Daily Briefing

What Changed

Declining clinician use makes sustained adoption, not availability, the key test for EHR-embedded generative AI [2].

Policy & Governance | AI oversight still lacks clear alert ownership and data readiness. The Signal Room podcast reported that healthcare has spent five years getting AI in the door, but oversight is not ready, and it is the meeting after the model that matters, according to Chris Hutchins, who describes that the hard part is determining who owns the alert and whether anyone will admit the data isn't ready [1].

Research & Evidence | Clinician use of an EHR-integrated AI chart-summary tool declined during a randomized preprint trial. A pragmatic randomized preprint trial included 284 ambulatory clinicians across forty-two specialties. Of 74,474 AI chart summaries generated, 14.2% were interacted with by a clinician; the proportion of generated summaries interacted with declined from 21.5% in month 1 to 10.5% in month 3, and the proportion of clinicians using the tool at least once per month declined from 88.7% to 66.2% [2].

Operations & Workflow | Ambient AI scribes remain ineffective for ventilated ICU documentation. The Pop Health Podcast reported that Beth Israel Lahey Health identified AI scribes as ineffective for ICU documentation due to lack of patient dialogue when patients are on ventilators, creating a gap in ambient scribing utility for critically ill patients [3].

Research & Evidence | ICU deep learning predicted mortality and renal-replacement therapy after cardiac surgery. A study analyzed 7,402 cardiac-surgery ICU admissions from 2008-2019 in one database, using a 70:20:10 training-validation-test split. DLMs achieved an AUPRC of 0.877 for in-hospital mortality and 0.906 for CRRT prediction, with F1-scores of 0.808 and 0.757 respectively [4].