Osteoporos Sarcopenia
Publication
OSTEO
Osteoporosis screening using chest radiographs with a deep learning model in a multi-ethnic population in Malaysia A single-center experience
Jack-Kang Tan, Muhammad Ridhwan Nasir, Wei-Lin Ng, Gaeun Lee, Junhyeok Park, Jemyoung Lee, Lee-Ling Lim
Abstract
Objectives
Osteoporosis is underdiagnosed, partly due to limited access to dual-energy X-ray absorptiometry. We evaluated the diagnostic performance of a deep learning-based opportunistic screening tool for identifying osteoporosis using chest radiographs in a multi-ethnic population.
Methods
In this prospective cross-sectional diagnostic accuracy study, 201 osteoporosis treatment-naïve participants aged ≥50 years (Malay, Chinese, and Indian ethnicities) were recruited from an urban academic institution in Kuala Lumpur, Malaysia. Posteroanterior chest radiographs were analyzed by the convolutional neural network-based model to generate osteoporosis probability scores, which were validated against DXA measurements of the lumbar spine and proximal femur. Area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity were calculated, stratified by ethnicity.
Results
The model achieved an overall AUROC of 0.90 (95% confidence interval [CI] 0.85–0.94). At the pre-specified probability threshold of 0.20, sensitivity was 0.69 (95% CI 0.58–0.80), specificity was 0.91 (95% CI 0.86–0.96), and negative predictive value was 0.85 (95% CI 0.79–0.91). In an exploratory analysis, a post hoc Youden-derived cutoff of >0.025 yielded sensitivity of 0.84 (95% CI 0.75–0.93) and specificity of 0.84 (95% CI 0.78–0.90). No significant differences in AUROC were observed across the 3 major ethnic groups, with AUROCs of 0.91 (95% CI 0.85–0.97) for Malays, 0.88 (95% CI 0.79–0.98) for Chinese, and 0.90 (95% CI 0.78–1.00) for Indians, although statistical power was limited in the smallest subgroup.
Conclusions
Integrating AI-assisted analysis into the routine chest radiograph workflow may enable opportunistic osteoporosis screening without additional radiation exposure.
Keywords
Osteoporosis; Opportunistic screening; Bone mineral density; Chest radiography; Convolutional neural network
