Abstract / Summary
Osteoporosis has a high prevalence among older adults and poses severe health risks. Traditional screening methods such as dual-energy X-ray absorptiometry (DXA) are costly and inaccessible for widespread application in community settings. This study aimed to develop and validate a machine learning model for screening for high risk of self-reported physician-diagnosed osteoporosis in community‑dwelling older adults, thereby addressing the limitations of traditional screening methods. A cross-sectional study was conducted to collect data from community-dwelling older adults in Nanjing. Osteoporosis was defined by self‑report of a prior physician diagnosis (i.e., participants’ responses to a question about whether they had ever been told by a doctor that they had osteoporosis). The dataset was randomly divided into a training set (70%) and a test set (30%). Least Absolute Shrinkage and Selection Operator (LASSO) regression was used for variable selection, and six machine learning models were constructed: Logistic Regression (LR), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Naive Bayes (NB), and Light Gradient Boosting Machine (LightGBM). All preprocessing steps, including handling of missing data, were performed exclusively on the training data before being applied to the test data. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, and F1-score. Model calibration was assessed using calibration plots and the Brier score, and clinical utility was evaluated using decision curve analysis. Shapley additive explanations (SHAP) values were employed to interpret the key predictive factors of the optimal model. A total of 1,896 participants were included in this study, and the prevalence of self-reported physician-diagnosed osteoporosis was 9.9% (187). Eight predictive variables were finally selected: sex, age, household status, marital status, arthritis history, fracture history, parental fracture history, and hormonal medication use. In the test set, the LightGBM model showed the best performance, with an AUC of 0.841 (95% confidence interval (CI): 0.777–0.905), sensitivity of 0.848, and specificity of 0.667. The performance of other models was as follows: LR (AUC: 0.829), RF (AUC: 0.810), SVM (AUC: 0.634), XGBoost (AUC: 0.829), and NB (AUC: 0.827). The LightGBM model demonstrated good calibration and positive net benefit across clinically relevant threshold probabilities in decision curve analysis. SHAP analysis revealed that arthritis history, sex, and age were the most influential predictive factors in the LightGBM model. The machine learning models constructed based on easily accessible clinical features can effectively screen for self-reported physician-diagnosed osteoporosis risk in community-dwelling older adults. Among these, the LightGBM model has the best predictive performance within the present dataset. Future multi-center studies with diverse populations are needed to improve the robustness and generalizability of the model, providing support for the establishment of an efficient community-based risk assessment and case-finding system for self-reported physician-diagnosed osteoporosis.