Generalizable and Explainable Machine Learning for Chronic Kidney Disease Screening

Abstract

Chronic Kidney Disease (CKD) represents a major global health challenge due to its progressive nature, high prevalence, and frequent underdiagnosis at early stages. Although machine learning (ML) techniques have shown promise for CKD screening using clinical data, most existing studies rely on single, homogeneous datasets and internal validation, limiting their real-world applicability. Moreover, limited attention has been given to model generalisation, dataset heterogeneity, and the stability of explainability mechanisms.

This study proposes a robust and interpretable machine learning framework for CKD screening that explicitly addresses cross-dataset generalisation and trustworthy model behaviour. Multiple classification models, including Logistic Regression, Random Forest, and Histogram-based Gradient Boosting, are evaluated using a leakage-aware experimental design combining stratified internal cross-validation with strict external cross-dataset testing. A harmonised preprocessing strategy is employed to align semantically equivalent clinical features across datasets with differing encodings, enabling realistic out-of-distribution evaluation.

Experimental results demonstrate that while predictive performance decreases under external validation as expected in heterogeneous clinical settings the proposed framework maintains clinically meaningful discrimination and sensitivity suitable for screening applications. Explainability analysis based on SHapley Additive exPlanations (SHAP) reveals consistent reliance on clinically relevant renal and hematological features across datasets. Furthermore, explanation stability analysis confirms that ensemble models exhibit more robust and consistent feature attribution behaviour under dataset shift.

Overall, this study highlights the necessity of cross-dataset evaluation, conservative performance reporting, and explainability stability analysis for developing reliable and deployable CKD screening systems. The proposed framework provides a practical pathway toward trustworthy clinical decision-support tools that prioritise robustness and transparency over unrealistic accuracy claims.