Prediction of Acute Kidney Injury after Extracorporeal Cardiac Surgery (CSA-AKI) by Machine Learning Algorithms

Abstract

Background: Acute renal failure after extracorporeal cardiac surgery under general anesthesia is high and unpredictable, but machine learning algorithms could change this. A feasible approach is to use machine learning models to construct models to predict acute kidney injury after extracorporeal cardiac surgery (CSA-AKI) and screen for the best predictive model. Method: From January 2014 to December 2021, 2187 patients undergoing extracorporeal cardiac surgery at the third hospital of Hebei Medical University and the first medical centre of Chinese PLA General Hospital were collected in this study. After excluding 923 patients who did not meet the inclusion criteria, a dataset of 1264 patients with 125 clinical indexes was constructed. After screening the feature variables using Least absolute shrinkage (LASSO) regression, the dataset was randomly divided into a training set (70%), test set (30%), and six machine learning algorithms, including extreme gradient boosting (XGBoost), logistic regression (LRC), light gradient boosting machine (LGBM), random forest classifier (RFC), adaptive boosting (AdaBoost), and K-nearest neighbor (KNN), were used in training set for predicting the CSA-AKI. The machine learning model with the best predictive performance was selected to complete external validation of the test set. The SHapley Additive exPlanations (SHAP) algorithm was used to interpret the model. Results: Of all 1264 patients, 372 (29.43%) patients presented with CSA-AKI. The LASSO regression eliminated 22 feature variables out of 125 before model development. Among the six prediction models, the RFC prediction model has the best prediction performance, with an Area Under Curve (AUC) value of 0.778 (95% CI: 0.726–0.830) in the test set and the best net benefit compared to the other tools. SHAP explained the impact of different feature variables on the predicted outcome, where the three most influential feature variables were creatinine clearance (CRC), intraoperative urine output (mL/kg/h) and age. Conclusion: We developed an RFC prediction model to predict the CSA-AKI, which has good predictive performance and can explain the factors affecting the prediction results of cases by integrating the SHAP method.