"GENEDIS 2026 – Paper 41 – Machine-Learning Approaches for Predicting Childhood Obesity: A Narrative Review of Data Modalities and Algorithms"

Background: Childhood obesity affects over 159 million children and adolescents worldwide and is a major driver of lifelong cardiometabolic disease. Advances in machine learning (ML) and the growing availability of electronic health records, longitudinal cohorts, and genomic data have produced numerous paediatric obesity prediction models, but the literature remains heterogeneous in data, algorithms, features, and methodological quality. This narrative review synthesises emerging trends and remaining challenges. Methods: Studies published 2019–2026 were included if they developed or evaluated any supervised, unsupervised, deep-learning, ensemble, or polygenic-prediction algorithm for overweight, obesity, or excess adiposity in ages 0–18 and reported at least one quantitative performance metric. Findings were organised thematically by dataset type and further analysed by algorithm family, feature domain, developmental window, and validation strategy. Results: Reported best-model AUROC ranged from 0.60 (a genome-wide association study) to 0.986 (ROC-AUC: 0.986 in the internal testing set., Guangzhou, China, Center for Disease Control and Prevention), with several studies exceeding 87–88% accuracy. Tree-based gradient boosting (XGBoost, LightGBM, Random Forest) was the best-performing family in most studies; deep learning dominated only when longitudinal trajectories were the principal input, while unsupervised clustering and polygenic risk scores defined an emerging genomic and multi-omics strand. Electronic health records and national cohorts were the most common data sources, and predictive features spanned anthropometric, parental, behavioural, socioeconomic, genetic, dietary, and gut microbiome domains. External validation, calibration, and fairness analyses were rarely reported. Conclusions: ML models show considerable potential for early prediction of childhood obesity, but current evidence is limited by insufficient external validation, inconsistent calibration reporting, minimal fairness assessment, and under-representation of low- and middle-income populations. Externally validated, equitable, and gener-alisable models are needed before routine paediatric use can be recommended.