Diabetes Prediction Using Stacking Ensemble and MLP: A Comparative Study with Hybrid Deep Learning Models

Abstract

Diabetes mellitus is one of the most prevalent chronic diseases globally, necessitating accurate predictive tools to support early clinical intervention. This paper presents a comprehensive evaluation of two advanced machine learning architectures—a Stacking Ensemble Classifier (CatBoost + LightGBM with Logistic Regression meta-learner) and a seven-layer Multilayer Perceptron (MLP) neural network—applied to the Kaggle Diabetes Prediction Dataset of approximately 100,000 patient records exhibiting a severe class imbalance of 91.5% non-diabetic vs 8.5% diabetic cases. The proposed Stacking Classifier achieves an accuracy of 97.59%, precision of 97.61%, recall of 97.59%, F1-Score of 97.59%, and AUC-ROC of 0.9974. The MLP achieves 95.48% accuracy and AUC of 0.9933. Both models are benchmarked against three hybrid deep learning models—RF+NN (96.81%), XGBoost+NN (96.75%), and Autoencoder+RF (96.58%)—demonstrating substantial superiority particularly in recall and F1-Score balance. SMOTE-based oversampling, 5-fold stratified cross-validation, and LIME explainability are integrated throughout. Confusion matrices, ROC curves, precision-recall curves, multi-metric comparisons, and LIME feature attribution plots derived from actual experimental results are presented in detail.