Healthcare AI Daily Briefing

Industry & Products

Charité builds AI platform Clay to retain control over data and models.

Podcast from KFF's The Business of Health with Chip Kahn

The interview describes how Charité developed its own AI platform, Clay, after finding no commercial product that met its combination of on-premises deployment, organization-wide access for sensitive patient data, and control over model selection and clinical integration. The platform runs entirely on Charité's servers; the speaker states they cannot accept a situation where a critical capability is turned off, so they need to decide how data are used, which models are deployed, and how applications connect to clinical processes. [1]

Operations & Workflow

AI scribes shift clinician burden from typing to cognitive oversight, author argues.

In an opinion piece, Dr. Ambrén Zaman Riaz, MD, MPA, argues that ambient AI scribes solve the visible pain point of documentation but do not reduce overall clinician burden — they move it upstream into verification, billing compliance, and anticipation of how payers and auditors will read the note. Riaz writes that ROI stories frame value around minutes saved documenting, yet harder measures such as after-hours charting, inbox volume, and decision fatigue often do not improve. She contrasts this with upstream intervention: in clinical trial eligibility screening, placing AI at the step where research staff once combed charts by hand removes most of that process altogether rather than accelerating it, which she says should be the standard for clinical AI tools. [2]

Policy & Governance

Google publishes health AI evaluation framework with expert partners.

Google Health released a whitepaper co-authored with Dr. Elizabeth McGlynn and Dr. Dana Gelb Safran detailing an end-to-end evaluation approach assessing models across accuracy, safety, and usefulness using closed-question benchmarks, autoraters, and expert human review for common and high-risk health scenarios. The company states it uses this framework internally and shares it to promote transparency across developers, researchers, and clinicians building health AI tools. [3]

Industry & Products

Medical education must teach clinicians to evaluate AI in clinical context.

In a MedCity News perspective, Ed Lee, MD, argues that medical schools and health systems must treat AI competency as a core clinical skill, not an optional technical one. The defining competency is fitting AI-synthesized information into each patient's clinical context — comorbidities, medication history, preferences, and social circumstances — rather than simply accepting AI-generated summaries. [4]

Research & Evidence

MedGemma shows hierarchical encoding where high-magnitude features track lexical identity and low-magnitude features track contextual meaning.

Journal article (npj Digital Medicine)

Researchers evaluated clinical reasoning stability in GPT-5, MedGemma-27B-Text-IT, and OpenBioLLM-Llama3-70B using 3035 systematic perturbations from physician-validated oncology SOAP, brief vignette, and MIMIC cases. Sparse autoencoder analysis on 1 billion tokens from 50,000 MIMIC-IV notes revealed hierarchical encoding in MedGemma, where high-magnitude features encode lexical identity and low-magnitude features encode contextual meaning. These internal encoding structures differentially affect retrieval and reasoning interventions, suggesting interventions effective for one architecture may harm another, motivating architecture-specific safety validation as a component of pre-deployment evaluation for medical LLMs. [5]

Medical Imaging

Patients prefer AI-simplified breast imaging reports over radiologist-written versions.

Journal article (Insights into imaging)

In a prospective multicenter study of 314 patients evaluating 10 breast imaging reports simplified by ChatGPT-4, AI-generated versions were preferred in 63.3% of 2,965 evaluations and received higher simplicity (70.0% vs 43.4%) and comprehensibility (68.1% vs 40.4%) ratings than radiologist-written reports. Higher comprehensibility of AI reports was the strongest predictor of preference, though the authors support supervised complementary use rather than replacement. [6]