XAI-Enhanced Hybrid Models for Effective Chronic Kidney Disease Prediction

Chronic Kidney Disease (CKD) is a major global health concern that requires early and accurate detection to prevent progression and improve patient outcomes. This study presents a predictive framework integrating traditional machine learning and deep learning with an ensemble approach for robust CKD diagnosis. To address data imbalancing, three resampling techniques, namely SMOTE, SMOTE-Tomek, and SMOTE-ENN, were proposed to balance the dataset. A single SHAP-based feature selection via XGBoost was applied to each resampled dataset, and the top features were aggregated across all three techniques. We also performed a comprehensive statistical analysis using SPSS software to identify the most important features. Nine machine learning classifiers and two ensemble strategies were trained and applied alongside several deep learning models, including Attention Autoencoder with XGBoost for latent features, TabNet, TabPFN, LightCNN, MLP, and Deep and Cross Network (DeepCrossNet). DeepCrossNet achieved the highest accuracy (97.38 %) among all deep learning models; meanwhile, the stacking ensembles model reached 97.50 %, and Random Forest led among traditional models with the accuracy of 97.71 %. Explainable AI, likely SHAP and LIME analysis, confirmed that the features GFR (glomerular filtration rate) and serum creatinine are the most influential predictors, supporting clinical interpretability. t-SNE and UMAP projections demonstrated clear class separation and highlighted ambiguous cases. These results demonstrate that combining aggregated, SHAP-selected features from resampled datasets with ensemble and deep learning methods yields accurate and interpretable CKD predictions.