Obesity is a global health crisis requiring accurate and interpretable prediction models for clinical use. Current ensemble approaches lack systematic comparisons and often use suboptimal configurations. This study introduces OBES-Net, a novel framework that integrates three methodological advances: (1) exhaustive hyperparameter optimization across 50 configurations of nine machine learning algorithms, (2) systematic comparison of Hybrid Majority Voting strategies (Majority Hard vs. Weighted Hard) against Ensemble Stacking, and (3) comprehensive Explainable AI (XAI) validation using SHAP and LIME. Utilizing two diverse datasets, OBES-Net demonstrates that Stacking achieves superior performance on complex data (Accuracy: 0.991 on Dataset-2), while voting strategies remain competitive on simpler tasks. Crucially, the proposed XAI analyses reveal that Stacking’s advantage stems from synthesizing broader clinical risk factors, withWeight and Age emerging as consistent primary drivers. Extensive ablation studies validate the robustness of the proposed design choices. The OBES-Net framework establishes that combining exhaustive optimization with systematic ensemble comparison and XAI delivers both state-of-the-art performance and clinical trustworthiness, providing an advancing methodology for healthcare decision-making.