This retrospective, multicenter systematic review asked how artificial intelligence can be applied across the life course to improve obesity prevention, risk stratification, and management, using a broad range of machine-learning, deep-learning, graph-based, causal-inference, and large-language-model approaches to synthesize heterogeneous biomedical, behavioral, and environmental data. Researchers analyzed 1,470 peer-reviewed human studies identified from PubMed, Web of Science, and CNKI published between August 31, 2020, and August 31, 2025, following PRISMA 2020 guidelines. Studies incorporated multi-omics data, medical imaging, EHRs, wearable-sensor streams, environmental and geospatial data, and unstructured clinical or lifestyle text, with preprocessing including feature scaling, class balancing, multimodal integration, and explainability analyses. The AI methods reviewed ranged from classical models (logistic regression, random forests, SVMs, gradient boosting) to deep neural networks (CNNs, LSTMs, transformers), graph neural networks, mechanistic digital twins, causal models (SEM, DAGs), federated learning frameworks, and LLMs. They were typically compared against standard statistical methods, manual clinical assessment, or existing clinical workflows rather than head-to-head clinician performance. Best-performing models in individual studies frequently achieved high discrimination (often >95% accuracy or AUROC ≈0.95–1.00 for structured clinical tasks), with more modest performance for complex gene–environment prediction (e.g., AUROC ≈0.70–0.75). Secondary analyses across studies included explainability, subgroup testing, causal modeling, geospatial mapping, and fairness considerations, with recurring emphasis on privacy-preserving approaches such as federated learning and differential privacy. Key limitations included selection bias, heterogeneous data sources, reliance on cross-sectional or region-specific cohorts, synthetic oversampling, limited external validation, and inconsistent demographic or fairness analyses. External validation was variably performed, and subgroup fairness assessments were often absent, underscoring that findings primarily reflect model performance and pattern discovery rather than direct improvements in patient outcomes or proven clinical efficacy.