Machine learning-driven clinical decision support for low bone mineral density: A web-based prediction model with explainable AI integration

• Published in: Bone (New York, N.Y.) Vol. 200; p. 117592
• Main Authors: Yang, Xing, Liu, Jianyuan, Huang, Xiaozhi, Liang, Hao, Cui, Ping, He, Shiran, Zhang, Heng, Liao, Wenping, Zhang, Guangkun, Huang, Qianqian, Ning, Huan, Luo, Tingyan, Luo, Yinghua, Li, Wei, Huang, Jiegang
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.11.2025
• Subjects: LIME; Logistic regression; Low bone mineral density; Machine learning; Prediction model; Risk factors; SHAP; Logistic regression; SHAP; Low bone mineral density; LIME; Machine learning; Prediction model; Risk factors
• ISSN: 8756-3282; 1873-2763; 1873-2763
• DOI: 10.1016/j.bone.2025.117592

Low bone mineral density (LBMD), which includes osteopenia and osteoporosis, is associated with substantial health care costs. However, current diagnostic methods for LBMD are limited in terms of accuracy and accessibility. This study aims to develop an interpretable machine learning model for LBMD risk assessment and implement it as a web-based clinical decision support tool. Data from subjects who underwent dual-energy X-ray absorptiometry (DXA) at the People's Hospital of Guangxi Zhuang Autonomous Region were collected and randomly divided into a training set (70 %) and an internal validation set (30 %). An external validation set was sourced from the National Health and Nutrition Examination Survey (NHANES) database. Least absolute shrinkage and selection operator (LASSO) regression and multiple logistic regression were used for feature selection. Ten common machine learning models were conducted based on the selected features. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), Matthews correlation coefficient (MCC), Brier score, and decision curve analysis (DCA). The decision mechanisms of the best-performing model were explained using SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). The optimal model was deployed as a web application using Streamlit. A total of 16,274 participants were included in this study. Age, body mass index (BMI), alkaline phosphatase, and total cholesterol were identified as key predictors of LBMD. The logistic regression (LR) model demonstrated superior prediction performance (internal validation set [AUC = 0.902, MCC = 0.684, Brier score = 0.123], external validation set [0.812, 0.358, 0.265]). DCA confirmed its clinical utility. Both SHAP and LIME showed consistent results in identifying predictive factors. The LR model was deployed as a web application to predict LBMD. Our interpretable machine learning model and web-based implementation provide a free and reliable tool for predicting LBMD, which represents a significant advancement in making LBMD screening more accessible and cost-effective. [Display omitted] •Compared ten machine learning models; best discrimination with logistic regression•Validated the model externally using the NHANES database (LR model AUC 0.812)•Used SHAP and LIME analyses to interpret and verify variable importance•Developed a web app for LBMD prediction based on the optimal model•Only four variables included (age, BMI, alkaline phosphatase and total cholesterol)