An Explainable Machine Learning Model to Predict Acute Kidney Injury After Cardiac Surgery: A Retrospective Cohort Study

Abstract

Background: To derive and validate a machine learning (ML) prediction model of acute kidney injury (AKI) that could be used for AKI surveillance and management to improve clinical outcomes. Methods: This retrospective cohort study was conducted in Fuwai Hospital, including patients aged 18 years and above undergoing cardiac surgery admitted between January 1, 2017, and December 31, 2018. Seventy percent of the observations were randomly selected for training and the remaining 30% for testing. The demographics, comorbidities, laboratory examination parameters, and operation details were used to construct a prediction model for AKI by logistic regression and eXtreme gradient boosting (Xgboost). The discrimination of each model was assessed on the test cohort by the area under the receiver operator characteristic (AUROC) curve, while calibration was performed by the calibration plot. Results: A total of 15,880 patients were enrolled in this study, and 4845 (30.5%) had developed AKI. Xgboost model had the higher discriminative ability compared with logistic regression (AUROC, 0.849 [95% CI, 0.837–0.861] vs 0.803[95% CI 0.790–0.817], P<0.001) in the test dataset. The estimated glomerular filtration (eGFR) and creatine on intensive care unit (ICU) arrival are the two most important prediction parameters. A SHAP summary plot was used to illustrate the effects of the top 15 features attributed to the Xgboost model. Conclusion: ML models can provide clinical decision support to determine which patients should focus on perioperative preventive treatment to preemptively reduce acute kidney injury by predicting which patients are not at risk.