Frontiers in Public Health · Nov 18, 2026
Development and Validation of a Machine Learning-Based Model for Identifying Metabolic Syndrome in Middle-Aged and Older Adults
By Liyuan Zhu, Jinjiang Xu
Background: Metabolic syndrome (MetS) is common in middle-aged and older adults, yet practical tools for identifying individuals with MetS in routine health assessment remain limited. This study aimed to identify factors associated with prevalent MetS and to develop and validate a machine learning-based classification model for identifying MetS in middle-aged and older adults. SHapley Additive exPlanations (SHAP) analysis was used to improve model interpretability. Methods: This cross-sectional study used the China Health and Retirement Longitudinal Study cohort of adults aged ≥45 years (n=9,867), randomly divided into training (70%) and testing (30%) sets. External validation was performed in 553 participants from the Health Management Center of the First Affiliated Hospital of Jinzhou Medical University. After excluding MetS diagnostic indicators, seven predictors (age, body mass index [BMI], pulse, blood urea nitrogen, total cholesterol, uric acid, and hemoglobin A1c [HbA1c]) were selected. Six machine learning algorithms were developed and tuned using 10-fold cross-validation, with CatBoost retained as the primary model. Performance was assessed using discrimination, calibration, and decision curve analysis. SHAP was used for model interpretation, and exploratory probability stratification and an online calculator were developed. Results: Among 9,867 CHARLS participants, 2,197 had MetS; 213 of 553 participants in the external validation cohort had MetS. CatBoost achieved an AUC of 0.829 (95% CI, 0.813–0.846) in the testing set and 0.842 (95% CI, 0.809–0.875) in external validation. SHAP identified BMI as the most influential feature and showed nonlinear model contributions for BMI, HbA1c, uric acid, and pulse. Observed MetS prevalence increased across the exploratory low-, intermediate-, and high-probability groups (5.9%, 23.7%, and 53.5%, respectively). Conclusion: The CatBoost model based on seven routine indicators showed good ability to identify individuals with prevalent MetS. Combined with model interpretation, exploratory probability stratification, and an online tool, the model may provide digital support for MetS screening and health management in primary care settings.