Derivation And Validation Models For Predicting Acute Kidney Injury After Cardiac Surgery In Patients With Normal Renal Function

BackgroundAcute kidney injury(AKI)is not an uncommon complication after cardiac surgery in patients with normal renal function,which is related with increased chronic kidney disease and mortality.Delayed recovery of renal function after AKI is common.Compared with patients underwent AKI who reversal within a short period of time,patients with persistent AKI had a worse prognosis.Currently,limited strategies to treat AKI or facilitate kidney recovery after kidney injury,which highlights the urgency of early identification of high-risk groups.Hence,tools that enable early identification of the high-risk subpopulation for AKI and recognition of persistent AKI need to be developed urgently.However,at present,no models are available for predicting AKI or persistent AKI after cardiac surgery in patients with normal renal function.ObjectiveThe present study had two primary objectives.First,based on clinically information that is easy accessible,derivation and validation a model for predicting AKI after cardiac surgery in patients with normal renal function.Second,developed a model for predicting persistent AKI after cardiac surgery in patients with normal renal function.MethodsSection 1:A total of 22348 consecutive patients with normal renal function at baseline and underwent cardiopulmonary bypass surgery were analyzed retrospectively in our hospital.Among them,15,701 patients were randomly chosen for model construction and the remaining for model validation.Clinical data of patients were extracted from electronic medical records of our hospital.The interesting outcome was AKI after cardiac surgery defined according to serum creatinine of the Kidney Disease Improving Global Outcomes(KDIGO)criterion.Based on preoperative data,four models were developed by means of logistic regression(LR),LR with least absolute shrinkage and selection operator(Lasso)regularization(Lasso-LR),random forest(RF)and extreme gradient boosting(XGBoost)and assessed the prediction performance of the models in validation group.The discrimination,calibration,and clinical value of the model were evaluated.The area under the receiver operating characteristic curve(AUC)was used to assess the discrimination of the model.The calibration and clinical value were estimated using calibration plot and decision curve analysis(DCA),respectively.Section 2:A retrospective analysis of 5368 consecutive patients with preoperative normal renal function and underwent AKI after cardiac surgery with cardiopulmonary bypass was performed at our hospital.Among them,3768 patients were randomly chosen for model construction and the remaining for model validation.Clinical data from the preoperative,intraoperative and postoperative period(up to the time of enrollment)were extracted from medical records.The interesting outcome was persistent AKI after cardiac surgery referenced from persistent AKI criterion defined by the Acute Dialysis Quality Initiative.We developed four models using from 34 clinical features using LR,Lasso-LR,RF and XGBoost and compared the prediction performance of the models in validation group.The discrimination,calibration,and clinical value of the model were evaluated.The AUC was used to assess the discrimination of the model.The calibration and clinical value were assessed using calibration plot and DCA,respectively.ResultsSection 1:AKI incidence rates in the training and validation groups were 25.2%(n=3955)and 24.4%(n=1621),respectively.The XGBoost model showed best discrimination in validation group with the AUC was 0.806,followed by RF(AUC=0.783),LR(AUC=0.747),and Lasso-LR(AUC=0.740).Lasso-LR had lower but good discriminative ability.The calibration plots indicated XGBoost,Lasso-LR and LR models shown good agreement between the model prediction and actual observation.Among them,the calibration of XGBoost was excellent.However,the RF model underestimate the actual rate in moderate-to high-risk areas.DC A demonstrated that within most of the range of prediction thresholds,using all models to predict AKI risk provided superior net benefits,the XGBoost model provided the greatest net benefits.Section 2:The persistent AKI rates of training and validation groups were 50.6%(n=1905)and 48.5%(n=776),respectively.The XGBoost model showed best discrimination in validation group with the AUC was 0.951,followed by RF(AUC=0.947),LR(AUC=0.769),and Lasso-LR(AUC=0.761).Lasso-LR had lower but good discriminative ability.The calibration plots indicated all models shown excellent agreement between prediction and actual observation.Among them,the calibration of XGBoost was best.DCA demonstrated that within the entire range of prediction thresholds,using all models to predict AKI risk provided superior net benefits,the XGBoost model provided the greatest net benefit.ConclusionsBased on easily available clinical data,we developed models using LR,Lasso-LR,RF and XGBoost for predicting AKI after cardiac surgery in patients with normal renal function,respectively.Models for predicting persistent AKI after cardiac surgery were also developed.Whether predicting AKI or persistent AKI,XGBoost model presented the best performance among these models and had a high clinical application value,which may help early identification of the high-risk subpopulation,optimizing treatment options and improving prognosis.