<p>Chronic kidney disease imposes a long-term decline in renal function that often goes unnoticed until advanced stages. In primary care, routine laboratory markers are available at scale but are underused for proactive risk triage. This study leverages machine-learning pipelines to transform routinely collected variables into timely predictions for clinical screening. In this study, we propose a CKD diagnostic system that integrates degree-2 polynomial feature expansion and a KMeans-guided, feature-aware stratified train–test split—validated by the Kolmogorov–Smirnov (KS) test)—to improve robustness and generalization of downstream classifiers. We employ a cluster-stratified split: samples are first grouped in feature space via K-means, then training and test folds are drawn proportionally from each cluster. This preserves covariate balance across splits—a detail seldom made explicit in prior CKD studies. To expose non-linear structure, we augment the design matrix with a polynomial basis, enabling even simple learners to model curvature and interactions. In combination, these steps improved predictive stability and discrimination. On the public UCI CKD dataset, we benchmarked six classifiers (RF, SVM, Naïve Bayes, LR, k-NN, XGBoost). Under our protocol, random forest (RF), XGBoost, SVM, and logistic regression (LR) achieved accuracy = 1.00, exceeding results typically reported in recent work. Beyond CKD screening, the workflow is lightweight, scalable, and transferable to routine clinical environments and other medical tabular problems. This work should be regarded as a proof-of-concept on a public benchmark. Any inference about clinical usefulness requires prospective, independent external validation on statistically virgin cohorts from multiple centers and countries, with a locked model and prespecified analysis, following best-practice guidance.</p>