Abstract
Abstract This research formulates a clinically interpretable and computationally scalable artificial intelligence framework for the early detection and stage-wise classification of chronic kidney disease (CKD) utilizing routinely gathered clinical and laboratory data. The proposed framework aims to achieve two complementary goals: (i) the early detection of CKD framed as a binary classification problem, and (ii) the stratification of severity through ordinal classification based on the estimated glomerular filtration rate (eGFR). Calibrated ensemble learning models are trained on the UCI CKD dataset to enable early risk stratification. These models are trained to deal with different types of features and missing records. We use both multiclass gradient boosting models and an ordinal logistic regression approach based on the CORAL paradigm to stage diseases. This lets us model ordered CKD stages based on eGFR in a structured way. In this study, the early prediction task reaches test AUROC values close to 0.98, with sensitivity over 90% at high-specificity operating points. Additionally, the stage-wise classification achieves macro-AUROC values above 0.90, with misclassifications mainly limited to adjacent disease stages. Feature importance analysis corroborates that recognized biomarkers, including serum creatinine, hemoglobin, albumin, and urine specific gravity, predominantly influence model decisions. The framework offers a clear and useful foundation for assessing CKD risk and facilitates incorporation into clinical decision support systems.