Prediction of acute kidney injury at the end of cardiac surgery using routine perioperative variables: a pre-specified benchmark of 210 machine learning combinations

Background Acute kidney injury after cardiac surgery (CSA-AKI) is common yet detected late. Risk stratification at the end of surgery could enable timely renal protection, but most machine learning (ML) models either use postoperative data or lack a transparent, pre-specified evaluation. Methods In a retrospective single-center cohort of 1,352 adult cardiac surgery patients (211 AKI, 15.6%), we predicted AKI from 32 routine preoperative and intraoperative (up to end-of-surgery) variables. We benchmarked 210 combinations (five feature-selection strategies × seven algorithms × six class weights) by repeated stratified 5-fold cross-validation (10 repeats) on a fixed development set ( n = 1,081), selected the final model by a pre-specified two-stage rule, and evaluated it once on a fixed test set ( n = 271). A screening-oriented threshold (sensitivity ≥0.80) was pre-specified. We additionally tested the few-shot transformer TabPFN, stability selection, bootstrap optimism correction, and simple ensembling, with SHAP for interpretability. Results The selected random forest (positive-class weighting, isotonic calibration) achieved a test-set AUC of 0.722 (95% CI 0.634–0.799), sensitivity 0.762, and specificity 0.515. No model family was decisively superior; penalized linear models generalized comparably, and none of the small-sample strategies or ensembling improved discrimination. Bootstrap correction reduced the apparent AUC from 0.965 to a corrected 0.724. Homocysteine ranked first in SHAP importance and was among ten independent factors in multivariable logistic regression. Conclusion Routine perioperative variables support a screening-oriented, interpretable prediction of AKI at the end of cardiac surgery, with modest discrimination. The lack of model superiority argues for parsimony, and homocysteine merits prospective study; external validation is required before clinical use.

Authors

Institutions

Publication Details

Journal
Frontiers in Medicine
Published
2026-11-16
DOI
https://doi.org/10.3389/fmed.2026.1992789
Primary Topic
Acute Kidney Injury Research
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Prediction of acute kidney injury at the end of cardiac surgery using routine perioperative variables: a pre-specified benchmark of 210 machine learning combinations

张凤伟, Wenqi Liu, Wenhui Zhang, Xiao Li et al.
1 links
Frontiers in Medicine
Acute Kidney Injury Research
article

Prediction of acute kidney injury at the end of cardiac surgery using routine perioperative variables: a pre-specified benchmark of 210 machine learning combinations

张凤伟, Wenqi Liu, Wenhui Zhang, Xiao Li, Yanjin Wei, Yangyang Han, Zhengrong Li, Shiming Wang
article en

Abstract

Background Acute kidney injury after cardiac surgery (CSA-AKI) is common yet detected late. Risk stratification at the end of surgery could enable timely renal protection, but most machine learning (ML) models either use postoperative data or lack a transparent, pre-specified evaluation. Methods In a retrospective single-center cohort of 1,352 adult cardiac surgery patients (211 AKI, 15.6%), we predicted AKI from 32 routine preoperative and intraoperative (up to end-of-surgery) variables. We benchmarked 210 combinations (five feature-selection strategies × seven algorithms × six class weights) by repeated stratified 5-fold cross-validation (10 repeats) on a fixed development set ( n = 1,081), selected the final model by a pre-specified two-stage rule, and evaluated it once on a fixed test set ( n = 271). A screening-oriented threshold (sensitivity ≥0.80) was pre-specified. We additionally tested the few-shot transformer TabPFN, stability selection, bootstrap optimism correction, and simple ensembling, with SHAP for interpretability. Results The selected random forest (positive-class weighting, isotonic calibration) achieved a test-set AUC of 0.722 (95% CI 0.634–0.799), sensitivity 0.762, and specificity 0.515. No model family was decisively superior; penalized linear models generalized comparably, and none of the small-sample strategies or ensembling improved discrimination. Bootstrap correction reduced the apparent AUC from 0.965 to a corrected 0.724. Homocysteine ranked first in SHAP importance and was among ten independent factors in multivariable logistic regression. Conclusion Routine perioperative variables support a screening-oriented, interpretable prediction of AKI at the end of cardiac surgery, with modest discrimination. The lack of model superiority argues for parsimony, and homocysteine merits prospective study; external validation is required before clinical use.

Frontiers in MedicineVol. 13
Linyi People's Hospital (CN), Shandong Provincial QianFoShan Hospital (CN), Shandong First Medical University (CN)
Good health and well-being
Acute Kidney Injury Research
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon