Osteoporosis on chest radiographs: what AI-OsRM is worth, validated on three external cohorts, before real-world evidence

A Korean team (Seoul National University Bundang Hospital and partners) trained AI-OsRM, a ResNet pretrained by image masking, to spot osteoporosis on a plain chest radiograph, then tested it on three external cohorts totalling 153,058 people. AUROCs range from 0.88 to 0.93, and a simulation suggests that using the model as a prescreen could cut bone densitometry (DXA) by about 50% while still catching 98 to 99% of osteoporosis cases. Discrimination is strong and the validation scale is rare, but the accessible version gives no thresholds, no comparator and no prospective test: this is a promise of opportunistic screening, not yet a demonstration.

The context

Osteoporosis is a weakening of the skeleton that raises fracture risk. Its diagnosis rests on DXA (dual-energy X-ray absorptiometry), which measures bone mineral density (BMD) and compares it with a reference as a "T-score". The problem is well known: DXA is under-prescribed, and many at-risk people are never tested.

Yet hundreds of millions of chest radiographs are taken every year for other reasons. The idea of "opportunistic" screening is to reuse these existing images to flag, in passing, the people who deserve a DXA. The bone visible on a chest radiograph (vertebrae, ribs, clavicles) carries information about density, but it is hard to read by eye. Deep learning models have already been proposed for this task; the question here is robustness, meaning how the model holds up outside the hospital where it was built.

The method

The model. AI-OsRM is a ResNet-type network (a convolutional network with residual connections, a classic vision architecture) pretrained as a masked autoencoder: large parts of each image are hidden and the network must reconstruct them, which forces it to learn the structure of radiographs without labels. It is then fine-tuned to predict two targets: osteoporosis and a broader "reduced BMD" category. The parameter count is not given in the accessible version.

Development data. 46,478 examinations pairing a chest radiograph with a DXA, in 26,757 patients of a tertiary referral hospital, between 2016 and 2022. DXA serves as the reference label.

External validation. Three separate cohorts: SNUH (6,391 people, a hospital cohort separated in time), HPC (24,132 people, health check-up) and YSMC (122,535 people, health check-up). External validation means testing the model on people who were not used to build it and who come from other places or periods; it is the most important test for an imaging model.

The main metric. The AUROC (area under the ROC curve) measures the ability to rank a person with the disease above one without: 0.5 is chance, 1 is perfect separation. An AUROC of 0.9 means that, for a random pair of one diseased and one non-diseased person, the model gives the higher score to the diseased one nine times out of ten.

What the accessible version does not state. We could only read the abstract and metadata of this early-access "accepted manuscript", which may still change. The exact definitions (T-score cut-offs), the interval between radiograph and DXA, exclusion criteria, the training / validation / internal test split, prevalence per cohort and operating thresholds are not there; they are probably in the PDF and its supplements (including a STROBE checklist), which we could not consult. What follows is therefore limited to what the abstract publishes.

The results

ML metrics. For osteoporosis, the AUROC is 0.88 (95% CI 0.87–0.89) on SNUH, 0.93 (0.92–0.93) on HPC and 0.92 (0.92–0.92) on YSMC. For reduced BMD it is 0.88 (0.87–0.89) on SNUH; the corresponding values for HPC and YSMC are not given in the abstract. The intervals are very narrow, which reflects sample size more than clinical certainty.

Prescreening simulation. The authors report that using the model as a filter before DXA could reduce DXA use by about 50% while detecting 98 to 99% of osteoporosis cases and 80 to 90% of reduced-BMD cases. The threshold used, the simulation method, and the associated sensitivity and specificity are not detailed in the abstract.

Clinical translation. These figures read as follows: out of 100 people who truly have osteoporosis going through this filter, 1 to 2 would be missed and wrongly routed away from DXA; out of 100 people with reduced BMD, 10 to 20 would be missed. In exchange, half of the DXA scans would not be performed. We cannot go further without prevalence and specificity: without them, we do not know how many healthy people would be sent to DXA needlessly, nor how many fractures that would prevent or cause.

What is good

Three large external cohorts. 6,391, 24,132 and 122,535 people, 153,058 in total, including one temporal hospital cohort and two health check-up cohorts. Few imaging models are tested at this scale outside their home centre.

A substantial development set and a solid reference. 46,478 radiograph–DXA pairs from 26,757 patients, with DXA, the reference examination, as the label rather than a subjective reading of images.

Performance that holds across different populations. The AUROC stays between 0.88 and 0.93 from a hospital setting to a health check-up setting, and the approach is designed for realistic use (reusing an existing image). Public funding (National Research Foundation of Korea, hospital research fund) and open access make critical reading easier.

What is less good

A simulated prescreen, not a measured one. The 50% DXA reduction comes from a simulation, with no threshold published in the abstract, no prospective study and no measured effect on decisions or fractures. The 98 to 99% detection depends entirely on the chosen threshold, which is not stated: this is the classic misleading-metric case, where a high AUROC says nothing about the real operating point. For reduced BMD, up to 20% of cases are missed.

No comparator and a risk of population bias. No comparison with simple tools (age, sex, body mass index, clinical risk scores) or with radiologists appears in the accessible version. Yet age and sex already predict osteoporosis well: without that comparator, we do not know what the image really adds. Moreover, all the data appear to come from South Korea; performance on other populations, other devices or other body types is unknown. Health check-up cohorts, made of relatively healthy people, may also have a different prevalence and profile from the patients one actually wants to screen.

Many ties of interest and incomplete transparency. The competing-interests statement says "no competing interests related to this article", but authors separately declare consulting roles, stock or grants with medical AI companies (Crescom, Coreline Soft, Lunit, Promedius, Yeji X), and one author is affiliated with Osstem Implant, a dental device company. Code, weights and data are not mentioned in the abstract; the article's licence (CC BY-NC-ND 4.0) says nothing about the model's. A risk of shortcut learning (the model relying on a spurious cue such as apparent age, sex or body build rather than bone) is not discussed in what is accessible.

What it changes

For the research community. The paper sets a reference level for opportunistic screening on chest radiographs, with an external-validation scale rarely reached. The logical next steps: compare against simple clinical variables, publish thresholds, test outside Korea, and run a prospective study measuring avoided DXA and missed fractures.

For clinicians. Nothing changes today: this is an early-access manuscript, a research model with no regulatory clearance mentioned, and the decision to order a DXA remains the physician's. In time, such a tool could act as a discreet signal added to radiographs already taken, provided its false-alarm rate is known.

For patients and the public. Osteoporosis is often silent until the first fracture. Screening that requires no extra radiation or visit is appealing, but it is not yet known how many people would be falsely reassured or needlessly worried.

Going further

The paper is published in npj Digital Medicine: doi.org/10.1038/s41746-026-03333-7 (received 3 April 2026, accepted 18 September 2026, published online 9 October 2026). The PDF and its supplements, including the STROBE checklist, contain the methodological details we could not read. No code repository is mentioned in the accessible parts.

Editorial transparency: this article was produced with the help of artificial intelligence tools from the published abstract and metadata (the full text was not accessible); the English, Spanish and Chinese versions are machine translations.