Accuracy of artificial intelligence-based prediction of osteoporosis from chest radiographs and association with risk of long-term mortality

Purpose: Osteoporotic fracture has become a big concern in elderly individuals. Identifying asymptomatic individuals with osteoporosis may help reduce this burden. We aimed to evaluate the accuracy of an artificial intelligence (AI) model to predict osteoporosis from chest radiographs (CXRs) and its association with long-term mortality.

Material and Methods: We retrospectively included consecutive adult (age ≥19 years) individuals who underwent CXRs and dual-energy x-ray absorptiometry (DEXA) scans on the same day for health checkups between 2014 and 2017 in a single institution. A commercialized AI tool (PROS® CXR: OSTEO, PROMEDIUS INC, Seoul, Korea) for predicting osteoporosis in the form of probability scores (0–100%) was retrospectively applied to CXRs. The accuracy of AI was evaluated based on osteoporosis diagnosis using DEXA scans as reference standards. We evaluated the area under receiver operating characteristic curves (AUCs) for the score provided by AI. We also evaluated the sensitivity and specificity of AI at two different thresholds: Predicted scores of 10% and 50%. We also investigated the association between the prediction of osteoporosis by AI and long-term all-cause mortality after adjustment of age, sex, and body mass index using Cox proportional hazard analysis.

Results: A total of 10,412 CXRs from 8,618 asymptomatic individuals (mean age, 58 years; 2882 men) were included. The prevalence of osteoporosis by DEXA scan was 4.2% (440/10,412). The AI’s prediction exhibited an AUC of 0.91 (95% confidence interval [CI], 0.90–0.92) for osteoporosis by DEXA scan. At a threshold predicted score of 10%, the sensitivity and specificity were 72.5% (319/440; 95% CI, 68.1–76.6%) and 89.4% (8,911/9,972; 95% CI, 88.7–90.0%), respectively. Meanwhile, at a score of 50%, the sensitivity and specificity were 61.1% (269/440; 95% CI, 56.4–65.7%) and 93.7% (9344/10,412; 93.2–94.2%), respectively. In multivariate Cox proportional hazard analyses of 8,167 individuals with available mortality information, a higher predicted score for osteoporosis by the AI was associated with a higher risk of long-term all-cause mortality (hazard ratio [HR], 1.01 per 1% point increase of score; 95% CI, 1.00–1.01; P=.016). Binary prediction results by both thresholds of 10% (HR, 1.40; 95% CI, 1.08–2.06; P=.014) and 50% (HR, 1.49; 95% CI, 1.05–2.12; P=.027) were also independently associated with long-term all-cause mortality

Conclusions: Osteoporosis prediction from CXR using AI was accurate compared to DEXA scan, and was associated with risk of long-term mortality.

Clinical relevance statement: An AI analysis of CXR can help opportunistic screening of asymptomatic osteoporosis, which can potentially provide a chance to reduce the risk of long-term mortality and morbidity.