An Explainable Hybrid CNN–Machine Learning Framework for Accurate Chronic Kidney Disease Classification: A Comparative Study

Abstract

Chronic kidney disease (CKD) is a major global health burden, emphasizing the importance of early detection and timely management. Conventional machine-learning methods often depend on manually engineered features, whereas end-to-end deep-learning models may have limited interpretability when applied to structured clinical data. This study proposes a hybrid CNN–machine learning framework that combines one-dimensional Convolutional Neural Network (1-D CNN) feature learning with Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), and Logistic Regression (LR). The framework is evaluated using 400 CKD patient records containing 24 clinical variables, which are transformed into 34 model-ready features through missing-value treatment, categorical encoding, numerical standardization, and feature engineering. A stratified 80:20 train–test split and five-fold stratified cross-validation are used. The CNN generates a 64-dimensional feature representation for subsequent classification. Among the evaluated models, CNN–XGBoost achieves the best independent test performance, with 98.75% accuracy, 100% precision, 98% recall, 100% specificity, 98.99% F1-score, 0.9993 ROC-AUC, and 0.9996 PR-AUC. The confusion matrix shows 49 true positives, 30 true negatives,  one false negative, and no false positives. SHAP analysis identifies serum creatinine, blood urea, hemoglobin, specific gravity, albumin, and packed cell volume as important predictors. Overall, combining CNN-based feature learning with XGBoost provides a potentially effective and interpretable approach to CKD classification and data-driven clinical decision-making. External validation using larger, multicentre, prospective datasets remains necessary to establish its clinical applicability.