A deep learning–based vertebral fracture assessment tool predicted future fracture risk in older women as effectively as physician-interpreted assessment, according to a prospective cohort study published in Osteoporosis International.  

“These findings support the clinical utility of deep learning–based vertebral fracture detection, which may enhance fracture risk assessment and management in routine practice,” wrote researchers, who were led Mattias Lorentzon, MD, from the University of Gothenburg in Sweden.  

Researchers  evaluated Xplainable Vertebral Fracture Assessment (XVFA), an artificial intelligence–based method designed to identify vertebral fractures on lateral dual-energy x-ray absorptiometry (DXA) scans. The testing group consisted of 423 Swedish women aged 75 to 80 years from the population-based SUPERB cohort who were followed for a median of approximately 8 years.  

Baseline vertebral fractures were identified using both conventional vertebral fracture assessments performed by physicians and XVFA. Incident fractures were confirmed through a radiographic review of regional imaging records. Cox proportional hazards models adjusted for age, anthropometric measures, Fracture Risk Assessment Tool (FRAX) clinical risk factors, and femoral neck bone mineral density were used to estimate fracture risk.  

After adjustment for FRAX clinical risk factors and femoral neck bone mineral density, vertebral fractures identified by either XVFA or manual assessment were associated with approximately twofold higher risk of subsequent major osteoporotic fracture. XVFA-detected vertebral fractures were also associated with a significantly increased risk of hip fracture, whereas manually identified vertebral fractures were not.   

Manual assessment identified vertebral fractures in 102 women (24%), whereas XVFA identified fractures in 187 women (44%). During follow-up, incident fractures occurred in 48% of women with manually identified vertebral fractures versus 20% of those without vertebral fractures; corresponding rates were 43% and 16%, respectively, among women with and without XVFA-detected vertebral fractures.   

After adjustment for FRAX clinical risk factors and femoral neck bone mineral density, vertebral fractures identified by either method were associated with approximately a twofold higher risk of major osteoporotic fracture. Manual VFA-detected vertebral fractures were associated with nearly a fivefold increased risk of incident vertebral fracture. In contrast, XVFA-detected vertebral fractures were associated with more than a fourfold increased risk of incident vertebral fracture and nearly a threefold increased risk of hip fracture.  

Notably, women with vertebral fractures detected by XVFA but not by manual assessment had more than twice the risk of incident fracture after adjustment for clinical risk factors and femoral neck bone mineral density, suggesting that XVFA may identify vertebral deformities with prognostic significance that are not captured by routine manual assessment.   

The researchers noted several limitations. XVFA performance depends on accurate detection of vertebral landmarks, and image quality may affect classification. The model analyzed more vertebrae than manual assessment and was not trained to exclude poorly visualized vertebrae, which may have increased sensitivity and also contributed to the substantially higher number of vertebral fractures detected by XVFA. The study population also consisted exclusively of ambulatory older women, possibly limiting generalizability to other populations. Statistical power was also limited for some hip fracture analyses.  

Lead researcher Mattias Lorentzon, MD, reported receiving lecture or consulting fees from multiple pharmaceutical companies, and Lisa Johansson reported receiving lecture fees from UCB Pharma. The remaining researchers reported no conflicts of interest.