Construction And Evaluation Of BP Neural Network-based Prediction Model For Multidrug-resistant Organism Infection In The Intensive Care Units

Purposes1.To explore the risk factors for the occurrence of multidrug-resistant organism(MDRO)infection in the intensive care unit(ICU),and develop the MDRO risk prediction model using BP neural network;2.To evaluate the prediction performance of BP neural network and analyze the characteristic importance of the risk factors to provide reference for early identification of high-risk groups and formulation of targeted prevention and control strategies.MethodsThis study developed the Data Extraction Table for Risk Factors of MDRO Infection in ICU patients through literature review and expert meetings.With the help of this tool,the data of 688 patients admitted to a tertiary care hospital comprehensive ICU in Qingdao from August 2021 to January 2022 were collected as a modeling group.The patients were divided into the MDRO infection group(109 cases)and the non-MDRO infection group(579 cases)based on whether MDRO occurred during ICU.The risk factors for MDRO infection were identified based on Lasso and stepwise regression analysis.The risk factors included were used as input variables to construct the BP neural network model.The model was validated internally and externally and the importance of predictors was analyzed.The internal validation of the model was performed using Bootstrap sampling and the modeling group data was randomized into training and test set on an 8: 2 basis.A total of 238 ICU patients in a tertiary care hospital in Qingdao from May 2022 to July 2022 were collected as an external validation group to externally validate the constructed MDRO infection prediction model.The calibration of the model was evaluated by the calibration curve.The AUC,sensitivity,specificity,accuracy were used to evaluate the discrimination of the model.Results1.A total of 38 articles were included to draw up the primary draft of the MDRO infection risk factor data extraction table for ICU patients.Subsequently,13 experts were invited to hold an expert meeting.The Cr was 0.896,the Ca was 0.900,and the Cs was0.892.The final draft of Data Extraction Table for Risk Factors of MDRO Infection in ICU patients was formed.It included six aspects: general information of patients,severity of disease,nutritional status,invasive operation,drug use and laboratory indicators.2.A total of 688 patients were included in the study,109(15.84%)in the MDRO infection group and 579(84.16%)in the non-MDRO infection group.The risk factors were screened by Lasso and stepwise regression analysis,including APACHE II(OR=1.06),numbers of antibiotics use(OR=1.81),comorbid chronic lung disease(OR=2.02),hypoproteinemia(OR=3.59),length of hospital stay(OR=1.04),length of ICU stay(OR=1.02),long-term bed rest(OR=3.51),use of antibiotics before ICU admission(OR=2.95),and invasive operation before ICU admission(OR=2.20).3.Based on the risk factors determined by Lasso and stepwise regression analysis,this study used BP neural network to develop a prediction model.The accuracy,sensitivity and specificity of the training set were 0.895,0.818 and 0.811,respectively.The AUC was 0.889(95%CI:0.852-0.925).The accuracy,sensitivity,specificity and AUC of the test set were 0.918,0.857,0.864 and 0.919(95%CI:0.856-0.983),respectively.In the BP neural network model,the importance of risk factors were ranked as length of hospital stay,length of ICU stay,long-term bed rest,use of antibiotics before ICU admission,APACHE Ⅱ,invasive operation before ICU admission,numbers of antibiotics use,combined with chronic lung disease,and hypoproteinemia.4.238 patients were enrolled in the external validation group,31(13.03%)were MDRO patients and 207(86.97%)were in the non-MDRO group.The specificity,sensitivity,accuracy and AUC of the external validation set were 0.715,0.806,0.852 和0.811(95%CI: 0.731-0.891),respectively.Conclusions1.The incidence of MDRO infections in the ICU of this hospital is low,and the most detected drug-resistant organism is carbapenem resistant Acinetobacter baumannii.The risk factors for the occurrence of MDRO infections are length of hospital stay,length of ICU stay,long-term bed rest,use of antibiotics before ICU admission,APACHE Ⅱ,invasive operation before ICU admission,numbers of antibiotics use,combined with chronic lung disease,and hypoproteinemia.2.The MDRO infection prediction model developed based on BP neural network is validated internally and externally to have good predictive effect and can be used for prediction of MDRO infection in ICU.Medical workers identify high-risk groups with high-risk factors early,use the model for early screening and take targeted measures to cut off the occurrence and transmission of MDRO infections from the source.