Study On MDROS Infection Prediction Model Of ICU Patients Based On Random Forest

Research Purpose:Through literature search and expert consultation,the influencing factors of MDROs infection in intensive care unit（ICU）patients were identified.The related data of ICU patients in two research centers in five years were retrospectively collected to understand the status of MDROs infection in the region.Logistic regression and random forest（RF）prediction models were constructed and verified,and the prediction models with good prediction efficiency and more suitable for clinical practice were selected to guide clinical early identification and early intervention of high-risk patients and reduce the infection rate of MDROs.Research Methods:The influencing factors of MDROs infection in ICU patients were determined by literature review and expert consultation.The MDROs infection in ICUs of two tertiary general hospitals in a city of Jiangsu Province from January 1,2016 to December 31,2020 was investigated.The patients were screened according to the inclusion criteria,and the data of selected patients were collected to form a target database containing 618 patients’ data.The target database patient data were randomly divided into training set and test set according to the ratio of 8:2.Using the training set data,the Logistic regression prediction model was established by using SPSS through single factor and multi-factor analysis,and the OR（95%CI）value of each influencing factor was obtained.RF prediction model was established using Spyder to obtain the importance of each influencing factor for MDROs infection in ICU patients.The test set data were substituted into the output predictive value of the model,and the discrimination and calibration of the two models were evaluated by F-measure,Area Under Curve（AUC）and Brier score.Research Results:1.Through literature retrieval and two rounds of expert consultation,26 indicators related to MDROs infection in ICU patients were determined.It included patient source,history of MDROs infection or colonization in the past,APACHE II（Acute Physiology and Chronic Health Evaluation II）,days of fever,this admission surgery or not,days of combined antibiotics,days of mechanical ventilation,times of bronchoscopy,procalcitonin,serum albumin,days of ICU stay,whether immune system diseases（or radiotherapy,chemotherapy,etc.）were combined,whether infection lesions were found in imaging data,types of antibiotics used,days of indwelling catheter,whether gastrointestinal integrity was damaged during operation,interleukin-6,days of antibiotics application,days of active / venous catheterization,whether tracheotomy was performed,whether diabetes was combined,age,times of indwelling catheter,times of indwelling catheter,times of partially activated thrombin,and white blood cell count.2.A total of 618 patients（309 cases of infection,309 cases of non-infection）were included,and the majority of MDROs infection patients were males（72.82%）.Gram-positive bacteria infection in 41 cases（13.27%）,Gram-negative bacteria infection in 268 cases（86.73%）;acinetobacter baumannii（36.57%）,Klebsiella pneumoniae（32.36%）and Pseudomonas aeruginosa（11.00%）were the three types of MDROs with the largest number of infections in the two hospitals.3.In the process of fitting with the training set data,the RF model selected the days of starting fever,indwelling catheter,mechanical ventilation,arteriovenous catheter,white blood cell count,ICU hospitalization days,activated partial thromboplastin time,antibiotic use days,APACH II,serum albumin,procalcitonin,age,and antibiotic combination days as the important influencing factors for MDROs infection in ICU patients.4.The results of multivariate logistic regression analysis showed that the intraoperative gastrointestinal damage（OR = 2.911）,the number of bronchoscopy（OR = 1.922）,the number of arteriovenous catheterization（OR = 1.509）,the number of days of fever（OR = 1.105）,the number of days of mechanical ventilation（OR = 1.053）,and WBC（OR = 1.028）were the relative risk factors for MDROs infection in ICU patients,and the operation during hospitalization（OR = 0.327）was the relative protective factor（p < 0.05）.Logistic regression equation model was established as follows: Logit（P）=-2.334 + 1.069 ×（whether the gastrointestinal tract was damaged during operation）+ 0.653 ×（number of bronchoscopy）+ 0.411 ×（number of arteriovenous catheterization）+ 0.100 ×（days of fever）+ 0.052 ×（days of mechanical ventilation）+ 0.028 ×（WBC）-1.119 ×（this admission surgery or not）.5.The AUC of Logistic regression model was 0.706（95 %CI 0.614 ～ 0.798,p < 0.001）,and the F-measure was 0.70.The calibration of Logistic regression prediction model: Brier score was0.23.The discrimination of RF model: AUC was 0.935（95 %CI 0.895 ～ 0.975,p < 0.001）,Fmeasure was 0.84;calibration of RF model: Brier score is 0.11.Research Conclusion:Indwelling catheter days,mechanical ventilation days,ICU hospitalization days,fever days,white blood cell count,arterial/venous catheterization days,serum albumin,activated partial thromboplastin time,antibiotic use days,APACHE II are important influencing factors for predicting MDROs infection in ICU patients.In the past five years,the majority of MDROs infection patients in the two research centers were males,and Gram-negative bacteria were more resistant.Acinetobacter baumannii,Klebsiella pneumoniae and Pseudomonas aeruginosa were the three types of MDROs with the largest number of infections in the two hospitals.RF model has higher discrimination and calibration than Logistic regression model,which is of great significance for accurately predicting high-risk patients and guiding clinical intervention as soon as possible to prevent MDROs infection.In the future,it is expected to integrate RF model into hospital information system for further research,to give full play to the professional initiative of ICU nursing in information integration and disease observation.