AI Screening of Routine Chest X-Rays for Osteoporosis Is Cost-Effective Only If Patients Follow Through

A cost-utility analysis published in Archives of Osteoporosis finds that using a deep learning model to analyze routine chest radiographs for opportunistic osteoporosis screening among Taiwanese postmenopausal women is cost-effective compared with no screening, but only under the assumption that every screen-positive patient completes confirmatory bone density testing and begins pharmacological treatment. When real-world completion rates observed in Taiwan were applied instead, the approach no longer met the prespecified economic threshold. The modeling study, which simulated a cohort of 50-year-old postmenopausal women over a 50-year horizon, concludes that strengthening the post-screening care pathway is central to realizing the economic value of this technology.

Osteoporosis is a systemic skeletal disorder characterized by reduced bone mineral density, leading to an increased risk of fragility fractures. It predominantly affects postmenopausal women and often remains undiagnosed until a fracture occurs, owing to its asymptomatic progression. These fractures substantially reduce quality of life and generate a significant economic burden. According to the study’s background, the global cost of osteoporosis treatment in 2018 was estimated at approximately $5.3 million, accounting for nearly 20% of gross domestic product. Early detection is paramount for initiating preventive strategies before a fracture occurs, yet the gold standard for diagnosis, dual-energy X-ray absorptiometry (DXA), remains underutilized due to limited accessibility, cost, and the lack of universal screening mandates. This diagnostic gap highlights an unmet need for more accessible screening approaches.

The research team, led by Bo-Xiang Shi and Wen-Hui Fang of institutions in Taiwan, employed a hybrid decision tree and Markov model from the Taiwan healthcare perspective. The deep learning model adopted in the analysis was developed and validated on chest radiographs obtained during routine clinical care at a medical center in Taiwan and was prospectively evaluated in the OPSCAN randomized controlled trial, in which the screening workflow was applied to patients receiving chest radiography for indications other than osteoporosis. In the modeled pathway, the model was applied to existing images rather than newly acquired ones, so no additional radiation exposure or separate imaging cost was included. Individuals with positive screening results were assumed to be notified or referred for confirmatory DXA, and pharmacological treatment was initiated only among those with DXA-confirmed osteoporosis. Based on clinical expert input, 75% of treated individuals were assumed to receive denosumab, with the remainder receiving bisphosphonates.

After initial management, all individuals entered a Markov model comprising seven health states to simulate long-term outcomes, including hip, vertebral, and other fragility fractures, as well as death. Fracture incidence, mortality, and direct medical costs were obtained from domestic Taiwanese sources, and utility values were synthesized from the literature with preference given to Taiwanese or Asian cohorts. All costs were reported in 2025 US dollars at a conversion rate of 31 New Taiwan dollars to 1 US dollar, and an annual discount rate of 3% was applied to both costs and effectiveness outcomes, in accordance with Taiwanese health economic evaluation guidelines. The cost per deep learning-based screening was assumed to be $6.45, equivalent to one chest radiograph, while the cost of DXA was sourced from the National Health Insurance pricing dataset in June 2025.

In the base-case scenario, which used a sensitivity of 92.9% and a specificity of 68.1% and assumed full care-pathway completion, the screening approach generated 1,194 additional life-years compared with the no-screening arm in a simulated cohort of 1,000 individuals. The average cost was higher in the screening group at $12,776 compared with $1,226, while the average quality-adjusted life year (QALY) value was 26.99 compared with 26.46. The resulting incremental cost-effectiveness ratio (ICER) was $21,827 per QALY gained, below the willingness-to-pay threshold of $39,492, corresponding to one gross domestic product per capita in Taiwan in 2025. Because Taiwan has no official willingness-to-pay threshold for economic evaluations, the study adopted this conservative reference value for interpreting its results.

The analysis also examined how variations in disease prevalence and diagnostic performance influenced economic value. Screening was considered cost-effective when osteoporosis prevalence exceeded 14.3%. To assess diagnostic thresholds, the researchers evaluated 371 unique sensitivity and specificity combinations generated by varying the classification threshold along the receiver operating characteristic curve while holding overall discriminatory performance constant. The greatest economic benefit was observed at a sensitivity of 86.6% and a specificity of 76.3%, yielding a net monetary benefit of $9,453 and an ICER of $20,319. Notably, the base-case operating point generated approximately 26 missed cases per 1,000 individuals, whereas the economically optimal operating point generated approximately 48 missed cases. The authors indicate that threshold selection should therefore consider not only cost-effective results but also the acceptable number of missed cases in the intended clinical setting.

Robustness checks supported the base-case conclusion. A deterministic one-way sensitivity analysis, which varied costs by ±20% and other parameters by ±10%, identified the cost of DXA as the most influential parameter affecting the ICER, followed by osteoporosis prevalence and the incidence of hip or vertebral fracture, although none of the varied parameters changed the primary conclusion. A probabilistic sensitivity analysis based on 5,000 Monte Carlo simulations indicated that the screening approach had a 90.9% probability of being cost-effective at the willingness-to-pay threshold of one gross domestic product per capita. The modeled prevalence of 36% falls within the range reported in international economic evaluations of similar screening approaches conducted in Japan, South Korea, Singapore, the United States, and Germany, which consistently reported cost-effective results under their respective modeling assumptions.

Scenario analyses, however, showed that the economic value of screening depends heavily on the completion of the post-screening care cascade. In a one-way analysis of confirmatory DXA examination rates, the most favorable outcome occurred under the base-case assumption that 100% of screen-positive individuals received DXA. As the examination proportion decreased, economic outcomes progressively worsened: screening remained cost-effective at a DXA examination rate of 20%, with an ICER of $38,146, but the ICER exceeded the willingness-to-pay threshold at lower rates, including the OPSCAN-observed rate of 14.6%, where it reached $45,691. In a two-way analysis, when the OPSCAN-observed DXA examination rate of 14.6% and the Taiwan-based pharmacological treatment initiation rate of 28.7% were applied simultaneously, the approach was no longer cost-effective.

Real-world evidence from the OPSCAN trial underscores this implementation challenge. In the trial, 12.8% of high-risk participants underwent the fully reimbursed DXA examination offered by the study, while an additional 1.8% underwent DXA arranged by their primary care physicians, resulting in an overall completion rate of 14.6%. The authors suggest that in a feasible implementation pathway, the deep learning model would be applied to existing chest radiographs within routine radiology or electronic medical record workflows, positive results would be flagged for physician review, and screen-positive patients would be informed that the result indicates elevated risk rather than a definitive diagnosis before referral for confirmatory DXA. Follow-up programs such as patient notification or regional coordination are identified as key targets for improving real-world implementation.

The study acknowledges several limitations. Some model inputs were derived from multiple studies or comparable settings, which may introduce uncertainty. The base-case assumptions of complete confirmatory testing and complete treatment initiation may not reflect real-world practice, and implementation data remain limited. Treatment effects were represented using medication class and persistence-specific hazard ratios, so changes in adherence, delayed dosing, treatment switching, or reinitiation were not simulated, and routine monitoring and adverse event costs were excluded, which may overestimate effectiveness or underestimate costs. Osteoporosis prevalence derived from claims data may underestimate the true burden because of underdiagnosis, and annual fracture incidence was applied as a fixed transition probability because suitable age-specific Taiwanese data were unavailable, a simplification that may underestimate fracture events in later life. The cohort Markov structure also did not capture patient-level heterogeneity, individual fracture history, or the cumulative effects of sequential fractures. The authors state that future studies incorporating prospective implementation data and real-world treatment pathways are warranted.

Subject of Research: Cost-utility analysis of deep learning-based opportunistic osteoporosis screening using chest radiographs in Taiwanese postmenopausal women

Article Title: Cost-utility analysis of deep learning-based opportunistic osteoporosis screening using chest radiographs: implications of diagnostic thresholds and confirmatory testing

Article References: Shi, B.-X., Tsai, D.-J., Lin, C., Lin, S.-H., Hsieh, P.-H., & Fang, W.-H. (2026). Cost-utility analysis of deep learning-based opportunistic osteoporosis screening using chest radiographs: implications of diagnostic thresholds and confirmatory testing. Archives of Osteoporosis, 21(1), Article 149. https://doi.org/10.1007/s11657-026-01788-5

Image Credits: AI Generated

DOI: 10.1007/s11657-026-01788-5

Keywords: osteoporosis, deep learning, chest radiographs, cost-utility analysis, opportunistic screening, DXA, postmenopausal women, Taiwan, health economics, fragility fractures, care pathway, diagnostic thresholds