Ensemble Machine Learning and SMOTE-NC forthe Multi-Stage Classification of Chronic KidneyDisease Using Routine Clinical Data

Chronic Kidney Disease (CKD) is a condition that becomes evident only after its symptoms have developed to advanced stages because doctors need special tests for serum creatinine and glomerular filtration rate to confirm its presence, which rural clinics and low-resource health facilities cannot provide. The moment patients acquire these diagnostic tests, their chances of receiving early treatment have already gone low. This research investigates whether routine check-up measurements can provide insights into assessing CKD severity. We evaluated six supervised learning models --- Random Forest, Logistic Regression, SVM, XGBoost, AdaBoost, and Gradient Boosting --- trained on a dataset of 400 patients using 11 standard clinical parameters including blood pressure, blood glucose, haemoglobin, and urine dipstick results, while omitting creatinine and eGFR. The study found that the classification problem definition requires more importance compared to the algorithm selection process. Grouping CKD into three broad severity classes (Early, Moderate, Advanced) instead of five detailed stages produced better results, increasing macro F1 scores by 0.19 to 0.28 across all tested models. XGBoost under the three-class formulation achieved the best test results with 80.52\% accuracy and a macro F1 score of 0.7655. An Optuna-based hyperparameter search reached an F1 score of 0.9400 during cross-validation, indicating that the model has enhanced performance capacity with additional data. This research shows that a well-structured screening process using only routine clinical data can function as an effective primary screening instrument, identifying patients who require additional assessment before their condition demands specialist clinical judgment.