Machine Learning for Childhood Obesity Risk Assessment: A Systematic Review of Design, Validation, and Reporting Practices, with a Feasibility Analysis for Peruvian Schoolchildren

Background: Machine-learning models for classifying or predicting overweight and obesity in children vary widely in design, validation, and reporting. Methods: We registered a protocol (OSF, doi:10.17605/OSF.IO/EP47Z) and searched PubMed, Scopus, and Web of Science on 15 October 2026 for original studies of machine-learning models for overweight or obesity in people under 19 years. One reviewer screened all records, and a second reviewer screened a 15% sample; one reviewer assessed risk of bias with PROBAST+AI. Results: Of 105 full texts sought, 36 (34%) could not be retrieved; 37 studies were included. None was judged at low risk of bias. External, temporal, or geographic validation was reported in 6 studies (16%) and calibration in 6 (16%); up to 4 (11%) used potentially outcome-defining inputs, and preprocessing preceded data partitioning in at least 11 (30%). In the 2021 Peruvian VIANEV survey, only 185 children, all in Lima Metropolitana, had measured weight and height, too few for multivariable model development; national implementation requires data with comprehensive anthropometric coverage. Conclusions: Reported accuracy and AUC values should be interpreted cautiously; studies should define the outcome with a named reference, avoid outcome-defining inputs, apply preprocessing inside the training data, and report calibration and external validation.