Applications of traditional machine learning and deep learning algorithms in obesity prediction or classification: a systematic review of comparative performance

This systematic review evaluated traditional machine learning (TML) and deep learning (DL) approaches for obesity prediction in longitudinal studies and obesity classification in cross-sectional studies. PubMed, Scopus, and Web of Science were searched from inception to June 15, 2025. Studies applying these algorithms and reporting performance metrics were included. Seventeen studies (three longitudinal and fourteen cross-sectional) met the inclusion criteria. Only one of the 17 included studies performed independent external validation, and calibration reporting was extremely limited, highlighting substantial methodological barriers to clinical implementation. Longitudinal studies showed modest performance, with only one reporting discrimination (AUC = 0.81), limiting comparison between TML and DL for future risk prediction. In contrast, cross-sectional studies reported high apparent performance, with DL models reaching AUCs up to 0.979 and well-tuned TML models also performing strongly; however, many estimates were influenced by methodological limitations, including circular predictor-outcome relationships and incomplete performance reporting. Both approaches performed well for obesity classification, but evidence for clinically meaningful future obesity risk prediction remains limited. Longitudinal designs and independent external validation are needed to improve clinical relevance. Until such evidence is available, AI-based obesity models should be viewed as exploratory decision-support tools rather than routinely implemented instruments for clinical or public health risk assessment.