| BackgroundNosocomial infection is one of the most important causes of death among inpatients,ranking the eighth in the United States.UTI is one of the nosocomial infections,accounting for 36%-40%of the total.Once UTI occurs,it often affects the patient's recovery and increases the cost of treatment.As the susceptible population of nosocomial infection,the incidence and mortality of nosocomial infection in elderly patients are high,which needs to be paid more attention by the medical staff.Therefore,it is very important to predict the risk of early postoperative complications in elderly fracture patients,which can help clinicians to understand the prognosis of patients and help doctors to make clinical decisions.ObjectiveThrough the analysis and mining of the text-based data and the digital data of the laboratory test of the elderly orthopedic surgery patients in the real world of his system,the risk prediction model of the postoperative UTI of the elderly orthopedic surgery patients was established by using the machine learning algorithm,in order to provide reference for the clinical prevention and treatment.MethodResearch object From September 1,2014 to November 30,2018,4611 clinical electronic medical records were collected from the hospital information management system,laboratory information management system and medical image information system of Wangjing Hospital of traditional Chinese medicine,which met the inclusion criteria.Research factors In order to ensure the multi angle and multi dimension of the research,we should extract as many features as possible from the medical record information that can be used for observation.The data mainly includes 273 related features in the electronic medical record,including 189 continuous features and 84 discrete features.Statistical method Use Excel to store the data,then transform and reduce the dimension of the data,finally carry out quality control,and finally complete the construction of the database.The features in the database are counted,and the features with the missing ratio greater than 30%are eliminated,and the features with the missing ratio less than 30%are interpolated multiple times.Chi square test is used for discrete feature,t test for continuous feature and Wilcoxon test for normal distribution.Five machine learning models(logistic regression,balanced bagging classifier,easy ensemble classifier,balanced random forest classifier,XG boost)with statistically significant features were fed into python.The best model was determined by comparing the model parameters(AUC,etc.)and dimensionality was reduced according to the model results.Part of the model features were removed and the model was improved Row optimization to get the final model.Research findings Ten risk characteristics were obtained by logistic model,including catheterization,history of malignancy,large platelet ratio,minimum value of ?2-microglobulin before operation,minimum value of pH before operation,dislocation and sprain of lumbar and pelvic joints and ligaments,thrombin clotting time,other heart rate disorders,and other diseases of urinary system.There were 7 protective features:intraspinal anesthesia,local anesthesia,postnatal deformation of fingers and toes,nerve block,shoulder damage,preoperative maximum urine specific gravity,and positive urine routine binary variables.Model parameter AUC=0.8163,accuracy 0.6241,specificity 0.8333,sensitivity 0.6170,positive predictive value 0.9910,negative predictive value 0.0681,positive release ratio 3.7020,negative release ratio 0.4590.Modelformulalogistic(C=1.0,class_weight='balanced',dual=False,fit_intercept=True,intercept_scalin g=1,max_iter=100,multi_class='warn',n_jobs=None,penalty='12',random_state=None,solver='warn',tol=0.0001,verbose=0,warm_start=False)Research conclusion Based on his real world clinical data,using machine learning algorithm,a UTI logistic risk prediction model was established for the elderly orthopedic patients after operation.Through internal verification,it was found that the model had a good prediction ability for the elderly orthopedic patients after operation.According to the model features,17 UTI related features were found.There was a positive correlation between catheterization,personal history of malignant tumor,large platelet ratio,? 2-microglobulin min,shoulder damage,urine pH minimum,dislocation and sprain of lumbar and pelvic joints and ligaments,other heart rate disorders and other diseases of urinary system.There was a negative correlation between auxiliary anesthesia,local anesthesia,postnatal deformation of fingers and toes,nerve block,shoulder damage,urine specific gravity Max and urine routine.InnovationBased on the real world clinical data of his,a large number of samples are included,which are patients of general orthopedic surgery.The observation features are comprehensive,the model features are not preset,and the high latitude clinical features are completely searched through statistical analysis.In addition,compared with the logistic regression in traditional SPSS,it can optimize the sample imbalance data,the model is more stable and reliable,and has better prediction ability for positive or negative outcomes.ICD-10 coding is used to standardize and reduce the dimension of text type diagnosis information,which improves the utilization rate of text type clinical information. |
PDF Full Text Request