Non-Invasive Prediction of Metabolic Syndrome Using Explainable Machine Learning

Background: Metabolic syndrome (MetS) is a complex health problem significantly associated with cardiovascular diseases and type 2 diabetes mellitus. Traditional diagnostic approaches rely on invasive biochemical markers, which limit their accessibility. Here, we developed an explainable machine learning (ML) framework for MetS prediction based on non-invasive demographic, anthropometric, and behavioral variables. Method: We conducted a retrospective observational study that included 1090 participants with MetS and 584 without MetS, using six supervised ML models, including Random Forest, AdaBoost, K-Nearest Neighbors, Bagging, Logistic Regression, and Ensemble learning. We trained and validated these models using stratified 10-fold cross-validation. Predictive performance was evaluated using accuracy, precision, recall, specificity, negative predictive value, area under the receiver operating characteristic curve (AUC) and F1-score. Shapley Additive Explanations (SHAP) applied to the Random Forest classifier evaluated interpretability. A 1:1 balanced sensitivity analysis was subsequently performed using the same leakage-safe predictor set and validation framework to assess whether model performance was materially influenced by outcome prevalence. Results: In the natural-prevalence primary cohort (n = 1674; 65.1% with MetS), leakage-safe models showed moderate discrimination. Ensemble achieved the highest mean AUC (0.751, 95% CI 0.731–0.771), followed by Random Forest (0.745, 95% CI 0.723–0.767) and Bagging (0.742, 95% CI 0.725–0.760). SHAP analysis identified (body mass index) BMI, age, and body weight as the dominant contributors, with additional contributions from physical activity, and dietary and sociodemographic variables. In the 1:1 balanced sensitivity cohort (n = 1000), discrimination remained similar, with Ensemble achieving the highest mean AUC (0.742, 95% CI 0.704–0.780). Conclusion: Explainable ML models based on accessible non-laboratory variables provided moderate discrimination of prevalent MetS and may serve as candidate pre-laboratory risk-stratification tools. They should not replace established diagnostic criteria, and external validation, recalibration, and prospective evaluation of clinically relevant thresholds are required before implementation.