This study systematically examined the impact of three feature selection techniques (Boruta, Extreme gradient boosting (XGBoost), and Lasso) for optimizing four machine learning models (Random forest (RF), XGBoost, Logistic regression (LR), and Support vector machine (SVM)) in predicting bone density prevalence. Our findings revealed that varying data partitioning ratios (training and test sets: 0.6:0.4; 0.7:0.3; 0.8:0.2; 0.9:0.1) minimally impacted the prediction accuracy across all four models, a conclusion reinforced by 10-fold cross validation. Besides, principal component analysis (PCA) led to substantial accuracy degradation (0.6-0.8 range), suggesting incompatibility with this study's requirements due to the inherent complex decision boundaries in the original high-dimensional data. Comparative analysis demonstrated that the Boruta-XGBoost combination achieved superior performance (accuracy: 0.9083 ± 0.0146), significantly outperforming the Lasso-LR combination (0.7480 ± 0.0157) across all evaluation frameworks. Regarding model evaluation metrics, the RF model exhibited enhanced discriminative capacity with Area under the receiver operating characteristic (AUROC) values of 0.85, 0.81, and 0.80 under different feature selection approaches, surpassing the SVM model (0.78, 0.76, and 0.76). This advantage likely stems from RF's native capability to capture non-linear relationships and feature interactions.