Development And Validation Of A Machine Learning Model For Predicting Acute Kidney Injury After Craniotomy For Brain Tumors

PurposePatients with brain tumors can suffer acute kidney injury(AKI)after craniotomy due to a variety of perioperative factors,which can prolong hospital stays and even result in increasing morbidity and mortality from remaining postoperative complications.Most established AKI prediction models do not adequately consider the impact of various intraoperative variables on predictive performance and lack the interpretation of prediction process and results.We aimed to use electronic healthy record(EHR)data combined with machine learning(ML)techniques to construct a perioperative AKI prediction framework to accurately assess the risk of postoperative AKI(PO-AKI)in patients undergoing craniotomy,and to combine model interpretation algorithms to provide insight into model prediction results and improve the interpretability and transparency of the model.MethodsWe extracted data from patients who had underwent craniotomy for brain tumors at Southern Medical University Nanfang Hospital between January 2001 to May 2018.The variables used for analysis included preoperative variables such as demographic characteristics,comorbidities,preoperative medications and preoperative laboratory tests,as well as intraoperative variables like intraoperative vital signs,intraoperative medications,and surgical information.The ML algorithms utilized included Decision Tree(DT),Random Forest(RF),Logistic Regression(LR),Support Vector Machine(SVM)and K-Nearest Neighbor(KNN)and extreme gradient boosting(XGBoost).A series of models were constructed by combining different ML algorithms with data at different perioperative periods,and the performance of the models was evaluated by the area under the receiver operating characteristic curve(AUC),aiming to verify the superiority of the ML models and the influence of intraoperative factors on the occurrence of PO-AKI.Finally,the SHapley Additive exPlanations(SHAP)algorithm was applied to interpret the model globally and locally by drawing SHAP summary diagrams,SHAP dependency diagrams and SHAP waterfall diagrams.ResultsData of 1999 patients were finally analyzed for this study,of which 81(4%)developed PO-AKI.The XGBoost model performed best when modelled using the preoperative dataset(AUC=0.946,95%CI 0.889-0.979).The XGBoost model was not optimal but performed consistently and robustly when combining the intraoperative dataset(AUC=0.914,95%CI 0.848-0.957)and the full dataset(AUC=0.951,95%CI 0.895-0.982).Except for the DT model,all ML models improved their AUC values when intraoperative data were added to the baseline model,with a statistically significant change for the SVM model in particular(0.939 vs 0.978,p<0.05).We can see from the SHAP summary plots of XGBoost models with three different periods,preoperative blood glucose and operative duration were highly positively associated with the risk of PO-AKI,while female,the intraoperative use of vasoactive drug and antibiotic were protective factors for the occurrence of PO-AKI.And intraoperative low end-tidal carbon dioxide partial pressure(PetCO2)and systolic blood pressure would increase the risk of PO-AKI.According to the SHAP dependency plots,the risk of PO-AKI was reduced by maintaining preoperative blood creatinine levels above 6070μmol/L,and increased progressively when preoperative blood glucose levels were above 5-5.5 mmol/L、the duration of surgery exceeded 4 hours or the value of PetCO2 below 25mmHg.ConclusionsThis study introduced an interpretable model framework to predict the risk of POAKI in patients undergoing brain tumor resection by using real-time EHR data.The framework has good predictive and interpretative performance,which can enable early warning of PO-AKI and provide analysis of relevant predictors to inform individualized perioperative management and treatment strategies for high-risk patients.