Opportunistic Risk Assessment of Major Osteoporotic Fractures from Routine Chest Radiographs Using Deep Learning

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 validated it in 85,284 patients from two independent institutions. Eligibility required a chest radiograph obtained within 1 year of dual-energy X-ray absorptiometry scans, so all participants had been referred for densitometry. The primary outcome was incident major osteoporotic fracture. Discrimination was assessed by using the concordance index (C-index) and area under the receiver operating characteristic curve (AUROC) at 2, 3, 5, and 10 years, compared with FRAX with BMD and with the lowest BMD T-score. RESULTS: In the external validation cohorts (mean 58.6 years of age; 68.9% female; and follow-up of 77 months), the deep-learning (DL) model achieved a C-index of 0.843 (95% confidence interval [CI], 0.837 to 0.850) and a censoring-adjusted 10-year AUROC of 0.855 (95% CI, 0.846 to 0.863), which was higher than that of FRAX with BMD (C-index, 0.779, 95% CI, 0.770 to 0.787; 10-year AUROC, 0.785, 95% CI, 0.775 to 0.794; both P<0.01). Adding FRAX to the DL model did not meaningfully improve discrimination (C-index, 0.847, 95% CI, 0.841 to 0.853). In the osteopenia subgroup (n=35,358), the DL model achieved a 10-year AUROC of 0.779 (95% CI, 0.764 to 0.795) versus 0.659 (95% CI, 0.641 to 0.676) for FRAX with BMD (P<0.01). This advantage was consistent across fracture types and BMD categories. CONCLUSIONS: A survival DL model applied to routine chest radiographs identified patients at high fracture risk with discrimination exceeding that of FRAX, with the greatest benefit in patients with osteopenia. As all participants had been referred for densitometry, absolute-risk and case-finding estimates require prospective evaluation in an unselected routine-radiography population. (Funded by the National Research Foundation of Korea and others.).

Authors

Institutions

Publication Details

Journal
NEJM AI
Published
2026-11-16
DOI
https://doi.org/10.1056/aioa2600590
Primary Topic
Bone health and osteoporosis research
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Opportunistic Risk Assessment of Major Osteoporotic Fractures from Routine Chest Radiographs Using Deep Learning

Jin Hwan Kim, Yisak Kim, Sung Hye Kong, Chang Min Park et al.
4 links
NEJM AI
Bone health and osteoporosis research
article

Opportunistic Risk Assessment of Major Osteoporotic Fractures from Routine Chest Radiographs Using Deep Learning

Jin Hwan Kim, Yisak Kim, Sung Hye Kong, Chang Min Park, Dongheon Lee, Hanbin Ko
article en

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 validated it in 85,284 patients from two independent institutions. Eligibility required a chest radiograph obtained within 1 year of dual-energy X-ray absorptiometry scans, so all participants had been referred for densitometry. The primary outcome was incident major osteoporotic fracture. Discrimination was assessed by using the concordance index (C-index) and area under the receiver operating characteristic curve (AUROC) at 2, 3, 5, and 10 years, compared with FRAX with BMD and with the lowest BMD T-score. RESULTS: In the external validation cohorts (mean 58.6 years of age; 68.9% female; and follow-up of 77 months), the deep-learning (DL) model achieved a C-index of 0.843 (95% confidence interval [CI], 0.837 to 0.850) and a censoring-adjusted 10-year AUROC of 0.855 (95% CI, 0.846 to 0.863), which was higher than that of FRAX with BMD (C-index, 0.779, 95% CI, 0.770 to 0.787; 10-year AUROC, 0.785, 95% CI, 0.775 to 0.794; both P<0.01). Adding FRAX to the DL model did not meaningfully improve discrimination (C-index, 0.847, 95% CI, 0.841 to 0.853). In the osteopenia subgroup (n=35,358), the DL model achieved a 10-year AUROC of 0.779 (95% CI, 0.764 to 0.795) versus 0.659 (95% CI, 0.641 to 0.676) for FRAX with BMD (P<0.01). This advantage was consistent across fracture types and BMD categories. CONCLUSIONS: A survival DL model applied to routine chest radiographs identified patients at high fracture risk with discrimination exceeding that of FRAX, with the greatest benefit in patients with osteopenia. As all participants had been referred for densitometry, absolute-risk and case-finding estimates require prospective evaluation in an unselected routine-radiography population. (Funded by the National Research Foundation of Korea and others.).

NEJM AI
Seoul National University (KR), Chungnam National University (KR), Seoul National University Hospital (KR), Seoul National University Bundang Hospital (KR)
Good health and well-being
Bone health and osteoporosis research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon