Beyond Accuracy: A Structural Validation Framework for Machine-Learning Benchmarks in Chronic Kidney Disease Classification

This study makes four contributions:
• A CKD-specific structural validation framework integrating
predictive signal, clinical-variable structure, and
reproducibility;
• A joint structural evidence criterion that combines predictive
and dependence null testing into a single benchmark level
assessment;
• A reproducibility analysis incorporating repeated cross-validation,
complete-case and feature-ablation sensitivity,
label permutation, and 10,000 paper-compatible random
partitions; and
• An empirical comparison demonstrating that CKD
Datasets with similar variable dictionaries can exhibit
markedly different predictive and multivariable structures.
Cross-dataset source classification and frozen transfer are
retained as secondary sensitivity analyses.