A Clinical Engineering-Oriented Explainable Artificial Intelligence Framework for Chronic Kidney Disease Risk Estimation and Multi-Parameter Staging

Abstract

Background/objectives: Chronic kidney disease (CKD) is a progressive and often asymptomatic disorder that requires early detection supported by interpretable clinical decision systems. Conventional machine-learning (ML) models provide strong predictive accuracy but frequently lack physiological transparency and clinical interpretability. This study aims to develop a composite explainable risk inference framework that integrates consensus-based biomarker discovery, interpretable ML, and rule-based clinical reasoning to enhance diagnostic transparency and clinical reliability in CKD risk estimation and staging. Methods: A structured CKD dataset containing 1,659 patient records and 54 recorded variables, including the outcome variable, was analyzed. The dataset was audited for completeness prior to analysis. No missing values were identified; therefore, neither numerical nor categorical imputation was required. Categorical variables were encoded, and numerical variables were standardized within the modeling pipeline. A consensus biomarker influence index was introduced by integrating mutual information, least absolute shrinkage and selection operator (LASSO)-based feature selection, permutation importance, and explainable boosting machine (EBM) contribution scores. Seven consensus biomarkers were retained and used to train an EBM for probabilistic CKD risk estimation. Clinical thresholds were encoded into binary risk indicators to compute a Clinical Risk Score (CRS). Model performance was evaluated using accuracy, precision, recall, F1-score/F-measure, receiver operating characteristic area under curve (ROC-AUC), and five-fold cross-validation. The analytical framework was additionally evaluated on an independent dataset containing 168 records. Results: The consensus approach has identified serum creatinine and glomerular filtration rate as key biomarkers, which have the highest influence indices. The proposed EBM model demonstrated promising results in terms of accuracy: 0.9277, precision: 0.9297, recall: 0.9967, F1-score: 0.9620, and ROC-AUC: 0.7650, compared to other baseline models, such as logistic regression, random forest, gradient boosting, etc. A monotonic relationship was established between CRS and ML risk probability, which showed a high degree of association between the proposed model and the actual physiological load. The proposed multi-parameter model was able to stratify patients at different clinically relevant stages of CKD. In an independent cohort, EBM achieved an ROC-AUC of 0.7222, with a bootstrap 95% confidence interval (CI) of 0.535–0.889. Conclusions: The proposed model of explainability is effective in integrating the predictions of ML models with the power of clinical reasoning, thereby enhancing the risk estimation of CKD. This model is reliable for the effective management of CKD.