Federated learning for privacy-preserving opportunistic screening of lumbar-spine DXA-defined osteoporotic-level bone mineral density using chest radiographs: A multi-center independent validation study.

Objectives: To evaluate the performance of an artificial intelligence (AI) system trained via federated learning for opportunistic screening of lumbar-spine DXA-defined osteoporotic-level bone mineral density (T-score ≤ -2.5) using routine chest X-rays (CXRs) in a multi-center independent validation setting. Methods: A deep learning model was developed via privacy-preserving federated learning on 5943 paired CXR and dual-energy X-ray absorptiometry (DXA) examinations and validated on an independent dual-center cohort of 3007 adults undergoing same-day CXR and DXA. Ground truth was a lumbar spine DXA T-score ≤ -2.5. The primary metric was AUROC, alongside sensitivity and specificity. Image-only and clinical-only ablation models were evaluated on the same cohort to disentangle imaging from demographic contributions. Results: Among 3007 validation cases, 200 (6.7%) had a T-score ≤ -2.5. The AI achieved an AUROC of 0.942 (95% CI, 0.923-0.962), with sensitivity of 91.0% and specificity of 82.7%, and performance was preserved across study sites, sexes, and the clinically relevant older stratum (≥60 years). In ablation analysis, the image-only model achieved AUROC 0.931 (95% CI 0.912-0.950) and the clinical-only baseline 0.780 (0.732-0.827); non-overlapping 95% CIs (image-only higher) support an independent imaging contribution beyond demographics. Conclusions: This federated learning-based AI demonstrates high discrimination for identifying lumbar-spine DXA-defined osteoporotic-level bone mineral density (T-score ≤ -2.5) from standard CXRs and shows temporal transportability across two institutions without centralizing raw imaging data. Positioned as a triage adjunct to confirmatory DXA, it offers a scalable, privacy-preserving opportunistic screening solution.