Explainable Machine Learning for Chronic Kidney Disease Classification from Routine Urinalysis and Diabetes Records

Chronic kidney disease (CKD) often advances without symptoms, and the measurements most likely to reveal it are the ones already collected in routine care. Machine learning has repeatedly reported near-perfect CKD classification on small public benchmarks, which raises a sharper question: which measurements carry the signal, and does the result survive an evaluation that never splits a patient's records, or indistinguishable records, between training and test data? We address this question in a urinalysis cohort of 380 records and a longitudinal cohort of 4,000 annual records from 400 patients with diabetes, using grouped nested cross-validation in which preprocessing, model selection, calibration, and thresholds are fitted on training data only. Urine measurements classified recorded CKD status with ROC AUC 0.964 (95% cluster bootstrap CI 0.942 to 0.982), logistic regression was selected in every fold, and the 18 urine fields alone reached 0.973 while demographics alone reached 0.642. Demographic, treatment, and lifestyle records reached only 0.687, and the full feature set did not outperform age and diabetes duration. Calibration reduced probability error, masking 30% of urine measurements lowered mean AUC to 0.920, and same-record attribution rankings agreed across partitions with median correlation 0.886. A next-record onset analysis found 127 of 185 first positive labels at the final observation, so apparent early-detection performance was largely timing. Routine measurements carry real information about recorded CKD status; the analysis shows where it lives and where clinical claims must stop.