Healthcare AI Daily Briefing

Industry & Products

Doximity internal test finds 2026 LLMs often miss dose adjustments; its DoseCheck tool claims 99.9% accuracy in company validation.

Doximity queried several 2026 large‑language models with more than 8,500 dosing questions covering over 2,200 FDA‑approved drugs. Errors were 13.2% minor, 43% significant and 3.4% potentially lethal, most often a failure to adjust for age, kidney function or drug interactions; newly approved drugs had five‑times the odds of error. The company’s own, non‑peer‑reviewed internal validation reported its DoseCheck verification tool was correct 99.9% of the time. [1]

Policy & Governance

AI models order three times more care than physicians in real‑visit study.

A NEJM AI study had 24 models write care plans for 200 real outpatient and ED visits; AI plans averaged $314 per visit versus $97 for physicians. On the 66% of visits where doctors ordered no tests, every model still ordered a standard panel of labs. Cost‑aware prompts cut the cost gap to 2.5× and a parsimonious‑but‑safe prompt to 1.4×, yet only about 20% of the extra workups were guideline‑concordant. Wachter says “what is enough” must be deliberately encoded by health systems, payers, regulators and vendors, with clinicians staying in the loop to judge when nothing is needed. [2]

Industry & Products

Prodigy refuses low-quality results, boosts abstraction efficiency.

MRO's Prodigy tool for clinical data abstraction will not present search results unless believed high quality, achieving 97% accuracy. Designed for high-consequence uses: clinical trial recruitment, infection control, accreditation. It automates harmonization of unstructured EHR text into CDISC SDTM format. Crucially, Prodigy handles 40% of an abstractor's traditional time spent finding/copying EHR data, working alongside human abstractors rather than replacing them, improving efficiency while maintaining quality in complex, context-dependent data extraction. [3]

Clinical Practice

Sully.ai co‑founders argue pausing health AI would lock in medication errors and physician shortages.

In an opinion piece, Sully.ai co‑founders Ahmed Omar and Amit Kumthekar write that halting health‑AI adoption would cement medication errors, rural physician gaps and documentation burnout. They cite the AAMC forecast of an 86,000‑physician shortfall by 2036, federal projections that non‑metro areas will have only 42 % of needed physicians by 2038, and a drop in public openness to AI from 52 % in 2024 to 42 % now, while 51 % have already used AI for important health decisions without a clinician. The authors call for careful testing, honest disclosure, explicit consent, keeping physicians in charge and transparent publication of results. [4]

Medical Imaging

Negative LVEF reserve linked to higher mortality and high‑risk coronary disease in 34,976‑patient PET registry.

Journal article (Circulation. Cardiovascular imaging)

In the REFINE PET registry of 34,976 patients from 13 sites undergoing vasodilator stress PET myocardial perfusion imaging, a lower left ventricular ejection fraction reserve (stress EF minus rest EF) showed a graded association with all-cause mortality over a median 4.1-year follow-up. Compared with LVEF reserve above 5%, adjusted hazard ratios for all-cause mortality were 1.20 for 0–5% (95% CI 1.13–1.28), 1.32 for −5% to 0% (95% CI 1.23–1.42), and 1.38 for ≤−5% (95% CI 1.24–1.52). In the 6,843-patient angiography subset, a negative LVEF reserve also identified higher likelihood of left-main or three-vessel obstructive coronary artery disease. [5]

Clinical Practice

AI‑ECG screening raises low ejection fraction detection in routine care, meta‑analysis of three RCTs finds.

Journal article (Annals of noninvasive electrocardiology : the official journal of the International Society for Holter and Noninvasive Electrocardiology, Inc)

A systematic review and meta‑analysis of three randomized trials involving 37,467 participants showed that AI‑enabled ECG screening increased new low ejection fraction (≤50%) diagnoses compared with usual care (risk ratio 1.36, 95% CI 1.12–1.64; number needed to screen 204). Echocardiography referral rates did not differ significantly (risk ratio 1.02, 95% CI 0.96–1.09). The AI tools’ positive predictive values ranged from about 20% to 34% across studies, and the authors call for further validation across diverse AI platforms. [6]

Clinical Practice

AI virtual patient platform speeds early colposcopy diagnostic and communication skills.

Journal article (Journal of gynecology obstetrics and human reproduction)

In a randomized, assessor‑blinded, wait‑list‑controlled trial of 60 PGY‑2/3 residents, adding an AI‑driven virtual patient platform to traditional teaching produced significantly higher scores at 1 month in Case Analysis (mean difference 7.54, p<0.001, d=1.54), Diagnostic Justification (d=1.15), Communication Performance (d=1.02) and Theoretical Knowledge (d=0.74). Operative competence did not differ (p=0.113). The control group crossed over to the AI platform after month 1, and both groups reached high proficiency (>85 points) by 3 months. Satisfaction was high (90% of participants, 91.7% of supervisors). [7]

Research & Evidence

Spatial deep‑learning model stratifies cancer risk in Barrett's esophagus using DNA‑content and immune features.

Journal article (Cancer research)

Researchers built BEACON, a spatially aware framework that first uses DACOR — a multi‑instance learning model trained on 777 Barrett's esophagus biopsies with flow‑cytometry ground truth from two centers — to predict DNA‑content abnormality from routine H&E slides. Among patients classified as DNA‑content abnormal, cancer progressors showed greater plasma‑cell clustering next to abnormal epithelium than non‑progressors. A logistic‑regression model that combined molecular, immune‑ecology, and morphology features then stratified only those DNA‑content‑abnormal patients into high‑ and low‑risk groups that aligned with progression and survival; performance metrics and external cohort size were not disclosed. [8]