Physician in the Loop
For youSign in
Endocrinology ·

AI reading chest X-rays flags osteoporosis in external cohorts; LLM system refines hyperthyroidism admission notes

An insulin pen lying at a slight angle with its cap beside it, showing a clear cartridge window, a small capped hub at the tip and an amber dose knob.
THE SHORT READ

A deep learning model that screens chest radiographs for osteoporosis reached an area under the curve of 0.88 to 0.93 in three external validation cohorts, Korean researchers reported in npj Digital Medicine. A Chinese group reported a large language model system for hyperthyroidism admission notes with a 6.67% hallucination rate, versus 26.67% for a general large language model.

THE DETAILS
  • AI Oct 9, 2026
    AI model reads chest X-rays for osteoporosis with area under the curve of 0.88 to 0.93 in three external cohorts
    Researchers at Seoul National University Bundang Hospital and Soongsil University reported AI-OsRM in npj Digital Medicine on Oct. 9, a model developed with 46,478 paired chest radiograph and dual-energy X-ray absorptiometry (DXA) examinations from 26,757 patients. In external validation, the area under the receiver operating characteristic curve was 0.88 in a temporally separated hospital cohort of 6,391 people, and 0.93 and 0.92 for osteoporosis detection in two health checkup cohorts of 24,132 and 122,535. In simulations, prescreening with the model could reduce DXA use by about 50%, the authors said. The study measured accuracy, not patient outcomes.
    Source: npj Digital Medicine
  • AI Oct 7, 2026
    LLM system has 6.67% hallucination rate versus 26.67% for a general LLM in 30 hyperthyroidism admission notes
    Researchers at Fuzhou University Affiliated Provincial Hospital in China reported in npj Digital Medicine on Oct. 7 a multi-agent large language model system with retrieval-augmented generation, HQ-ANR, that refines the history of present illness in admission notes. In a blinded evaluation of 30 resident-drafted hyperthyroidism notes, it reduced positive factual hallucinations to 6.67%, compared with 26.67% for a general large language model with chain-of-thought prompting, and raised median Physician Documentation Quality Instrument scores from 39.00 to 42.00. Physician refinement time fell by an average of 112.9 seconds per note, the authors reported.
    Source: npj Digital Medicine

The weekly innovation roundup for endocrinology, published every Sunday for the week before; every item links to its sources. Latest endocrinology news · All specialties