Development and validation of supervised machine learning models for the non-invasive screening of osteosarcopenia: Bushehr Elderly Health (BEH) program

Osteosarcopenia, the co-occurrence of osteoporosis and sarcopenia, severely impacts older adults but remains underdiagnosed due to complex diagnostics. We aimed to develop and internally validate supervised machine learning (ML) models using accessible clinical data from a large cohort study to create an exploratory framework for the concurrent identification of this high-risk syndrome. In this cross-sectional study, we developed and internally validated ML pipelines to identify musculoskeletal risk using a large Iranian cohort. Four outcomes—low bone mass, osteoporosis, sarcopenia, and osteosarcopenia—were defined via standard consensus criteria. To prevent data leakage, feature selection and resampling were embedded within cross-validation folds, isolating a maximum of ≤ 15 clinically accessible predictors. A diverse ensemble of algorithms was optimized and subjected to post-hoc probability calibration to ensure reliable, risk estimates. To maximize screening utility, classification thresholds were constrained to prioritize high sensitivity. Finally, for osteosarcopenia, we evaluated both direct classifiers and a combined meta-learner framework. The median age of the cohort (n = 4110) was 66.0 years [IQR 62.0–70.5], and 54.2% of participants were female. For the identification of osteoporosis, calibrated tree-based models significantly outperformed standard baselines; the Random Forest classifier achieved a Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.821, a sensitivity of 0.946, and a Negative Predictive Value (NPV) of 0.926. For sarcopenia, XGBoost yielded an ROC-AUC of 0.865, a sensitivity of 0.862, and an NPV of 0.959. For identifying osteosarcopenia, direct classifiers and combined meta-learner frameworks achieved similar performance (ROC-AUC ~ 0.85, NPV ~ 0.95). Crucially, calibration ensured highly reliable probability risk estimates across all optimal models (Brier scores 0.10–0.17). Interpretability analysis via SHAP identified age, weight/BMI, forearm circumference, and functional limitations as the primary drivers of musculoskeletal risk. We demonstrated that distinct ML models, using routine clinical parameters, can identify osteoporosis, sarcopenia, and osteosarcopenia with high sensitivity. Future research should focus on external validation of these frameworks across diverse populations to translate these into deployable tools.