Scientific Reports · Nov 19, 2026

HMCKD-Net: hybrid machine learning for chronic kidney disease screening and probability calibration assessment

By Suman, Singh Y, Gulati N et al.

Chronic kidney disease (CKD) can remain undetected until renal damage progresses. HMCKD-Net combines training-only preprocessing, six principal components, four PCA-compressed autoencoder components, and logistic stacking of XGBoost, random forest and LightGBM for CKD classification. The development dataset contains 400 records (320 training; 80 hold-out). On the hold-out, HMCKD-Net achieved AUROC 0.9973 (95% confidence interval 0.9880–1.0000), accuracy 0.9875 and Brier score 0.0146. Hybrid-feature LightGBM achieved the same accuracy and AUROC 0.9967; the additional stacking stage did not establish a meaningful discrimination advantage. The frozen model was separately assessed against the time-to-event-derived EventCKD35 endpoint in 491 Abu Dhabi records. Its reported AUROC was 0.6786 and Brier score 0.3133. Seven of ten selected predictors were unavailable as measured external inputs. An independent assessment of baseline eGFR on the same public external records gave AUROC 0.8370 and average precision 0.4159, exceeding the reported transported discrimination. The external test involves predictor and endpoint shift and is not conventional external validation of the CKD classification model. The results emphasize leakage control, transparent comparisons, calibration assessment and the need for same-task prospective evaluation before clinical application.

doi.org/10.1038/s41598-026-80931-1Open accessBiomarkers

AgingResearch.News