A chest X-ray could cut bone-density scans in half—if a simulation survives clinical practice

A chest radiograph is usually ordered to examine the lungs, heart or acute symptoms. It also contains the ribs, spine and other skeletal information. AI-OsRM was trained to use that existing image as a prescreen for osteoporosis and reduced bone mineral density, potentially identifying people who should receive the dedicated DXA test without adding another initial scan.[1]

The distinction between prescreening and diagnosis is essential. DXA remains the reference examination in this study. The AI produces a risk score from the chest image; it does not directly measure bone density or decide treatment. Its clinical value would come from changing who is referred for DXA while missing very few people with disease.

The same model was tested in different settings

In the SNUH cohort, the model reached an AUROC of 0.88 for both osteoporosis and reduced BMD. For osteoporosis in the two health-checkup cohorts, it reached 0.93 in HPC and 0.92 in YSMC. AUROC measures how well scores rank positive cases above negative cases across thresholds. It does not specify the sensitivity, specificity or workload at the threshold a hospital would actually use.[1]

The different cohorts matter because hospital patients and people attending health checks can differ in age, disease prevalence, image acquisition and referral patterns. Stable discrimination across them is stronger evidence than a single internal split. It still does not establish performance across countries, equipment, ethnic groups or portable radiography without further validation.

Why reducing scans is not the only objective

A screening system changes two queues. It reduces tests for people below the threshold, but it may also uncover high-risk people who would never have been considered for osteoporosis assessment. The appropriate measure is therefore not simply DXA scans saved. It is fractures prevented, time to confirmed diagnosis, false reassurance, unnecessary referrals and access for groups currently underscreened.

Prevalence will also affect the workload. AUROC can remain similar while the number of false positives changes sharply between a hospital population and a broad screening population. A threshold selected to retain 99% of osteoporosis cases may refer many people without osteoporosis when prevalence is low.

The image can contain shortcuts as well as bone information

The model may learn true skeletal signals, correlates such as age and body composition, or image patterns associated with who receives both CXR and DXA. External validation reduces concern that one hospital artifact explains the result, but prospective monitoring should still examine performance by age, sex, equipment, clinical indication and image quality.

A safe implementation would show the score as a referral aid and keep the reason for the original chest radiograph separate. Clinicians should not reinterpret a normal-looking lung study as a normal bone-density result, or treat a high AI score as confirmed osteoporosis.

A clinical trial needs an operational denominator

The next study should begin with everyone eligible for the workflow, not only people who already have paired CXR and DXA. It should record how many chest radiographs were scored, how many alerts reached clinicians, how many people accepted DXA, how many diagnoses were confirmed and how treatment or fracture outcomes changed. That denominator reveals where an apparently accurate model loses value in practice.

Opportunistic screening is attractive because the image already exists. The important question is whether the second use creates a reliable route to care rather than another silent score in the record. AI-OsRM has unusually large external cohorts and a promising simulation. Its decisive test will be prospective: does it help the right people receive DXA and fracture prevention sooner without creating a new burden of missed or unnecessary care?

Report an error or suggest a correction