Healthcare AI Daily Briefing

Research & Evidence

Machine‑learning early warning score markedly improves mortality prediction in high‑risk patients.

Preprint

A retrospective medRxiv preprint using 2.08 million hospital encounters compared XGB-EWS, a gradient-boosted model predicting 24-hour mortality, with NEWS. Researchers adjusted for intervention bias using propensity scores. Decision-curve analysis showed a net benefit gain of up to 1.9 correct identifications per 10,000 patients overall, but up to 78.1 in the high-risk group. [2]

Clinical Practice

Nolla Health’s Utah pilot sets phased physician review for AI‑acne prescriptions.

Nolla Health announced a Utah pilot that lets its AI scan users’ faces, assess acne, and write prescriptions. Two physicians will approve each AI‑generated prescription before it is issued for the first 100 patients; after up to 500 patients, physicians will review at least 10% of prescriptions each month plus any case involving an escalation or side effect. [4]

Policy & Governance

Interview: Houston Methodist built AI oversight with guiding principles and risk‑based rollout.

Podcast from Healthcare AI Pioneers

The interview describes that three years ago Houston Methodist decided to design its AI oversight, creating guiding principles and holding crucial conversations about risk. The health system agreed not to accept certain AI risks during the first one‑to‑two years, then established objective methods to reevaluate those principles and the risk‑benefit profiles it deems acceptable. [5]

Policy & Governance

Feynman Xu advises defining the AI employee’s queue before measuring performance.

In an opinion piece, Feynman Xu, Ph.D., Founder and CEO of MedArise, argues that practices must first identify the administrative queue—its trigger, required inputs, completion state, evidence, and escalation path—before measuring completed in‑scope outcomes, turnaround time, backlog age, reopened cases, administrative exceptions resolved, and practice interventions still required to compare AI, outsourced, or human options on a common scale. [3]

Clinical Practice

BALL reduces trajectory error by 7% and improves psychiatric score prediction.

Preprint

BALL estimates state trajectories by combining sparse reference measurements with denser records. Across 30 simulations it cut composite trajectory error by 7% versus direct transformer training, and in psychiatric records from 1,609 patients improved withheld questionnaire score prediction at 21- and 28-day schedules, with uncertainty-guided queries adding further gain; it also links retrospective learning to prospective estimation. [1]

Industry & Products

Will an AI agent ever win a Nobel Prize for medicine?

In an opinion piece by Jeffrey Flier, the former Harvard Medical School dean wonders whether Claude or a future AI agent could win a Nobel Prize for medicine. The Nobel statutes refer to persons; Flier remains uncertain how credit for AI discoveries should be assigned. [6]