Global validation of osteoporosis screening on chest radiographs

Purpose: Osteoporosis, a global health concern, affects 19.7% of the population worldwide, with a higher prevalence in developing countries. Previous studies suggest that osteoporosis screening can effectively reduce societal costs by preventing osteoporotic fractures. The incidence of fragility fractures and their associated healthcare expenditures are expected to increase in the coming years due to demographic shifts towards an aging population. To address this issue, we propose a cost-effective screening method utilizing chest radiographs, which offer superior accessibility and cost-efficiency compared to other diagnostic tools.

Materials and Methods: We evaluated our osteoporosis screening model using four external datasets, which comprised chest radiograph (CXR) data paired with dual-energy X-ray absorptiometry (DXA) results from individuals aged over 50 years. We utilized data from four distinct external datasets. Dataset A is sourced from a tertiary university hospital. Dataset B originates from a secondary care hospital specializing in spine and joint care, which has a relatively lower proportion of osteoporosis patients. Dataset C is from a hospital for veterans, representing diverse settings within the healthcare system and featuring a higher proportion of males and elderly individuals. Datasets A, B, and C are all based in South Korea. Conversely, dataset D was obtained from a global medical imaging platform, with most data sourced from the United States and Brazil. A commercial AI tool (PROS® CXR: OSTEO, PROMEDIUS INC, Seoul, Korea) was retrospectively applied to CXRs to predict osteoporosis. CXR manufactures of each hospital were very different .The accuracy of this AI tool was assessed by comparing its predictions to osteoporosis diagnoses made using DEXA scans, which served as the reference standard. We evaluated the area under the receiver operating characteristic curves (AUCs), sensitivity, specificity, and F1 score. 

Results: Dataset A comprised 1,089 individuals, with 969 females (mean age 58 years) and 120 males (mean age 59 years), and an osteoporosis prevalence of 29.2%. Dataset B included 3,338 individuals, with 1,854 females (mean age 59 years) and 1,484 males (mean age 59 years), showing an osteoporosis prevalence of 6.0%. Dataset C consisted of 937 individuals, with 411 females (mean age 72 years) and 526 males (mean age 74 years), with an osteoporosis prevalence of 24.9%. Dataset D included 295 individuals, with 288 females (mean age 66 years) and 7 males (mean age 69 years), and an osteoporosis prevalence of 18.3%. In the A dataset, the performance was AUC 0.89, sensitivity 0.84, specificity 0.76, and F1 score 0.70. In the datasets B, C, and D, AUC was 0.92, 0.89, and 0.81, sensitivity was 0.71, 0.85, and 0.39, specificity was 0.92, 0.77, and 0.94, and F1 score was 0.48, 0.67, and 0.47, respectively. 

Conclusion: Osteoporosis screening using CXR demonstrated excellent performances of general usability across a variety of quality and multiple machines of chest radiographs from multiple hospitals with various conditions.

Clinical relevance: Osteoporosis diagnosis is critical for initiating timely interventions to prevent fractures. This study proposes leveraging CXRs, which are widely accessible and routinely performed, as an innovative screening tool for osteoporosis, particularly in resource-constrained settings. This approach has the potential to expand diagnostic capabilities and promote earlier treatment initiation across a broader patient demographic, thereby improving detection of osteoporosis and reducing the burden of osteoporotic fractures especially in developing countries.