Abstract

Chronic Kidney Disease (CKD) is an enduring medical ailment capable of causing a failure of kidney and disastrous health issues unless promptly recognized at an early age. Proper and timely CKD prediction is thus critical towards enhancing clinical decision-making and patient outcomes. Despite the recent research on the prediction of CKD using machine learning methodologies, a significant number of methods fail to offer robust validation, do not offer adequate model interpretability, and often lack the analysis of computational efficiency. To overcome these issues, this research suggests an explicable ensemble learning framework to predict CKD based on a sample of 400 patient records and various clinical features. The systematic data preprocessing, chi-square based feature filtering, Recursive Feature Elimination with Cross-Validation (RFECV), stratified dataset splitting, and SMOTE-based training set balancing are all proposed frameworks. Four ensemble models are trained: Random Forest, Extra Trees, XGBoost, and LightGBM and they are optimized with the help of RandomizedSearchCV and measured with the help of several performance metrics. The results of the experiment indicate that the optimized Random Forest model performed the best with an accuracy of 0.9900, precision of 0.9865, recall of 0.9922, F1-score of 0.9892, ROC-AUC of 0.9989, and MCC of 0.9889. SHAP interpretation, Morris sensitivity analysis, and permutation importance were used to determine the presence of red blood cell status, urine specific gravity, hemoglobin level, and hypertension as the most significant clinical indicators of CKD prediction.