Application of machine learning to identify predictors of bone health in older male populations in Taiwan

Approximately 1 in 8 men over the age of 50 will experience a fragility fracture. Unlike women, men exhibit a gradual decline in bone mineral density beginning in their 60s. While multiple linear regression (MLR) has been the standard approach in osteoporosis risk modeling, machine learning (ML) may better capture complex predictor–outcome relationships. We analyzed data from 2,870 healthy Taiwanese men aged 65–70 years and compared MLR with three ML algorithms—Stochastic Gradient Boosting (SGB), XGBoost, and Elastic Net (EN)-for prediction of T-scores. Model performance was evaluated using three continuous-error metrics (RAE, RRSE, and RMSE) across 10 repeated train–test splits with internal cross-validation for hyperparameter tuning. Key predictors were interpreted using Shapley Additive Explanations (SHAP) applied to the final XGBoost model. EN significantly outperformed MLR across all three continuous-error metrics, reducing RMSE to 1.245 (vs. 1.252, p = 0.029), RAE to 0.962 (vs. 0.974, p = 0.008), and RRSE to 0.958 (vs. 0.964, p = 0.031), whereas SGB and XGBoost showed no significant improvement over MLR (all p > 0.99). Across models, the six most influential predictors were alkaline phosphatase (ALP), body fat (BF), phosphorus (P), fasting plasma glucose (FPG), uric acid (UA), and estimated glomerular filtration rate (eGFR). SHAP interpretation indicated that higher ALP, BF, FPG, UA, and eGFR were associated with lower predicted T-scores (median SHAP values: ALP − 0.24, BF − 0.19, FPG − 0.17, UA − 0.15, eGFR − 0.12), while P was associated with higher predicted T-scores (median SHAP + 0.11). The Elastic Net model achieved a test set R² of 0.623 with an RMSE of 1.245, demonstrating moderate-to-strong predictive accuracy. EN provides superior prediction of bone health in older men compared with conventional regression, whereas the evaluated tree-based ML models (SGB and XGBoost) showed comparable performance to regression. The identified metabolic and anthropometric markers may support early risk stratification and guide future mechanistic and longitudinal studies.