Healthcare AI Daily Briefing

Research & Evidence

EMR model predicts frailty at 72 hours with F2 0.755 in retrospective Adelaide cohort.

Journal article (JMIR medical informatics)

A retrospective cohort of 26,824 overnight admissions at Royal Adelaide Hospital (July 2023–December 2024) used EMR data available within the first 72 hours to predict frailty defined by Hospital Frailty Risk Score >4. The final XGBoost model achieved an F2‑score of 0.755 (95% CI 0.741–0.768), AUROC 0.777 (95% CI 0.764–0.790), and recall 0.873 (95% CI 0.857–0.887) on a held‑out test set of 5,388 admissions, with good calibration (expected‑to‑observed ratio 0.996). Performance was lower at 24 hours and improved by discharge; the authors note reduced performance in Indigenous patients, highlighting the need for culturally validated frailty indices. [1]

Clinical Practice

AI diagnostic recommendations cost 3.2× physician care in NEJM AI study.

In a NEJM AI study, Dvir Aran and colleagues compared 24 AI systems with treating physicians across 200 outpatient and ED visits and found AI-suggested care averaged $314 per visit versus $97 for physicians. The excess split roughly evenly between more testing and more specialty referrals. When prompted to be cost-aware, AI-recommended care fell to 2.5 times physician cost; when prompted to be parsimonious, it fell to 1.4 times. In about 20% of cases where AI recommended more workup than physicians, the additional testing was guideline-concordant; in the majority it was not. [5]

Clinical Practice

Voice-guided AI navigation improves biliopancreatic EUS completeness but lengthens procedure time in randomized trial.

Journal article (Clinical gastroenterology and hepatology : the official clinical practice journal of the American Gastroenterological Association)

In a single-center randomized trial of 266 patients, the AI system EUS-GUIDEQ raised median station completeness to 1.00 [IQR 0.94-1.00] from 0.81 [0.75-0.88] versus standard endosonography (P<0.001). Procedure duration increased to 12.70±3.23 minutes from 10.55±3.62 minutes (P<0.001). The station-classification model reached 94.2% internal and 78.8% external accuracy. No serious adverse events occurred; authors call for multicenter, multi-vendor prospective validation. [4]

Clinical Practice

Live-streamed thrombectomy training shows dose-response gains strongest in already-skilled clinicians.

Journal article (Journal of neurosurgery)

In a prospective multicenter cohort of 1,010 clinicians, standardized live-streamed mechanical thrombectomy training was linked to better 90-day functional outcomes and procedural performance. Double/debiased machine learning models found consistent treatment effects (0.83–0.85, p<0.001). A Bayesian generalized linear mixed model mapped personalized dose-response trajectories stratified by baseline competency and surgical experience, confirming strong associations between training duration (OR 2.2, 95% CI 1.58–3.02) and competency (OR 3.7, 95% CI 2.79–4.86). Improvements were linear across all levels but most pronounced among high-competency practitioners. [3]

Research & Evidence

LLM-assisted adaptive weight optimization for head-and-neck seed implant brachytherapy shows feasible CTV coverage in retrospective study.

Journal article (Annals of the New York Academy of Sciences)

A locally deployed DeepSeek-R1 LLM integrated with an automatic planning algorithm iteratively evaluated plan quality and recommended new objective function weights until convergence, using a clinical knowledge base via retrieval-augmented generation. Retrospective validation on 23 head-and-neck seed implant brachytherapy cases found the LLM-assisted approach maintained clinical target volume coverage and reduced excessive target dose versus fixed-weight planning; in the eight patients with available organs-at-risk contours, OAR dose metrics showed favorable numerical trends. The method is proposed as an assistive tool under clinician supervision. [2]

Industry & Products

Nvidia releases open-weight Nemotron 3.5 Transcribe for 40 languages.

Nvidia on Wednesday released Nemotron 3.5 Transcribe, a free, open-weight model that transcribes audio in up to 40 languages and can translate 31 languages to and from English. The model handles recordings up to an hour long and distinguishes up to eight speakers, but works only with recorded audio, not livestreams. Nvidia also released version 2 of its open-source Granary speech dataset and says developers can customize the model for industry jargon using NeMo Speech tools. [6]