OpenRounds Editorial
Daily Briefing
Tuesday, May 5, 2026
What Changed
Major medical journals now prohibit AI as an author while requiring transparent disclosure for AI-generated figures, establishing editorial standards that research teams must navigate as generative AI becomes routine in medical illustration workflows [1].
Research
•[AI in Medical Imaging] The MUST-Sub multimodal transformer integrates B-mode ultrasound morphology with contrast-enhanced ultrasound hemodynamics to classify breast cancer molecular subtypes, achieving macro-average AUCs of 0.94, 0.90, and 0.92 for Luminal, HER2-enriched, and triple-negative subtypes across internal, prospective, and multicenter external validation cohorts published in NPJ Digital Medicine [2]. Oncology programs gain a potential pathway to molecular subtyping without tissue biopsy, though the approach requires paired B-mode and CEUS imaging capabilities that may limit deployment scope in centers that do not routinely perform contrast-enhanced ultrasound.
•[AI in Clinical Operations] Combining genetic association data with machine learning improves type 1 diabetes risk prediction beyond traditional polygenic scoring approaches, according to research published in Nature Genetics [3]. Pediatric endocrinology and prevention programs evaluating AI-enhanced risk stratification now have a peer-reviewed framework to reference when assessing whether hybrid genetic-ML models offer meaningful lift over existing screening tools.
Policy & Ops
•[AI in Clinical Policy] The International Committee of Medical Journal Editors and journals including the Journal of Korean Neurosurgical Society prohibit AI as an author while permitting AI-generated conceptual figures accompanied by explicit disclosure, a policy framed around protecting research integrity and the particular privacy constraints that limit real patient photography in pediatric cases [1]. Research teams using generative AI for medical illustrations will need to implement human-in-the-loop editing workflows and disclosure language that satisfies these requirements across submission venues, not only in pediatric neurosurgery.
•[AI in Clinical Policy] A systematic review of 956 FDA-approved radiology AI devices published in NPJ Digital Medicine found that 30.65% of software-defect recalls were resolved through post-recall software updates, pointing to lifecycle planning gaps where corrections arrive reactively rather than through pre-market risk management [4]. Health system procurement teams and radiology AI vendors should treat post-market update patterns as a due-diligence data point when evaluating device maturity, given that adverse events related to software defects affected roughly a third of recalled products without triggering version updates.
One to Watch
•[AI in Clinical Operations] Eric Topol's analysis of medical AI implementation paradoxes examines the gap between demonstrated AI capabilities and actual clinical adoption, framing the core problem as insufficient implementation evidence rather than insufficient technical performance [5]. Research programs and funding bodies tracking this line of argument may find it increasingly difficult to justify purely benchmark-oriented study designs as the field shifts attention toward deployment conditions and real-world outcome measurement.