Predicting Low Bone Density Based on Interpretable Machine Learning Models

Osteoporosis has an insidious onset and low bone density is an early stage of osteoporosis, in order to recognize and intervene in osteoporosis at an early stage, this study aimed to develop a predictive model for low bone density. This study used National Health and Nutrition Examination Survey (NHANES) data from 2013-2014 and 2017-2018 with Decision Tree (DT), Random Forest (RF), Gradient Boosting Decision Tree (GBDT) and XGBoost algorithms to construct the prediction model, and SHAP algorithm for visual interpretation. The results showed that the best prediction model, XGBoost, had an AUC value of 0.777, which ultimately determined that the characteristics of gender, age, obesity or not, race, marital status, number of co-morbidities, and dietary pattern had a good prediction effect. The prediction model of low BMD constructed in this study can help healthcare professionals to recognize low BMD as early as possible and develop targeted measures to prevent osteoporosis.