Interpretable machine learning model for low bone density screening in older adults using demographic and anthropometric data: findings from 2005 to 2020 NHANES

Rising osteoporosis from low bone mineral density (BMD) due to an aging population presents a major health challenge. Timely risk identification is vital for prevention. Traditional Dual-energy X-ray absorptiometry (DXA) diagnostics are expensive and inaccessible to all. This study aimed to create a machine learning (ML) model for low BMD screening using readily available demographic and anthropometric data. We analyzed the National Health and Nutritional Examination Survey (NHANES) 2005–2020 data of adults over 50, extracting demographic and anthropometric features. BMD was assessed via DXA, with low BMD defined as T-scores ≤ -1. We trained four ML algorithms—Logistic Regression (LR), Support Vector Machine with a linear kernel (SVM), Extreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost)—on this data, evaluating model performance using the area under the ROC curve (AUC), accuracy, sensitivity, specificity, precision, and F1-score, and used SHapley Additive exPlanations (SHAP) analysis to identify key features. A total of 13,133 participants were included in our study, with an average age of 64.88 years and a higher prevalence in males (52.2%) and Non-Hispanic Whites (46.9%). There were 7,209 individuals defined as low BMD (54.9%). Among the algorithms used, the CatBoost model showed the highest performance with an AUC of 0.840 in the training set and 0.822 in the testing set, accuracy of 0.757 and 0.751, sensitivity of 0.801 and 0.775, specificity of 0.703 and 0.720, precision of 0.765 and 0.782, and F1-score of 0.783 and 0.779, respectively. The CatBoost model identified weight, gender, age, waist circumference, and ethnicity as the top five most significant predictors of low BMD. The CatBoost model offers a hopeful method for early low BMD screening using accessible demographic and anthropometric data. Further validation in various populations and consideration of extra variables to boost performance are required. Not applicable.