an artificial intelligence model for automated classification of whole-body bone scintigraphy in oncology patients Scientific Reports · 29 Oct

    Trained on 5,040 oncology bone scans at one Thai hospital, OsseoScan AI classified normal versus abnormal scans with test AUC 0.963, sensitivity 0.907 and specificity 0.886; sensitivity fell to 0.725 for solitary metastases. External and prospective validation is still required.