This study develops an explainable hybrid machine learning framework, which is optimized by an improved Chaotic Dung Beetle Optimization (CSDBO) algorithm, to enhance the accuracy of osteoporosis (OP) risk prediction. Based on 1,537 clinical samples and 39 clinical and biochemical variables obtained from the Harvard Dataverse, the Boruta algorithm was employed to identify 12 key predictors. CSDBO was then used to perform intelligent hyperparameter optimization and model selection for multiple ensemble learning algorithms, including LightGBM, GBDT, and XGBoost. After optimization, LightGBM achieved the best performance, with a training AUC of approximately 0.998 and a test AUC exceeding 0.85. SHAP-based interpretability analysis indicated that femoral neck BMD, total lumbar T-score, and calcitriol were among the most influential factors. The proposed framework improves predictive accuracy and model stability while maintaining high interpretability, demonstrating potential value for clinical risk assessment and individualized intervention. Detailed algorithmic formulations and implementation procedures are provided in the Methods section.
Meng et al. (2026) studied this question.