Prediction of acute kidney injury following cardiac surgery by machine learning

  Background  Cardiac surgery-associated acute kidney injury (CSA-AKI) is one of the major complications after cardiac surgery, and its course has a negative impact on the short- and long-term survival of patients.

  Objective  To develop predictive models based on machine learning techniques, so as to identify patients at high risk for CSA-AKI after cardiac surgery.

  Methods  A total of 638 patients who underwent cardiac surgery in the First Medical Center of Chinese PLA General Hospital from January 1, 2017 to June 1, 2018 were enrolled in the study. Totally, 78 variables including demographic characteristics, complications, preoperative medication, laboratory test results, and operation-related data were included in the analysis for modelling. In this study, the support vector machine (SVM), decision tree (DT) and random forest (RF) were used to develop the prediction model of CSA-AKI. The performance of the above predictive models was evaluated by area under the receiver operating characteristic (ROC) curve and decision curve analysis (DCA). Shapley additive exPlanation (SHAP) was used for model visualization.

  Results  Of the 638 patients included in the analysis, 188 cases (29.5%) developed CSA-AKI in the first week after surgery. Among the three machine learning algorithms, RF model achieved the best performance in AUC and DCA, with sensitivity of 0.784, specificity of 0.934, accuracy of 0.927 and AUC value of 0.890 (95% CI: 0.762-1.000), which were higher than those of DT model and SVM model. The SHAP plots visualized the risk of developing CSA-AKI at the individual level. In the importance matrix of RF model, the top 10 variables were as follows: creatinine clearance, hemoglobin, operation time, ejection fraction, intraoperative urine output, left atrial diameter, surgical weight of the intervention, serum creatinine, intraoperative blood loss, and cardiopulmonary bypass time.

  Conclusion  This study successfully establishes machine learning models for predicting patients at high-risk for CSA-AKI, thus enabling clinicians to optimize treatment strategies and minimize postoperative complications.