Machine Learning Prediction of Obesity Development in Children With Overweight Using Longitudinal Body Composition Data.

BACKGROUND

Children with overweight (body mass index [BMI] between the 85th and 95th percentiles) represent a critical target for obesity prevention, with obesity progression rates 10- to 20-fold higher than normal-weight peers.

OBJECTIVES

This study developed machine learning models to predict obesity in this high-risk group using anthropometric and bioelectrical impedance analysis parameters.

METHODS

We analysed longitudinal data from 2801 overweight Korean children aged 7-9 years. XGBoost models integrated anthropometric measurements, bioelectrical impedance-derived body composition, standardised deviation scores and growth velocity parameters to predict obesity (BMI ≥ 95th percentile). Sex-stratified models were evaluated using area under the receiver operating characteristic curve (AUROC) with bootstrap validation. Shapley Additive exPlanations (SHAP) identified key predictive features.

RESULTS

Obesity developed in 32.3% of males and 25.4% of females during follow-up. Models achieved AUROC scores of 0.671 (95% CI: 0.619-0.721) for males and 0.652 (95% CI: 0.589-0.700) for females. Key predictors included standardised weight and adiposity measures, height-adjusted skeletal muscle mass and growth velocity parameters.

CONCLUSIONS

Machine learning models demonstrated effective predictive performance for obesity in children with overweight. Incorporating standardised adiposity measures and growth velocity beyond static anthropometric data provides a robust framework for risk stratification in this high-risk population.