Vertebral fractures identified on lateral DXA images by deep learning predict incident fractures in older women

Journal article, 2026

XVFA is an AI-based method for identifying vertebral fractures on DXA images. In 423 women followed for 8 years, vertebral fractures identified by XVFA or manual assessment were associated with a twofold increased risk of incident fractures. XVFA predicted fracture risk comparably to manual assessment, supporting automated vertebral fracture detection. Purpose Vertebral fractures (VFs), identified by vertebral fracture assessment (VFA) using dual-energy X-ray absorptiometry (DXA), predict incident fractures independently of clinical risk factors (CRFs) and bone mineral density (BMD). Most VFs remain clinically unrecognized. This study evaluated whether VFs identified using a deep learning-based method on lateral DXA images predict incident fractures comparably to manual VFA. Methods Associations between prevalent VFs and incident fractures were investigated in 423 women from the population-based SUPERB study who were not included in development of the explainable deep learning model (XVFA). Vertebrae were classified by manual VFA and XVFA. Incident fractures were X-ray verified. Cox proportional hazards models assessed fracture risk adjusted for CRFs and femoral neck (FN) BMD. Results Manual VFA reading and XVFA were used on baseline lateral images and classified 4563 and 5532 vertebrae, respectively, with numerical differences partly reflecting image quality limitations. VFs were identified in 102 women by manual VFA and 187 by XVFA. During 8 years of follow-up, incident fractures occurred in 48% of women with manual VFA VFs and 43% with XVFA VFs, vs 20% and 16% of women without VFs. Women with VFs had a higher fracture risk whether identified manually (HR 2.04; 95% CI, 1.35-3.07) or by XVFA (HR 2.32; 95% CI, 1.55-3.48), compared with women without VFs. Results remained significant after adjustment for CRFs and FN BMD. Conclusion Automated XVFA predicted incident fractures similarly to manual assessment. These findings support the clinical utility of deep learning-based VF detection, which may enhance fracture risk assessment and management in routine practice.