The authors developed a convolutional neural network named Osteo-AI to enable opportunistic screening for osteoporosis using routinely obtained chest x-rays. The stated motivation was to address barriers to conventional screening in regions such as the Philippines, where access to DXA—the diagnostic gold standard for bone mineral density—is limited. By leveraging chest radiographs, which are more frequently performed and relatively inexpensive, the investigators aimed to create a tool that could flag patients at risk and prompt further evaluation.
The development and initial testing of Osteo-AI used a dataset of 406 paired chest radiographs and DXA scans from Filipino patients aged 50 years and older. The sample included 369 female patients and 37 male patients. Patient ages ranged from 50 to 89 years, with a mean age of 63 years. All ground-truth labels were based on DXA results, which served as the reference standard for osteoporosis detection in this study.
To expand the effective size of the training corpus, the research team applied data augmentation techniques to the original 406 image–DXA pairs. Augmentation increased the training set to a total of 6,300 paired examples. The authors report using these expanded datasets to train Osteo-AI, acknowledging that augmentation was intended to improve model generalizability in the context of a relatively small original cohort.
The model is described as a convolutional neural network specifically trained to detect imaging patterns on chest radiographs that correlate with DXA-defined osteoporosis. To provide visual explanations of model predictions and to localize image areas that contributed to classification, the investigators applied gradient-weighted class activation mapping (Grad-CAM). Grad-CAM outputs were used to identify regions and patterns on the chest x-rays that the network associated with osteoporosis, offering an interpretability layer that may help clinicians assess the plausibility of the model’s focus areas.
In initial testing, Osteo-AI achieved a diagnostic accuracy of 85.71%. The authors note that this performance substantially exceeded a benchmark accuracy of 33.33% used for comparison in their analysis. The report presents these aggregate accuracy figures but does not provide additional performance metrics such as sensitivity, specificity, positive predictive value, negative predictive value, receiver operating characteristic curves, confidence intervals, or subgroup analyses by sex or age in the source text.
Ethical oversight for this work was provided by the Ethics committee/IRB of The Medical City, and the authors state that all relevant ethical guidelines were followed and that patient consent procedures were completed. The authors declared no competing interests. The preprint indicates that data are available upon request; no public dataset or external repository link is provided in the source document.
The authors conclude that Osteo-AI shows promise as an opportunistic screening tool to increase accessibility of osteoporosis detection, with potential to support earlier intervention and to reduce risk of fragility fractures. They emphasize that further research using larger and more diverse datasets is necessary to refine and optimize the model, to improve detection accuracy, and to evaluate generalizability across populations and imaging settings. The study is presented as a preprint on medRxiv (posted August 24, 2026), and the authors recommend additional validation before clinical deployment.
The source highlights the limited original sample size (406 paired studies) and reliance on augmentation to expand training data, which implies a need for external validation. The published preprint provides aggregate accuracy but does not report other diagnostic performance metrics or detailed error analysis in the text provided. No information on how the benchmark of 33.33% was derived is included in the source article. The authors call for larger datasets and further research to address these gaps.
If validated in larger and more varied cohorts, an automated screening approach using chest x-rays could support case-finding for osteoporosis in clinical settings where DXA access is constrained. The Grad-CAM visualizations reported by the authors could facilitate clinician review of flagged cases by showing image regions driving model predictions. The source emphasizes potential to expand screening reach and enable earlier intervention, while noting that model refinement and broader validation are required before routine clinical use.
The authors confirmed adherence to research reporting guidelines and stated that any necessary IRB approvals were obtained. Data are available on request. The preprint reports no competing interests. The report does not include trial registration identifiers or a link to a public code repository in the provided text.