| Objective:1.To analyze the epidemiological characteristics of acute kidney injury in patients who used loop diuretics and mannitol during hospitalization.2.To construct three risk prediction models of logistic regression,random forest and neural network for acute kidney injury caused by combined use of loop diuretic and mannitol during hospitalization,and evaluate the model predictive value,and select the model with the best predictive ability.Methods:1.Multi-center retrospective collection of people who used combination of loop diuretics and mannitol during hospitalization from the First Affiliated Hospital of Xinjiang Medical University from January 2013 to December 2017and from December 2007 to May 2017,the Central Asian University Xiangya Third Hospital used the patients with loop diuretics and mannitol during hospitalization.The patient's gender,age,ethnicity,and various biochemical indicators before and after the combination were recorded in detail,and 70 factors including other drugs,combined diseases and medical history were combined.2.Statistical analysis was performed using SPSS21.0 software,and the primary variables of the included variables were screened.P<0.05 was considered statistically significant.The samples were randomly divided into a model development cohort and an external verification cohort according to 7:3.3.The univariate analysis preliminarily selected meaningful factors(P<0.5)into multivariate logistic regression analysis to determine the risk factors affecting the occurrence of acute kidney injury,and established a logistic regression prediction model.The predictive value of the model was assessed using Hosmer-Lemeshow test(H-L),self-validation,receiver operating characteristic(ROC)curves and external validation.4.Establish a random forest prediction model using the R language version 2.11.1,and perform external validation,and calculate accuracy,sensitivity,and specificity,plot the ROC curve,and calculate the area under the curve(AUC).5.The neural network prediction model is established using the multi-layer perceptron of the neural network module in the SPSS21.0 software,and using the accuracy,sensitivity,and specificity,the ROC curve was drawn and the area under the curve(AUC)was calculated.Using the accuracy,sensitivity,and specificity,the ROC curve is plotted and the area under the curve(AUC)is calculated for evaluation.Results:1.There were 9673 cases of data from two centers,and 3248 cases of patients using only loop diuretics,the incidence of AKI was 11.79%.A total of 3369 patients were treated with mannitol alone(excluding any type of diuretic),and the incidence of AKI was 3.41%.A total of 2634 patients with loop diuretics and mannitol were used in combination(using mannitol for 3 days before and after the use of a loop diuretic),and the incidence of AKI was 16.40%.There were 287 cases of secondary medication(using mannitol within 3-7 days before and after the use of loop diuretics),and the incidence of AKI was 7.67%.There were 134 cases of tertiary medication(using mannitol within 7 to 15 days before and after the use of myelin diuretics),and the incidence of AKI was 8.40%.2.Logistic regression analysis showed serum sodium(P=0.033,OR=1.354,95%CI 1.025-1.788),mean corpuscular volume(P=0.042,OR=1.358,95%CI 1.011-1.825),urea(P=0.047,OR=1.328,95%CI1.004-1.757),baseline creatinine(P=0.000,OR=2.678,95%CI 2.033-3.527),insulin(P=0.000,OR=1.829,95%CI 1.390-2.406),digoxin(P=0.012,OR=2.426,95%CI1.211-4.857)is the risk factor in the risk prediction model of acute kidney injury.The prediction model is Logit(P)=In(P/1-P)=-2.620+0.303×serum sodium+0.306×mean corpuscular volume+0.284×urea+0.985×baseline creatinine+0.604×insulin+0.886×digoxin.Model evaluation:H-L test?~2=9.213,P=0.238,accuracy was 68.5%,specificity was 70.04%and sensitivity was 60.93%.The AUC of ROC curve was 0.633?0.029(95%confidence interval 0.575-0.690).3.The results of the random forest model shown that the number of 9 variables was the optimal number of variables in the model,the variable importance was pre-medication creatinine(baseline creatinine),platelet distribution width,urea,total bilirubin,direct bilirubin,leukocytes,aspartate transferase,serum albumin,serum chlorine.Model evaluation:the prediction accuracy was 61.7%,the sensitivity was 56.1%,and the specificity was 73.8%.The AUC of ROC curve was 0.688.4.The results of neural network model showed that the importance of variables were aspartate transferase,alanine aminotransferase,urea,direct bilirubin,hematocrit,total protein,serum calcium and percentage of monocytes,etc.Model evaluation:the training set accuracy is 83.1%,the test set is 86.7%,the verification set is 81.4%,and the AUC of ROC curve was 0.656.5.According to the AUC evaluation index and other indicators comprehensive evaluation,the random forest prediction model is the model with better prediction ability among the three risk prediction models.Conclusion:1.The incidence of acute kidney injury induced by loop diuretics(11.79%)was higher than mannitol(3.41%)during hospitalization(P<0.05).The combined use of loop diuretics and mannitol(16.40%)during hospitalization was associated with a higher incidence of acute kidney injury than the use of loop diuretics(11.79%)and mannitol alone(3.41%)(P<0.05).The incidence of acute kidney injury was higher in combination(16.40%)than in secondary medication(7.67%)and tertiary medication(8.40%)(P<0.05).The combined use of loop diuretics and mannitol during hospitalization has the highest incidence of acute kidney injury and the highest risk.2.Serum sodium,mean corpuscular volume,urea,baseline creatinine,insulin,digoxin are risk factors in the logistic regression risk prediction model of acute kidney injury,through the accuracy,specificity,sensitivity and external validation AUC results get the predictive power of the model;3.Random forest prediction model showed baseline creatinine,platelet distribution width,urea,total bilirubin,direct bilirubin,hemameba,aspartate transferase,serum albumin,serum chlorine as risk factors,through the accuracy,specificity,sensitivity and external validation AUC results get the predictive power of the model;4.The neural network model shows that aspartate transferase,alanine aminotransferase,urea,direct bilirubin,hematocrit,etc.are important influencing factors of this model,the accuracy of the model is higher,but the accuracy is lower,and the overall predictive ability of the model is medium;5.Both urea nitrogen and baseline creatinine appeared in three models,AST,direct bilirubin appeared in the random forest model and the neural network model of the first few variables,explain that these four variables are essential for influencing the occurrence of AKI;6.Compared with the other two models,the random forest model is the optimal model after accuracy,sensitivity,specificity,ROC curve and AUC calculation and comprehensive score evaluation,and has a good predictive ability... |
PDF Full Text Request