Explainable AI in Healthcare: A Stacking-Based Approach for Diabetes Classification

Diabetes classification remains a challenge in healthcare, demanding accurate and interpretable machine learning models for timely diagnosis. This study introduces a novel stacking ensemble framework combining Random Forest, Gradient Boosting, XGBoost, and Support Vector Machines to enhance prediction performance. To address class imbalance, the hybrid SMOTEENN technique ensures balanced representation. Feature engineering extracts meaningful patterns from the data, including polynomial interactions, feature ratios, and PCA. SHAP and LIME achieve explainability, providing insights into feature importance and prediction behavior. Experimental results show the framework achieves 97.1 % accuracy and 0.998 ROC-AUC, outperforming standalone models. Its transparency, supported by visual aids, makes it highly practical for clinical use. Future work will explore larger datasets and other medical prediction tasks.