Artificial intelligence prediction of osteoporosis from chest radiographs: Accuracy and association with long-term mortality in patients with chronic obstructive pulmonary disease

E. J. Hwang, J. Y. Lee

Seoul National University Hospital, Department of Radiology, Seoul, South Korea

Objectives

Osteoporosis is highly prevalent in patients with chronic obstructive pulmonary disease (COPD) but often underdiagnosed. It increases the risk of fracture and adverse health outcomes. We aimed to evaluate the accuracy of an artificial intelligence (AI) tool for predicting osteoporosis from chest radiographs (CXRs) and its association with long-term mortality.

Methods & Materials

For this retrospective, single-center study, we included patients with the following eligibility criteria: a) age of 60 years or older; b) diagnosis of COPD; c) available CXRs and pulmonary function test results conducted within 30 days interval between 2011 and 2018; and d) available survival status information. For the analysis of CXRs, we used a commercialized AI tool (PROS CXR: OSTEO, PROMEDIUS INC, Seoul, Korea), which provides a probability score between 0 and 100% for the presence of osteoporosis for a single frontal CXR image as an input. The accuracy of the AI tool was evaluated in patients with available dual-energy X-ray absorptiometry (DEXA) scans conducted within 90 days from the CXRs, which were reference standards for the presence of osteoporosis. The association between AI-predicted osteoporosis and long-term all-cause mortality was evaluated with multivariable Cox proportional hazard analysis, with covariates of age, sex, low body mass index (<18.5 kg/m2), and Global Initiative for Obstructive Lung Disease (GOLD) stage.

Results

4,176 patients (3,548 male and 628 female; median age 72 years) were included. Among 171 patients with available DEXA results, 70 (35.1%) had osteoporosis. The AI tool exhibited an area under the receiver operating characteristic curve of 0.804 (95% confidence interval [CI], 0.734–0.872) for the prediction of osteoporosis. At threshold probability scores of 20% and 30%, the sensitivity of the AI tool was 95.0% and 91.7%, respectively, while corresponding specificities were 37.8% and 47.7%, respectively. The AI-predicted probability score for osteoporosis was an independent risk factor for the all-cause mortality (hazard ratio [HR], 1.002 for a 1% increase of probability; 95% CI, 1.001–1.003; P=.004). Binary prediction results by both thresholds of 20% (HR, 1.119; 95% CI, 1.014–1.235; P=.025) and 30% (HR, 1.130; 95% CI, 1.027–1.243;P=.013) were also independently associated with long-term all-cause mortality.

Conclusion

Among patients with COPD, an AI tool could identify patients with osteoporosis from CXR with reasonable accuracy, and the AI-predicted osteoporosis was an independent risk factor of long-term mortality. The AI tool may help screen for osteoporosis, a correctable risk factor for adverse clinical outcomes in COPD patients.

Authors

First author: Eui Jin Hwang

Presented by: Eui Jin Hwang

Submitted by: Eui Jin Hwang