Predicting abdominal obesity in children and adolescents via machine learning: a longitudinal cohort study

OBJECTIVES: This study aimed to construct and verify machine learning (ML) models to predict long-term abdominal obesity (AO) risk in children and adolescents. METHODS: We trained and externally validated ML models to predict pediatric AO 4-5 years later using a publicly available longitudinal cohort. The body roundness index (BRI) was used as the diagnostic criterion for AO. The training/internal-validation cohort comprised 635 youths (2011→2015); an independent 2006→2011 cohort (n=456) provided external validation. After screening 25 routine variables and multiple imputation, four ML algorithms were evaluated by area under the receiver operating characteristic curve (AUC), sensitivity, specificity and F1-score across all three datasets. RESULTS: Among 635 children and adolescents in the raw-training set followed for four years, 164 (25.83 %) developed AO. Univariable analysis identified 12 significant baseline predictors (p<0.1); multivariable logistic regression (LR) retained five independent factors: body mass index (BMI), height, carbohydrate intake, reading/writing activity and urban residence. Four machine-learning algorithms were trained and validated; LR demonstrated the most stable performance across training (AUC=0.705), internal validation (AUC=0.742) and external validation (AUC=0.667) datasets. CONCLUSIONS: The LR model stands out as a potential tool in predicting long-term AO risk in children and adolescents.