OpenRounds Editorial
Daily Briefing
Sunday, July 26, 2026
What Changed
NHS England announced £10 billion over three years to accelerate AI rollout across its hospitals, and a BMJ commentary published alongside the pledge argues the evidence base still needs building before deployment can be justified at scale [1].
Policy & Ops
•[AI in Clinical Operations] The BMJ commentary frames NHS England's £10bn commitment as a promise that outpaces the evidence supporting broad AI use in clinical care [1]. NHS-affiliated trusts will receive capital earmarked for AI tools, but the commentary signals that local outcome data — not the funding announcement — will determine whether deployments improve waiting times and care quality.
•[AI in Clinical Operations] The American Geriatrics Society issued a position statement identifying polypharmacy, capacity-sensitive consent, and functional assessment as geriatrics care tasks where generative AI risks are amplified, and calls on LLM developers to embed bias-mitigation safeguards for these scenarios [2]. The statement signals what geriatrics leaders may demand from clinical decision support and agentic systems before adopting them — a domain-specific risk checklist vendors can expect to encounter when pitching to older-adult care programs.
Research
•[AI in Medical Imaging] Across 1,911 thyroid nodule ultrasound images read by seven radiologists under three major guidelines — ACR-TIRADS, ATA, and C-TIRADS — AI assistance improved diagnostic accuracy for intermediate-suspicion nodules from a range of 66.4%–71.4% to 74.1%–79.4% [3]. The intermediate-suspicion category is where clinicians most often disagree on whether to biopsy; gains that hold across all three guidelines and all seven readers suggest the tool addresses a genuine ambiguity point rather than inflating one benchmark.
•[AI in Clinical Practice] Workflow design determines whether AI-based autism screening tools fire at all, according to a JAMIA Open observational study of 8 clinicians and 20 caregivers at Duke-affiliated well-child visits [4]. The study found that electronic screening integrated into the visit, early-intervention provider input, and staff referral coordination were the conditions that made the tool usable — more actionable for deployment planners than another accuracy metric.
One to Watch
•[AI in Clinical Practice] UniPert-G2CP, a two-stage deep learning framework published in Cell, unifies genetic and chemical perturbation data so that phenotype predictions trained on one modality can transfer to the other [5]. The work addresses a core obstacle in building an AI virtual cell — the disparity between perturbagen modalities and assay formats — but remains a methods paper about cross-modal transfer, not a deployed therapeutic tool.