AI-Driven Early Detection of Chronic Kidney Disease: A Comparative Benchmark of Ensemble Learning Models with Clinical Explainability

Abstract

Chronic kidney disease (CKD) is an emerging global health burden, and a significant volume of recent machine-learning research reports automated detection with near-perfect accuracy; in fact, many studies claim 96 100% accuracy on benchmark datasets. These results deserve further examination, however, because CKD is itself classified according to the estimated glomerular filtration rate (GFR) when GFR or a GFR-derived variable is a feature of the model input, a classifier can imitate the labeling rule rather than learn an actual diagnostic signal. We perform a controlled comparative benchmark of thirteen classifiers (five conventional baselines and eight ensemble learners) on a staged CKD dataset comprising 4,000 records, under two prediction framings (six-class stage classification and balanced binary early-vs-progressed formulation) in a rigorous, leakage-controlled protocol using stratified ten-fold cross-validation.

To explicitly measure the impact of label leakage, we performed an ablation analysis where GFR and cluster variables are alternately removed. In the presence of these variables, ensemble models achieve 97.8% accuracy and a Matthews correlation coefficient (=1) > 0.97; upon their removal, every single model (including gradient-boosted ensembles) collapses to chance-level performance (multiclass accuracy 21.9%, below the baseline of majority class at 25.1%; binary Matthews correlation coefficient ≈0 with all models collapsing to defaulting on majority class).

Indeed, the absence of a learnable signal is confirmed by SHAP and LIME analyses, where near-zero feature attributions are observed. The pairwise Wilcoxon signed-rank tests do not indicate any statistically significant differences between the leading models. We show that the state-of-the-art headline accuracies reported for CKD detection are in fact, on this dataset, an artefact of label leakage, and that the clinical and lifestyle features which remain carry no reliable predictive information about CKD stage. We suggest that auditing leakage, establishing proper baselines and validation on real-world clinical cohorts as preconditions for trustworthy CKD prediction models and release our entire protocol to encourage reproducible, plain sight benchmarking.