Abstract
Obesity is a major public health concern. Predicting obesity risk from lifestyle data can guide targeted interventions, but current models remain limited. This study first evaluates ensemble learning methods and then combines approaches to improve prediction accuracy and generalizability. Four ensemble techniques-boosting, bagging, stacking, and voting-were tested. Five boosting and five bagging models were constructed alongside voting and stacking models. Hyperparameter tuning optimized performance, and feature importance analysis guided potential feature elemination. In phase two, hybrid stacking and voting models integrated the best-performing boosting and bagging models to enhance predictive capability. Model robustness was ensured through k-fold cross-validation and statistical validation. SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) improved interpretability by analyzing feature contributions. Hybrid stacking and voting models outperformed other ensemble methods, with stacking achieving the best performance (accuracy: 96.88%, precision: 97.01%, and recall: 96.88%). Feature importance analysis identified key predictors, including sex, weight, food habits, and alcohol consumption. The results demonstrated that hybrid ensembles significantly improved obesity risk prediction while preserving interpretability. Integrating multiple ensemble and explainability techniques provides a reliable framework for obesity prediction, supporting clinical decisions and personalized healthcare strategies to mitigate obesity risk.