Abstract
Machine-learning methods are increasingly applied to childhood obesity risk assessment, yet feature selection, validation strategy, calibration, and benchmarking against simpler models remain inconsistently reported. This critical review examines thirteen primary model-development studies (predominantly pediatric), distinguishing genuinely prospective prediction from cross-sectional classification: only four used baseline predictors to forecast a later outcome in the same children. Two systematic reviews corroborate these concerns; this review extends that evidence by integrating target-leakage analysis with digital-health usability evidence into a pre-specified, falsifiable implementation framework. Identified concerns include target leakage from outcome-constitutive anthropometric predictors, limited independent or temporal validation, sparse calibration reporting, and inconsistent superiority of machine learning over logistic regression. We also review digital-health interventions and usability assessment via the System Usability Scale. Building on these findings, we propose a two-phase translational framework for Peruvian schoolchildren, specifying VIANEV as the primary individual-level dataset with area-level contextual enrichment from ENDES and INS/MINSA. We would evaluate candidate models (logistic regression, decision trees, random forest, XGBoost) through nested cross-validation with training-fold-only preprocessing, a sealed external validation set treated as a deployment gate, calibration and subgroup-fairness assessment, and pre-specified stopping criteria. The framework guides future development of transparent, context-sensitive, school-based decision-support systems while explicitly distinguishing cross-sectional risk classification from prospective prediction.