Healthcare AI Daily Briefing

Policy & Governance

AMA tells FDA developers should monitor generative AI devices, not providers.

In comments to the FDA, the AMA and other provider groups argued that AI developers, not healthcare providers, should be responsible for monitoring generative AI-enabled medical devices, saying the burden would strain small and rural organizations. The AMA's framework also emphasizes that users need access to information about intended use, performance, and limitations rather than just warnings and disclaimers, and a 2026 survey found a quarter of physicians received no AI training while 92% wanted more. [5]

Medical Imaging

Deep learning CT changes after one year pirfenidone predict IPF outcomes.

Journal article (Insights into imaging)

In 436 IPF patients across two hospitals, each 5% increase in preserved normal lung extent on deep learning CT quantification after one year of pirfenidone associated with lower composite mortality-transplant risk (adjusted HR 0.90, 95% CI 0.85-0.96), while each 5% fibrosis increase predicted higher risk (HR 1.16, 95% CI 1.06-1.27). Fibrosis change alone predicted ≥10% FVC decline (HR 1.26, 95% CI 1.09-1.45). [1]

Research & Evidence

Machine learning model predicts severe fetal acidemia from CTG and clinical data.

Journal article (American journal of obstetrics & gynecology MFM)

A bicenter study of 3,574 deliveries found a machine learning model combining intrapartum CTG signals and clinical variables achieved AUC 0.78 (95% CI 0.71-0.85) for severe neonatal acidemia (pH <7.05) using the final hour before birth, outperforming clinical practice sensitivity at the same false-positive rate. CTG-derived features drove predictions while clinical variables added limited value. [2]

Operations & Workflow

LLM extracts explicit OA pain scores; narrative inference matches physician estimates.

Journal article (Arthritis care & research)

In a retrospective analysis of 163 OA clinic notes, a prompted Llama 3.2-90B model achieved strong agreement with abstracted pain scores in internal medicine notes (τ=0.76) and excellent in orthopedic notes (τ=0.95), also flagging 15 EMR default zeros. For notes without explicit scores, LLM inference correlated weakly with registry scores but matched physician estimates, indicating insufficient documentation rather than model failure. [7]

Industry & Products

FDA receives 500+ comments on generative AI device regulation framework.

Stakeholders submitted contrasting views on the FDA's proposed generative AI framework: AdvaMed urged building on existing premarket pathways and risk-based postmarket monitoring, while physicians and patients demanded stronger safeguards and manufacturer-led postmarket surveillance. Oura sought wellness exemptions for informational outputs, Generation Patient opposed replacing premarket evidence with postmarket signals, and Dr. Vasan recommended separate pass/fail criteria for high-consequence failures like missed sepsis or suicide risk. [4]

Research & Evidence

Multimodal AMD agent links imaging to guidelines for longitudinal decision support.

Journal article (npj Digital Medicine)

In a 20-case proof-of-concept, AMDagent integrated quantitative retinal imaging analysis with guideline retrieval to produce structured decision-support outputs rated clinically appropriate in 92% of predefined units under expert adjudication. Eighty-three percent of audited major clinical claims were explicitly grounded in tool-derived measurements or retrieved guideline evidence, enabling clinician verification, though larger prospective studies are needed. [8]

Industry & Products

LLMs reference correct dosing info but compute wrong math for complex patients.

A pediatric nephrologist testing six LLMs on antibiotic dosing for dialysis and kidney-impaired patients found models often cited correct guidelines but then miscalculated doses, with errors that could prove fatal. The findings suggest fit-for-purpose tools or improved prompting approaches are needed rather than general-purpose chatbots for clinical dosing decisions. Correction (2026-11-05): The physician's specialty has been corrected to pediatric nephrology. [3]

Clinical Practice

Interview: Pediatricians should ask about AI chatbot use in history taking.

Podcast from Advancing Health

The interview describes incorporating questions about AI LLM use into routine history taking for youth, similar to how social media use is assessed, to understand patients' full experience world. The account suggests this applies especially to pediatricians and youth-serving health professionals. [6]