World Health AIRegister

All insights

Healthcare

AI reading routine chest X-rays beats FRAX at predicting fractures in 85,284 patients

In patients with osteopenia, the deep-learning model's 10-year AUROC was 0.779 against 0.659 for FRAX with bone density, NEJM AI reports.

By World Health AI Editorial2 min read1 source

A deep-learning model that reads routine chest X-rays ranked patients' risk of a major osteoporotic fracture more accurately than FRAX with bone density, according to a Korean-funded study published in NEJM AI on 16 November 20261. The researchers built the model on 53,109 adults at one tertiary hospital and then tested it on 85,284 patients from two other institutions1.

What the study found

The model is a convolutional neural network survival model, and its outcome was a new major osteoporotic fracture1. To be included, a patient needed a chest radiograph taken within a year of a dual-energy X-ray absorptiometry scan, so the comparison with FRAX with bone mineral density and with the lowest T-score could be made on the same people1. The external cohorts had a mean age of 58.6, were 68.9% female and were followed for 77 months1.

In those cohorts the model reached a C-index of 0.843 and a censoring-adjusted 10-year AUROC of 0.8551. FRAX with bone density scored a C-index of 0.779 and a 10-year AUROC of 0.785, and both gaps were significant at P<0.011. Putting FRAX into the model added little, moving the C-index only to 0.8471.

The widest gap came in the 35,358 patients with osteopenia, the group in which the authors say bone density and FRAX miss much of the risk even though much of the fracture burden falls there1. Among them the model's 10-year AUROC was 0.779 against 0.659 for FRAX with bone density1. The authors report that the advantage held across fracture types and bone density categories1.

What it means for the market

The commercial case rests on the word "opportunistic" in the title: chest radiographs are already taken across a wide range of patients, so a model of this kind would draw a fracture risk estimate from images health systems already hold, with no extra scan1. For radiology departments, osteoporosis services and imaging AI vendors, the study suggests fracture risk could become a by-product of everyday imaging rather than a separate test.

The result also bears on how far such a model could replace existing tools. Because adding FRAX barely changed performance, the image alone carried most of the signal the study measured1. The paper does not say whether the model is cleared by any regulator, who would sell it or what it would cost, so routes to purchase remain unknown.

Caveats

Every patient in the study had already been referred for a bone density scan, which is not the population a hospital would screen opportunistically1. The authors say absolute risk and case-finding estimates therefore need prospective evaluation in an unselected population having routine radiography before the model's practical yield is known1. The work was funded by the National Research Foundation of Korea among others1.

Sources

  1. Opportunistic Risk Assessment of Major Osteoporotic Fractures from Routine Chest Radiographs Using Deep Learning. (opens in a new tab)NEJM AI · doi.org

This briefing summarises publicly available research and reporting for information only. It is not medical, investment or legal advice. Follow the sources listed for the primary research and reporting.