Healthcare AI Daily Briefing

Medical Imaging

Lunit AI matches double-reader AUC in Singapore screening but misses microcalcifications.

Journal article (Breast disease)

In a retrospective Singapore screening cohort of 1,982 women (2010–2017), Lunit AI achieved an AUC of 0.874, outperforming single readers (0.811) and matching double readers (0.863), with higher specificity (81.0% vs 72.6%) but lower sensitivity (80.3% vs 89.6% and 100%). The double-reader 100% sensitivity reflects the retrospective design, and Lunit missed 34 cancers, most being microcalcifications, suggesting an adjunct role rather than independent reading. [1]

Industry & Products

NEJM study finds AI care plans cost substantially more than physician plans.

A New England Journal of Medicine study of 24 AI systems on 200 clinical notes found general-purpose LLMs' plans averaged $314 more per visit than physicians', while specialized medical AI recommended care costing 2.7 to 3.5 times as much. Extra costs came mainly from diagnostic testing and referrals: in 66.5% of visits where physicians ordered no workup, every AI system suggested testing, though nearly a fifth of AI plans aligned with guidelines and caught omitted workups. [2]

Industry & Products

OLIS clinical knowledge system reaches 90% of acute NHS trusts and 500 global sites.

Podcast from The Signal Room

The interview reports that OLIS, an AI system indexing hospital formularies, policies, and guidelines for point-of-care clinician queries, is now used by 90% of acute NHS trusts and approximately 500 sites worldwide including Stanford, Mass General, and Boston Children's Hospital. The system launched roughly four years ago. [3]

Operations & Workflow

Cleveland Clinic and Mayo Clinic leaders detail AI revenue-cycle deployment at AHIMA.

In an AHIMA conference interview, Nick Judd of Cleveland Clinic and Stacey Bui of Mayo Clinic said both organizations have implemented AI for revenue-cycle management, with all three leaders — including AKASA's Dr. Nirav Kamdar — citing routine-case automation to free staff for complex work as the primary impact. Kamdar stressed that AI must be customized to local workflows and prioritize accuracy and compliance, while Judd and Bui highlighted the need for tighter collaboration between revenue and clinical teams, which they described as sometimes at odds. [4]

Clinical Practice

Opinion: isolating AI agents like contract nurses sacrifices trust and longitudinal learning, author argues.

In a MedCity Influencers opinion piece, Hygia Chief Clinical Officer Gennady Gekht argues that Imprivata CEO Fran Rosch's analogy of AI agents to contract nurses misses the point by prioritizing cybersecurity isolation over patient engagement. Gekht writes that treating AI as temporary, self-limiting transactions minimizes risk but also prevents the longitudinal learning and trust-building essential for chronic care, where patients need consistent, familiar companions to share information and follow recommendations. He contends that an 'adopt and drop' model creates repeated new risks, while gradually expanding a monitored AI's responsibilities builds stronger safety and trust. [5]

Research & Evidence

Open-weights LLMs match or exceed GPT and Claude for clinical note labeling in 46-model benchmark.

Journal article (JMIR)

A retrospective study tested 46 LLMs — including GPT and Claude families and 42 open-weights models from Llama and Qwen families — on labeling 380 clinical notes from two U.S. institutions for 29 Elixhauser comorbidity conditions. The top performers, including Llama 3.3 70B, Llama 3.1 405B, and Qwen 2 72B, achieved F1-scores above 70%, performing on par with or better than closed-source models. Performance improved significantly with model size and recency, mainly through gains in precision. Results varied widely across the 29 conditions. The study used manual labels by a radiology resident and medical student under senior radiologist supervision as the reference standard. [6]

Research & Evidence

Harvard summit consensus calls for AI curriculum, dedicated leadership, and deployment across dental education.

Journal article (Journal of dental education)

At the 2026 Reimagining Dental Education Summit at Harvard University, eight academic and organization leaders reached unanimous consensus on recommendations: create an AI and technology curriculum across all dental education levels; establish a dedicated leadership role to develop integration strategies for AI and emerging technologies; and deploy AI to improve institutional effectiveness, faculty calibration, patient care, and student learning. For practice readiness, they recommended pathways to reduce time to degree, enhance external clinical experiences, and expand incentivized postgraduate residencies. The working group, including two co-chairs and a summit scholar, used future-back facilitated discussions and a final report-out process. [7]

Medical Imaging

Systematic review: Multimodal AI outperforms unimodal models for oral cancer diagnosis and prognosis.

Journal article (Healthcare technology letters)

A PRISMA-guided systematic review of 82 studies from 2015 to 2025 found that multimodal AI models — integrating histopathology, radiology, genomics, and clinical data using convolutional neural networks, transformers, and graph neural networks — consistently outperformed unimodal approaches in diagnostic accuracy, AUC, and survival prediction. Architectures employing attention mechanisms, graph embedding, and fusion transformers showed superior performance for early lesion detection and treatment outcome modeling. The review notes persistent challenges in generalizability, data standardization, and clinical validation, calling for prospective validation and ethical deployment research. [8]