Cost-Effectiveness of Opportunistic Osteoporosis Screening using Chest Radiographs with Deep Learning in Germany

Objective(s): Deep learning models applied to chest radiographs can support opportunistic osteoporosis screening. This study evaluates the cost-effectiveness of this approach in German women aged 50 years and over.

Materials and Methods: A microsimulation Markov model estimated the cost per quality-adjusted life year (QALY) gained (€2024) for screening via chest radiographs with deep learning, followed by treatment, versus no screening and treatment. The patient pathways were based on the sensitivity and specificity of AI-enhanced radiographs and risk stratification per German osteoporosis guidelines. Women with a fracture risk below 5% received no treatment; those with 5–10% risk were prescribed alendronate, and those aged 65+ with risk >10% received sequential treatment starting with romosozumab. Assumptions and data were validated by a German clinical expert. Real-world medication persistence, realistic assumptions for probabilities of DXA examination post-screening detection and for treatment initiation rates were incorporated. Parameter uncertainty was analyzed through sensitivity analyses.

Results: Screening improved health outcomes, yielding more QALYs and fewer fractures while increasing treatment costs. The cost per QALY gained of opportunistic screening was estimated at €13,340 for all women aged 50+, well below the commonly accepted cost-effectiveness thresholds, which are generally in the range of €60,000. Optimizing follow-up, treatment initiation, and adherence enhanced cost-effectiveness, with dominance (higher QALYs at lower costs) achievable by halving medication non-persistence.

Conclusion(s): Opportunistic osteoporosis screening using chest radiographs with deep learning is a cost-effective strategy for German women aged 50 and older under real-world conditions. Improving screening follow-up and medication adherence could further enhance cost-effectiveness and achieve dominance.

Acknowledgments: This study was funded by an unrestricted educational grant from Promedius.