Healthcare AI Daily Briefing

Medical Imaging

CAPTOR vision‑language model beats baselines in oncologic CAP CT classification.

Preprint

CAPTOR was trained on 25,512 multicenter chest‑abdomen‑pelvis CT‑report pairs from North America, Europe and Asia and evaluated on 10,794 held‑out pairs. On the CAP subset it achieved AUC 0.767, exceeding the next‑best model’s AUC 0.695 on TumorImagingBench, which comprises up to 421 studies per task. [1]

Operations & Workflow

Preprint shows ProCare beats LLM agents on 36 EHR benchmark tasks.

Preprint

ProCARE converts data profiles into Evidence Contracts and compiles study objectives into an intermediate representation (RWS‑IR). Deterministic validation and localized repair check workflows cohort construction, variable derivation, analysis, and interpretation. On 36 EHR‑based benchmark tasks, ProCARE exceeds strong LLM‑agent baselines in code executability, semantic consistency, and conclusion traceability, with automated violation diagnoses showing moderate‑to‑substantial agreement with blinded expert review. [2]

Policy & Governance

Lobbyists shape state medical AI regulation behind closed doors.

In October a 22‑member workgroup convened by the Federation of State Medical Boards—including AMA, AHA, Anthropic and Elevance Health—met to draft guidelines for AI tools that perform clinical functions with limited or no direct physician supervision. The confidential discussions aim to produce recommendations that could become the basis of state legislation regulating medical AI. [5]

Operations & Workflow

Correction: Nursing documentation claim removed.

Correction, October 15, 2026: We removed a claim about Intermountain nurses’ documentation time because we could not verify its attribution and scope from the cited source. [3]

Policy & Governance

Interview: developers should bear liability for flawed medical AI.

Podcast from AI and Healthcare

The interview describes that developers should be liable for flawed autonomous AI, citing their deep pockets, role in creating the systems, potential inadequate guardrails, and the need for them to act cautiously and learn from problems to improve their programs. [4]