Answers are grounded in this paper and cite their sources.
Background Chronic kidney disease (CKD) is progressive, and early diagnosis improves outcomes. Objective We assessed how reliably routine clinical and laboratory variables identify CKD, and whether differences between common classifiers exceed the sampling error of a 400-patient dataset. Methods We analysed the UCI CKD dataset (250 CKD and 150 non-CKD patients, 24 predictors) with ten classifiers: logistic regression (LR), K-nearest neighbours (KNN), decision tree (DT), random forest (RF), AdaBoost, support vector machine (SVM), gradient boosting (GB), extra trees (ETs), XGBoost and LightGBM. A majority-class baseline served as reference. Imputation (KNN, k = 5), scaling and hyperparameter search were fitted inside the training folds of a nested, repeated stratified 5-fold cross-validation (CV) (three repeats). We report mean ± SD with Nadeau–Bengio 95% confidence intervals and compared models with corrected paired t -tests and Holm adjustment. Results Nine classifiers reached a cross-validated accuracy of 98.25–99.50% (AUC ≥ 0.990); KNN reached 93.33% and the baseline 62.50%. LightGBM (99.50 ± 0.63%), RF (99.42 ± 0.80%) and XGBoost (99.25 ± 0.79%) had the highest mean accuracy, but no pair among the nine non-KNN classifiers differed significantly. Permutation importance and SHAP highlighted specific gravity, haemoglobin, serum creatinine, albumin and packed cell volume. Conclusion No classifier could be ranked above another. The data come from one public dataset with almost completely separable classes, so external validation is needed.
No takes yet. Share an insight, caveat, or question.
Bader et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: