Risk Model To Predict AKI In ICU Based On Big Data

Objectives:To develop a model that can predict the risk of acute kidney injury in critically ill patients 48 h in advance by integrating data from electronic medical record systems in the ICU,and to validate its validity and accuracy.Method:A total of 9628 critically ill patients admitted to ICU from March 2015 to December 2019 were included in the study.The diagnostic criteria for AKI were serumcreatinine or urine criteria provided by the 2012 KDIGO clinical practice guideline.Data on demographics,vital signs,examination tests,clinical medication,and clinical interventions of all patients were collected and analyzed,accumulating 242 variables.Screening by exclusion criteria resulted in 2441 patients suitable for study entry and split 3:1 into derivation and validation groups.Serum creatinine and urine were not entered into the model construction as variables.Univariate analysis was performed to identify the susceptibility factors for AKI.With the help of R Studio software,logistic regression,random forest,SVM,GBM,AdaBoost,Ensemble Learning were used to try to construct the model,respectively.Explanatory logistic regression algorithms were selected to screen the indicators for entry into the model by stepwise regression forward and backward,and the final model and corresponding ROC curve were derived.Finally,Hosmer lemeshow test and decision curve analysis were used to evaluate the performance of the modelResults:Among 2441 patients,the incidence of AKI was 46.64%,which was centered within 3 days of ICU admission.the hospital mortality rate was significantly higher in the AKI group than in the non AKI group(12.7%vs 1.3%,P<0.001),which was statistically different.A number of risk factors for AKI were identified on univariate analysis:anemia,hypertension,diabetes,low chloride,high calcium,low pH,low PO2,high lactate,low white blood cell count,low MCHC,high PDW,high thrombin time and low fibrinogen,contrast exposure,ACEI,glycerfructose,midazolam,dexmedetomidine,and others.It was found in the subgroup analysis of blood pressure that the risk of AKI was more related to the extremes of blood pressure than average.The multivariable model including albumin infusion,fluid balance,diastolic blood pressure,PO2,blood glucose,platelets,base creatinine,serum sodium,age,norepinephrine,PPI,intra-abdominal infection,anemia,diabetes,glycerfructose,nutritional route was finally derived,with derivation AUCs of 0.822(95%CI,0.803-0.840)and validation 0.821(95%CI,0.787-0.855).In the Hosmer lemeshow test,data set in low probability regions the agreement was good,the model fit was excellent.In the decision curve analysis,the curves of derivation and validation cohorts were also higher than the extreme curves over a wide range of threshold intervals,and the models had a higher rate of benefit and a better fit.Conclusions:Derived a new prediction model of AKI in critically ill patients,which could predict the risk of AKI in patients 48 h earlier than the changes of urine volume or serumcreatinine,providing more possibilities for early diagnosis and intervention of AKI.