Risk Analysis in the Lower Silesia Healthy Donors Cohort: Statistical Insights and Machine Learning Classification

Background/Objectives: Metabolic syndrome (MetS) increases the risk of type 2 diabetes and cardiovascular disease. We aimed to identify the key metabolic predictors of MetS in a Central European cohort and to compare classical statistics with modern machine learning (ML) models. Methods: We analysed 956 adults from the Lower Silesia Healthy Donors cohort. Clinical, anthropometric, biochemical, and lifestyle variables were collected using standardised procedures. Group differences were tested with Mann–Whitney U tests and effect sizes. A multivariable logistic regression (outcome: binary MetS defined as ≥3 harmonised components, MetS_bin) estimated adjusted odds ratios. In parallel, ML models (logistic regression, Random Forest, XGBoost, LightGBM, CatBoost) were trained with stratified 5-fold cross-validation. Performance was evaluated by accuracy, F1-macro, and area under the receiver-operating characteristic curve (ROC AUC). Model interpretability used SHAP values. Results: Overweight/obese participants had higher fasting glucose (median 92.0 vs. 84.6 mg/dL), fasting insulin (9.9 vs. 6.6 µU/mL), and systolic blood pressure (134 vs. 121 mmHg) and lower HDL cholesterol (53 vs. 66 mg/dL) compared to normal-BMI individuals (all p < 0.001, r ≈ 0.39–0.41). Participants with a higher waist circumference also showed markedly increased HOMA-IR (2.16 vs. 1.34; p < 0.001). In multivariable logistic regression, waist circumference, BMI, triglycerides, HDL cholesterol, fasting glucose, and systolic blood pressure were independently associated with MetS, yielding a test ROC-AUC of 0.98 and PR-AUC of 0.88. Machine learning models further improved discrimination: Random Forest, XGBoost, LightGBM, and CatBoost all achieved very high performance (test ROC-AUC ≥ 0.99, PR-AUC ≥ 0.98), with CatBoost showing the best cross-validated PR-AUC (~0.99) and favourable calibration. SHAP analyses consistently highlighted fasting glucose, triglycerides, HDL cholesterol, waist circumference, and systolic blood pressure as the most influential predictors. Conclusions: Combining classical regression with modern gradient-boosting models substantially improves the identification of individuals at risk of MetS. CatBoost, XGBoost, and LightGBM delivered near-perfect discrimination in this Central European cohort while remaining explainable with SHAP. This framework supports clinically meaningful risk stratification—including a “subclinical” probability zone—and may inform targeted prevention strategies rather than purely reactive treatment.