Metabolic syndrome (MetS) is a cluster of interrelated metabolic abnormalities—central obesity, dyslipidemia, hypertension, and hyperglycemia—associated with increased risk of cardiovascular disease and type 2 diabetes. The prevalence of MetS has risen in Thailand and wider Asia, prompting interest in earlier identification of metabolic dysfunction. An intermediate state, termed pre-metabolic syndrome (pre-MetS), denotes presence of some metabolic abnormalities without meeting full diagnostic criteria and may represent an opportunity for early intervention.
Conventional diagnosis uses fixed threshold-based definitions such as the International Diabetes Federation (IDF) criteria, which dichotomize a continuous biological process. Machine learning (ML) can integrate multidimensional clinical data and model nonlinear relationships among predictors, potentially offering complementary information for distinguishing metabolic states. This study tested ten supervised ML algorithms to classify pre-MetS versus MetS in a Thai adult cohort and examined which features most strongly influenced model-based classification.
This single-center retrospective cross-sectional study enrolled 657 Thai adults. Participants were classified as pre-MetS (n = 359) or MetS (n = 298). For research purposes, pre-MetS was operationally defined as presence of at least one IDF-defined metabolic abnormality without fulfilling the complete IDF criteria for MetS; central obesity was not required for pre-MetS classification in this operational definition.
Twenty-four clinical and biochemical features were selected using the Boruta feature-selection algorithm. The selected features included routine measures and several derived indices (for example, triglyceride-glucose derived indices). All preprocessing steps—including z-score normalization and feature selection—were applied exclusively within the training dataset to prevent data leakage into model evaluation.
Ten supervised ML algorithms were evaluated. The dataset was split into an 80% training set and a 20% independent hold-out test set (n = 132). Models were trained on the training set with repeated 10-fold cross-validation (10 repetitions) used for hyperparameter tuning. Final model discrimination and calibration were assessed on the independent hold-out test dataset. Performance metrics reported included area under the receiver operating characteristic curve (AUC), accuracy, and Brier score, with 95% confidence intervals.
Performance varied substantially across algorithms. Ensemble methods achieved the highest classification performance. XGBoost (extreme gradient boosting) demonstrated the strongest discrimination with an AUC of 0.986 (95% CI: 0.970–0.997), the highest accuracy of 0.932 (95% CI: 0.885–0.969), and the lowest Brier score at 0.050 (95% CI: 0.026–0.080). A neural network and random forest also yielded high discrimination (AUCs 0.948 and 0.942, respectively). Support vector machine models showed the lowest discrimination among the evaluated algorithms.
Explainability analysis using SHAP (SHapley Additive exPlanations) for the XGBoost model identified the features that contributed most strongly to classification. The top contributors were systolic blood pressure, waist circumference, fasting plasma glucose, sex, and TyG-WHtR (triglyceride-glucose waist-to-height ratio). These variables—some of which directly overlap with MetS diagnostic components or are derived from them—had the greatest impact on model output as measured by SHAP values.
Because several influential predictors overlapped with the components used to define MetS or were derived indices related to those components, sensitivity analyses were performed excluding such predictors. Discrimination decreased progressively after removal of overlapping predictors; however, substantial discrimination was retained after excluding these variables. Exact performance metrics for each sensitivity condition were reported in the source but are not reproduced here in full.
The authors emphasize several caveats. First, several predictors that strongly influenced classification overlapped with or were derived from the IDF MetS diagnostic criteria; this overlap may inflate apparent discrimination when classifying cases defined by those same criteria. Second, the study design was single-center and retrospective and lacked external validation in independent cohorts, limiting immediate generalizability. Third, although analysis scripts and reproducibility materials were made publicly available via a Figshare repository, the de-identified participant dataset is accessible only through a controlled-access mechanism due to ethical restrictions.
Given these limitations, the authors describe findings as exploratory and advise cautious interpretation. They recommend validation in multicenter cohorts and suggest using predictor sets that minimize overlap with diagnostic definitions before considering clinical implementation.
Among ten supervised ML algorithms evaluated on routine clinical and biochemical variables in Thai adults, ensemble methods—particularly XGBoost—achieved the highest performance for classifying pre-MetS versus MetS. SHAP analysis highlighted systolic blood pressure, waist circumference, fasting plasma glucose, sex, and TyG-WHtR as the most influential features. However, because influential predictors overlapped with diagnostic components and external validation was not performed, these results are exploratory. The authors call for multicenter validation using predictor sets that reduce overlap with MetS definitions before clinical application is considered.