Development and Validation of a Machine Learning Model for Predicting Serum Creatinine-Defined Acute Kidney Injury in Older Adults After Cardiac Surgery.

BACKGROUND

Cardiac surgery-associated acute kidney injury (CSA-AKI) is a frequent and devastating postoperative complication, particularly among older adults. Accurate risk stratification and early prediction of CSA-AKI are essential for guiding preventive strategies and optimizing clinical decision-making.

METHODS

In this retrospective study, data from two centers (n=623) were utilized for model training and internal validation, whereas data from a third, distinct center (n=110) were reserved for external validation. CSA-AKI was defined according to the Kidney Disease: Improving Global Outcomes (KDIGO) Serum creatinine criteria. Key predictors were identified using a consensus of four methods: Least Absolute Shrinkage and Selection Operator (LASSO), Recursive Feature Elimination (RFE), Boruta, and Random Forest-based filtering. Six machine learning (ML) models, including Logistic Regression (LR), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM), were developed utilizing five-fold cross-validation. Predictive performance was assessed using the area under the receiver operating characteristic curve (AUC). The SHapley Additive exPlanations (SHAP) approach was applied to interpret the best-performing model.

RESULTS

Development of CSA-AKI was noted in 177 patients (24.1%) during the first postoperative week. In terms of comparative performance, LightGBM exhibited the greatest AUC (0.784, 95% confidence interval [CI]: 0.702-0.859). The most influential features were lactate, surgical duration, activated partial thromboplastin time (APTT), transfusion volume, and Prothrombin Time (PT). SHAP-based summary and force plots interpreted the model at global and local levels. Furthermore, SHAP dependence plots elucidated non-linear effects of single features on CSA-AKI risk.

CONCLUSION

Machine learning models demonstrate high efficacy in predicting CSA-AKI risk in older adults. The LightGBM model outperformed other algorithms; coupled with interpretability tools, it can assist clinicians to identify high-risk patients earlier and optimize perioperative management.