Leveraging Machine Learning and Genetic Risk Scores for the Prediction of Metabolic Syndrome in Children with Obesity ^†

Abstract

Background and objectives: Obesity is a growing global epidemic, associated with increased cardiometabolic disorders. Metabolic syndrome (MS) is defined by altered insulin, blood pressure, glucose, and lipid levels. Pubertal children with obesity are highly susceptible to developing MS, necessitating its early identification. This study aims to compute phenotype-specific genetic risk scores for MS-related biochemical markers and evaluate their clinical utility using machine learning-based models. Methods: Longitudinal data from the PUBMEP Spanish cohort were analyzed, including 138 children (71 girls and 67 boys) at two time points, spanning from prepuberty to puberty. Clinical, endogenous, environmental, and omics variables were measured. Genetic risk scores were generated using GWAS data and PRSice-2 software. These scores, alongside non-genetic prepubertal data (e.g., biochemical, anthropometric, and physical activity data), were integrated into predictive models using machine learning techniques to forecast the MS status during puberty. Linear models explored interactions between environmental factors, genetic risk scores, and disease risk. Results: Strong associations were observed between each genetic risk score and its corresponding phenotypic biomarker. Notably, certain scores related to obesity and high-density lipoprotein levels exhibited significant interactions with environmental factors, such as sedentary lifestyle, modulating disease effects. The predictive machine learning models incorporating prepubertal genetics, high-density lipoprotein, and sedentary lifestyle achieved reasonable performance in predicting pubertal obesity (AUC, accuracy, and sensitivity of 0.89). These models strike a favorable balance between risk scores derived from genetic factors and clinical variables. However, when individual risk scores were considered in isolation, limited predictive results were observed for MS and associated altered components. Discussion: This study demonstrates the importance of the early identification of at-risk children for MS. The integration of genetic risk scores, clinical variables, and machine learning techniques offers promising avenues for predicting pubertal MS. While individual risk scores have limitations in isolation, polygenic risk scores serve as valuable tools for investigating gene–environment interactions. Following our results, polygenic risk scores lacked sufficient predictive ability in most clinical traits, limiting their clinical application. Nevertheless, they remain valuable analytical tools for exploring the association with the environment, by consolidating the effects of multiple single nucleotide polymorphisms into a single variable.