An explainable machine learning framework for cross-dataset chronic kidney disease risk prediction using hybrid data balancing

Chronic kidney disease (CKD) is a major global health problem, and early detection is important for reducing complications and improving treatment. Machine learning has shown strong potential for CKD prediction, but many existing studies rely on single datasets, face severe class imbalance, and provide limited interpretability. These challenges reduce model reliability and make cross-dataset use in real clinical settings more difficult. Therefore, this study aimed to develop an interpretable and adaptable machine learning framework for CKD risk prediction across different datasets. The proposed hybrid explainable machine learning (HXAI-ML) framework used two public CKD datasets: the UCI CKD Risk Factor Prediction dataset as the source dataset and the Kaggle CKD dataset as the target dataset. After preprocessing, label binarization, imputation, and standard scaling, seven balancing methods and six classifiers were evaluated. A target adaptation setting was applied by using 10% of the target dataset for calibration and 90% for blind testing. SHAP, LIME, and Permutation Importance Analysis were used to explain the model predictions. Among all tested combinations, RO + TL + ETC achieved the highest overall accuracy among the evaluated target-domain adaptation configurations and was therefore selected as the primary model under the accuracy-based selection criterion. On the 1,494-sample blind holdout it obtained 91.97% accuracy, 92.24% precision, 99.64% recall, 95.80% F1-score, and 67.52% ROC-AUC. However, its specificity of 5.74% and balanced accuracy of 52.69% indicate substantial class-wise imbalance in predictive performance. The findings demonstrate the methodological feasibility of cross-dataset target-domain adaptation and explainable machine learning under limited target-domain data. However, the relatively low ROC-AUC values and low specificity for the non-CKD class indicate limited discriminative ability. Therefore, the proposed framework should be regarded as an exploratory cross-dataset evaluation rather than a clinically deployable screening or decision-support system.