Moderate evidence
More
Study finds AI reading chest X-rays predicts bone fractures better than FRAX
Study title: Opportunistic Risk Assessment of Major Osteoporotic Fractures from Routine Chest Radiographs Using Deep Learning
- Evidence type
- Controlled or validated comparison
- Sample
- n = 85,284
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2981cite or check it
Study title: Opportunistic Risk Assessment of Major Osteoporotic Fractures from Routine Chest Radiographs Using Deep Learning
A convolutional neural network survival model that reads routine chest radiographs identified people at high risk of major osteoporotic fracture better than FRAX combined with bone mineral density. In external validation it reached a C-index of 0.843 and a 10-year AUROC of 0.855, against 0.779 and 0.785 for FRAX with BMD.
Read the study abstract
Background Bone mineral density (BMD) and the Fracture Risk Assessment Tool (FRAX) incompletely capture fracture risk in the osteopenic range, where much of the fracture burden occurs. Routine chest radiographs are acquired across a broad range of patients and may contain information relevant to fracture risk, offering a potential opportunistic complement without additional imaging burden. Methods We developed a convolutional neural network survival model in 53,109 adults at a tertiary hospital and externally va…

The gain
In an external validation of 85,284 patients, a deep-learning model reading routine chest radiographs predicted major osteoporotic fracture with a C-index of 0.843, versus 0.779 for FRAX with bone mineral density.
In brief
Bone density and FRAX incompletely capture fracture risk in the osteopenic range, where much of the fracture burden occurs. Chest radiographs are already taken in a broad range of patients, so the approach could add risk information without extra imaging. The authors report the greatest benefit in patients with osteopenia.
Everyone in the study had been referred for densitometry, so the results may not carry over to an unselected population. The authors say absolute-risk and case-finding estimates need prospective evaluation in routine radiography patients.
Main points
- The model was developed in 53,109 adults at a tertiary hospital.
- It was externally validated in 85,284 patients from two independent institutions.
- External cohorts had a mean age of 58.6 years, were 68.9% female and had 77 months of follow-up.
- The model's C-index was 0.843 (95% CI 0.837 to 0.850), versus 0.779 for FRAX with BMD.
- Its censoring-adjusted 10-year AUROC was 0.855, versus 0.785 for FRAX with BMD.
- The advantage was greatest in patients with osteopenia.
The full story
Researchers built a convolutional neural network survival model in 53,109 adults at one tertiary hospital. Eligibility required a chest radiograph within 1 year of a dual-energy X-ray absorptiometry scan. The primary outcome was incident major osteoporotic fracture.
The model was externally validated in 85,284 patients from two independent institutions. Discrimination was assessed with the C-index and AUROC at 2, 3, 5 and 10 years, compared with FRAX with BMD and with the lowest BMD T-score.
In the validation cohorts, the model scored a C-index of 0.843 and a 10-year AUROC of 0.855, higher than FRAX with BMD. The authors conclude the benefit was greatest in patients with osteopenia, a group where standard tools miss much of the fracture risk.
Sources
1 on file- NEJM AIdoi:10.1056/aioa2600590
Talk it over
The model reached a C-index of 0.843 in 85,284 patients from two independent institutions, versus 0.779 for FRAX with bone mineral density. What does a validation at other institutions tell you, and how much weight would you give it?
Study finds AI reading chest X-rays predicts bone fractures better than FRAXhttps://truvace.com/item/opportunistic-risk-assessment-of-major-osteoporotic-fractures-from-routine-chest
The model reached a C-index of 0.843 in 85,284 patients from two independent institutions, versus 0.779 for FRAX with bone mineral density. What does a validation at other institutions tell you, and how much weight would you give it?
Answer in the threadEveryone in the study had been referred for densitometry. How could the results differ for people having routine chest radiographs who were never referred for a bone scan?
Answer in the threadThe authors say absolute-risk and case-finding estimates need prospective evaluation. Who should run that evaluation, and what result would persuade you to use chest radiographs this way?
Answer in the thread
If you have more time
The advantage was greatest in patients with osteopenia. If you were a patient in that group, how would you want a doctor to explain a risk score drawn from your chest X-ray?
Answer in the threadThe model was built in 53,109 adults at one tertiary hospital. What would you ask about differences between that hospital and the two validation institutions before trusting the result elsewhere?
Answer in the thread
ace
What readers say