Healthcare AI Daily Briefing

Clinical Practice

Opinion: Medical records omit women's trust-breaking encounters, risking biased AI.

In an opinion piece, Annemarie G. Hirsch and Jennifer Silva argue that electronic medical records fail to capture women's reported experiences of dismissed symptoms, misdiagnoses, and disrespect, citing their study pairing records with interviews of eighty-seven women where fifty-three described trust-breaking encounters but only seven had this documented. They warn that AI scribes and diagnostic tools trained on such records will reproduce these biases unless clinicians document delayed or incorrect diagnoses and create safer spaces for women to be heard. [1]

Operations & Workflow

Assort Health began as first agentic voice AI for specialty scheduling, starting with orthopedics in Epic.

Podcast from Claims Denied: A Hospitalogy Podcast

The interview describes Assort Health's thesis: the company started three years ago as the first agentic voice system to schedule appointments in specialty care, beginning with orthopedics in an Epic environment. The account does not specify which workflows the platform now enables beyond that origin. [2]

Policy & Governance

Medical AI Education Needs Mandatory Multi-Tiered Reform, Not Two-Week Courses.

In an opinion piece, Prof. Dr. Gheorghe Ioan Mihalas argues that current medical AI education has been hijacked by a market of quick fixes promising AI expertise in two weeks while stripping away theoretical foundations. He calls for immediate enforceable regulatory action mandating tailored, multi-tiered education across the medical continuum, warning that superficial prompt-building tricks create a dangerous illusion of knowledge that threatens patient safety and dissolves clinical authority. [3]

Policy & Governance

Midterm voters reject AI data centers, raising questions for health AI policy.

Tuesday's midterm elections saw voters roundly reject data centers needed to power AI on dozens of ballot measures across several states, reflecting polling that many Americans believe AI does more harm than good. The results leave open how this negative sentiment trickles into health AI legislation, including state sandbox bills and federal efforts to clarify FDA authority and reimbursement, with one legal expert calling AI a 'boogeyman' and data centers a symbol of perceived industry coziness. [4]

Diagnostics & Pathology

AI assistance improves pathologist agreement and sensitivity in HER2-low breast cancer interpretation.

Journal article (Cancer research and treatment)

In a nationwide multireader study of 168 breast cancer cases, thirty-four board-certified pathologists showed significantly improved interobserver agreement with AI assistance (Krippendorff's alpha 0.7415 to 0.8849; Fleiss' kappa 0.4492 to 0.7524). Sensitivity gains were largest for HER2-ultralow (48.5% to 73.9%) and HER2 IHC 1+ (64.9% to 87.7%) categories, with AI also reducing interpretation time for HER2-ultralow cases by 3.12 seconds. [5]

Research & Evidence

Multi-agent EHR system earns high clinician ratings for cognitive decline summaries and decision support.

Preprint, not peer reviewed

In a preprint, MERIT — a multi-agent CDSS co-designed with primary care physicians — used retrieval-augmented generation on multimodal EHR data from one PCP's panel for development and four others for testing. Clinician evaluators rated virtual cognitive status summaries 4.0–4.3 and decision support 3.9–4.3 on a 5-point Likert scale for correctness, completeness, and adoptability; automated readability showed provider-facing summaries at college level while patient education reached high-school level, though the evaluation covered only the provider-facing EHR component without patient conversation or prospective outcomes. [6]