Abstract

Background: Low bone mineral density (BMD) is a common age-related condition that elevates the risk of fractures and mortality. Machine learning (ML) techniques offer a promising approach for early prediction using readily available clinical, biochemical, and demographic data.

Aims: To evaluate the predictive performance of eleven ML models in identifying low BMD and to determine the most influential risk factors using the best-performing model.

Study Design: Cross-sectional study.

Methods: Data were obtained from National Health and Nutrition Examination Survey (2005-2010, 2013-2014, and 2017-2020), focusing on individuals aged ≥ 50 years with available femoral neck or total femur BMD data. After applying exclusion criteria, 12,108 participants were included. Supervised ML algorithms were trained using 57 clinical, biochemical, demographic, and behavioral features. Model performance was assessed using accuracy, area under the curve (AUC), recall, precision, and F1 score. SHAP analysis was employed to interpret model outputs and rank predictors.

Results: The extra trees classifier outperformed other ML methods, achieving an accuracy of 76.7% and an AUC of 0.85. Recursive Feature Elimination with Cross-Validation identified 14 key predictors of low BMD in descending order of importance: sex, age, body mass index, race, family income-to-poverty ratio, serum uric acid, diabetes status, HDL cholesterol, urinary creatinine, alkaline phosphatase, mean cell volume, lymphocyte count, diastolic blood pressure, and glycohemoglobin.

Conclusion: Tree-based ML models, particularly Extra Trees, can effectively predict low BMD. The identified risk factors include both established and lesser-studied predictors. These findings support the use of ML for personalized osteoporosis and osteopenia screening and highlight its ability to capture complex, multifactorial relationships in population health data.