A machine learning model predicted osteoporotic fracture risk in postmenopausal women with high accuracy over 8 to 10 years, according to a study published in Scientific Reports.
Researchers evaluated 576 postmenopausal women across two independent cohorts with long-term clinical follow-up. The HURH cohort included 276 postmenopausal women diagnosed with osteoporosis, 72 of whom experienced an osteoporotic fracture. The Camargo cohort consisted of 300 postmenopausal women from the general population, and 91 fractures were observed during follow-up.
The predictive model was developed in the HURH cohort and externally validated in the Camargo cohort. The researchers tested two groups of variables. One incorporates all available clinical and densitometric variables, and the other is limited to measurements that are more readily available in general medical consultations.
Across the tested approaches, Extreme Gradient Boosting demonstrated the strongest predictive performance. In postmenopausal women with osteoporosis, the area under the curve (AUC) reached 0.88 using all variables and 0.92 using a streamlined variable set. During external validation in a general population cohort, performance remained stable at 0.88 for both variable groups, noted Ricardo Usategui Martín, Ph.D., of the University of Valladolid, Spain, and colleagues.
In postmenopausal women with osteoporosis, the most influential factor predicting future fractures was previous fractures. Other influential variables include parathormone levels and lumbar T-scores. When the model was limited to more readily available clinical measures, the most influential predictors were previous fractures, parathormone, lumbar T-score, and vitamin D levels.
More complex models incorporating trabecular bone scores and three-dimensional dual-energy X-ray absorptiometry did not improve predictive performance compared to simpler models based on commonly available clinical variables.
The researchers noted that despite its role in bone metabolism, parathormone is not included in widely used fracture risk algorithms for postmenopausal women. Vitamin D levels were also identified as an important contributor to fracture risk prediction in both cohorts.
Bone densitometry across the spine, femoral neck, and hip contributed to the simplified model. Although bone mineral density is commonly used for diagnosis and treatment decisions, fractures frequently occur in patients with osteopenia or normal bone mineral density.
The researchers acknowledged some limitations. The cohort was recruited in Spain, which may limit generalizability to other populations. Fracture occurrence was modeled as a dichotomous outcome at the end of follow-up without incorporating the timing of fracture. This means that the model does not capture temporal dynamics such as impending fracture risk. Additionally, the sample size did not allow detailed analysis by fracture type or location.
“Machine learning should be used to identify postmenopausal women at high risk of fractures. This study summarizes that previous fractures, DXA, PTH, and vitamin D play important roles in identifying these women,” said Dr. Usategui-Martín and colleagues.
The researchers reported no competing interests.
Source: https://doi.org/10.1038/s41598-025-27226-z
