ESPE Yearbook of Paediatric Endocrinology

arXiv: 2504.06987 (2025) (preprint) doi: 10.48550/arXiv.2504.06987

Brief Summary: Using machine learning models, this study addresses the challenges of accurately predicting the Metabolic syndrome (MetS). Probability analysis revealed that elevated blood glucose had the highest likelihood (85.5%), while triglycerides showed the strongest predictive power (75%).

Comment: As the global prevalence of MetS has increased exponentially over the past decade, accurate prediction remains a challenge due to data scarcity and methodological consistencies, which can affect model reliability and clinical applicability. To address these challenges, this study used advanced machine learning models to overcome these limitations.

Following data preprocessing, which included demographic, clinical, and laboratory measurements from 2,402 individuals, several approaches were applied: model selection and evaluation, oversampling strategies, class imbalance handling, and counterfactual analysis. The counterfactual analysis identified the minimal changes required to shift individuals from MetS-positive and MetS-negative classifications. Accordingly, the most frequently modified features were blood glucose (50.3%) and triglycerides (46.7%), followed by waist circumference and HDL cholesterol. Triglycerides (74.9%) and blood glucose (58.7%) also emerged as strong predictors of MetS.

The authors conclude that combining data balancing methods with counterfactual analysis can enhance the prediction of MetS. This approach has the potential to be applied in other settings, where it may perform as well as or even better than traditional methods.