<![CDATA[Obesity is a global health problem that requires accurate, stable, and transparent multi-class prediction methods to support early clinical intervention. Previous studies used a Hybrid Stacking architecture with a linear Meta-Learner, which achieved 96.88% accuracy but had limitations in capturing complex non-linear interactions between basic model predictions. The main problem lies in the limitations of the linear Meta-Learner (Logistic Regression), which is not optimal in integrating non-linear signals from tree-based models at Level-1. The purpose of this study is to improve the performance, stability, and transparency of multi-class obesity predictions through the development of a Hybrid Stacking architecture with a non-linear Meta-Learner and the implementation of model interpretability techniques. To address this gap, this study proposes a new Hybrid Stacking Ensemble model by replacing the linear Meta-Learner with a powerful boosting model, namely CatBoost. The proposed model was evaluated on a multi-class obesity dataset and successfully surpassed state-of-the-art (SOTA) performance. The main performance improvement is demonstrated by an increase in Accuracy to 97.83% (an absolute increase of +0.95%) and a significant improvement in multi-class stability metrics: MCC (reaching 97.30%) and Cohen's Kappa (reaching 97.39%). This superiority validates the hypothesis that non-linear Meta-Learners are more effective. Furthermore, we included the technical innovation of Manual Padding on Level-1 outputs to ensure feature consistency, enabling a valid SHAP Level-2 analysis. The SHAP analysis revealed a strategic synergy, where CatBoost relied on Logistic Regression (a linear model) to predict high-risk class probabilities (Obesity Type II & III), while utilizing tree-based models for other classes. This model provides a superior, stable, and transparent methodology for obesity level prediction.]]>
