| With the advent of an aging society,the prevalence of lumbar degenerative diseases(LDD)will continue to increase,with lumbar spondylolisthesis,lumbar disc herniation and lumbar spinal stenosis being the most common.Recent years significant advances have been made in spinal surgical techniques and implants for lumbar degeneration,including open interbody fusion and minimally invasive interbody fusion.Among the many current surgical treatment methods,posterior lumbar interbody fusion(PLIF)is still widely used in the spinal surgery treatment of lumbar spondylolisthesis and lumbar spinal stenosis because of its relatively simple approach,relatively safe,high vertebral fusion rate and good surgical effect.However,as with other procedures,the surgical effects of PLIF are occasionally affected by postoperative complications.Surgical site infection(SSI)is a common postoperative complication,but it may have serious consequences in spinal surgery.Postoperative administration of infection treatment progressed slowly,resulting in unplanned multiple surgeries,prolonged hospitalization,increased patient pain,reduced patient treatment satisfaction,and an increased risk of readmission.The increased readmission rate leading to the increased burden of public health care costs is another issue of current concern.The retrospective study was designed to collect a large number of relevant medical history data,use data mining technology to explore the relevant risk factors of surgical site infection(SSI)after lumbar degenerative disease patients with posterior lumbar interbody fusion(PLIF),establish a risk assessment prediction model,and preliminary try to verify the clinical utility value of the model,so as to provide some reference for the clinical development of infection risk after PLIF.The study is divided into three parts,and each parts are summarized as follows.Part 1 Analysis of risk factors for infection after lumbar posterior spinal fusionObjective: To statistically analyze the relevant risk factors of infection after lumbar posterior vertebral fusion,provide a data basis for developing a later prediction model,and provide a certain reference for formulating measures to prevent infection after PLIF.Methods: A retrospective study was conducted on patients treated with PLIF in part of third-class hospitals in Beijing and Hebei Province from June 2019 to June 2021(patient-related case data results were provided by Zhong Wei Yun Institute of Medical Data Analytics & Application Technology Research Institute after data processing).Statistical analysis was performed using the SPSS Modeler 20.First,the K-means algorithm in the cluster division method was used to cluster a large number of uninfected patients in the data,and each category was randomly selected for class members in a certain proportion,with the highest similarity between members within the class and the lowest similarity between classes.Each category was randomly selected at a certain percentage,so that the continuous and classified variables were characterized by the mean ± standard deviation(SD)and count(percentage)after the preprocessed target dimension.Categorical data for the SSI and non-SSI groups were compared using either chi-square or Fisher’s exact test and continuous data using the Student-t test or Mann Whitney-U test as appropriate.In the univariate analysis,the variables that were tested as significant at a statistical level of P<0.05 were found,and these variables were subsequently included as independent variables in the multivariate Logistic regression models for fitting.Nonsignificant variables were removed by stepwise backward elimination,and variables with significant effects on infection were obtained.Hosmer-Lemeshow(H-L)test was applied to confirm the goodness of fit of the final model.P> 0.05 suggests acceptable results;further quantified by Nagelkerke R2,the larger the index between 0-1 indicates the higher accuracy of the prediction.Results: For the data included in this study,373 were diagnosed with SSI with an incidence of 4.4%(95%CI,2.2% to 6.5%).Following the K-means clustering,the SSI group consisted of 124 men and 249 women,with a mean age of 55.9 years(SD,14.6 years).The median time of onset of SSI was 9 days after surgery,with the earliest postoperative day 3 and the latest postoperative day 76.In SSI group,268 were superficial infections and 105 were deep infections.All 373 patients in SSI group were routinely microbiocultured,and 357 patients(95.7%)were cultured as positive.In 12 patients with deep SSI,and four patients with superficial SSI,no microbes were isolated.Univariate analysis indicated that the SSI and non-SSI groups showed significant differences in age classification variables,BMI,diabetes mellitus,chronic heart disease,renal insufficiency,preoperative stay time,total hospital length,ASA calss Ⅲ and above,surgical duration,postoperative drainage(ml),duration of postoperative use antibiotics(days),red blood cell(RBC)count,lymphocyte(LYM)count classification variables,fasting blood glucose(FBG)classification variables(P<0.05).The two groups showed no significant differences in gender,smoking,hypertension,cerebrovascular disease,lung disease,chronic liver disease,total protein(TP)count and classification variables,albumin(ALB)count and classification variables,white blood cell(WBC)and neutrophil(NEUT)count and classification variables,hemoglobin,hematocrit(HCT),platelet(PLT)count,and classification variables.Conclusions: In this study,the incidence of SSI was 4.4% after PLIF was treated with lumbar degenerative diseases.Body mass index(BMI),ASA calssⅢ and above,chronic heart disease(CHD),diabetes mellitus(DM),renal insufficiency(RI)and prolonged surgical duration were determined to be independently associated with SSI.Part 2 Construction of a prediction model for the risk of infection after posterior lumbar interbody fusion based on data mining technologyObjective: On the basis of the first part of the research,using decision tree classification,random forest classification,artificial neural network classification algorithm for the analysis of postoperative infection variables and accuracy,we try to build objective and accurate postoperative infection prediction model,aiming to.Methods: In this part we returned to the original sample using K-means cluster analysis to preprocess the total datas,on the basis of six independent influencing factors for the occurrence of infection in the first part of the study.The data dimensions were assigned,and SPSS Modeler 20 data modeling system was used as a tool.The C4.5 decision tree,C5.0 decision tree,random forest(RF),support vector machine(SVM)and artificial Neural network(ANN)algorithms were used here for comparative analysis.The minimum sample threshold in the branch nodes of the decision tree and random forest was controlled in 10,and then 70%,80%,90% and 100% of the total samples were randomly selected for classification model construction to improve the objectivity of the model.After comparison and selection of the most accurate classification method,the data mining results were brought into 100% sample data and summarized to obtain the characteristics of patients susceptible to infection after surgery,namely the infection model.Results: The K-means cluster analysis was reapplied in the total sample,After data pre-processing and clustering of 4 classes of the raw samples,data from 453 uninfected patients were obtained.The infected samples were 373,From 826 pieces of data,The ratio of uninfected to infected is roughly 1.2:1,After assigning the data dimension,C4.5 decision tree,C5.0 decision tree,random forest(RF),support vector machine(SVM)and artificial neural network(ANN)algorithms were used for comparative analysis.The results indicated that that 90.6% high accuracy was obtained by applying the random forest classification algorithm in this study.We aggregated the results of the random forest classification obtained by SPSS Modeler 20.Finally,the characteristics of patients prone to postoperative infection can be obtained(two modes of infection):((BMI=1)and(SD=1)and(ASA=1)and(RI=1))or(BMI=0)and(SD=1)and(DM=1)and(RI=1).Conclusions: In this study,more objective and accurate risk prediction models were obtained by using data mining techniques compared to applying logistic regression analysis.The random forest classification algorithm applied to this study yielded an average accuracy of 90.6%,Patients are prone to infection if obesity,renal insufficiency,and two modes of infection was obtained:(1)Patients with BMI≥28.0,renal insufficiency,heart disease,ASA class Ⅲ or above,and the surgical duration≥3 hours at the same time,are prone to infection.(2)Patients with diabetes,renal insufficiency and t the surgical duration≥3 hours,are prone to infection.Part 3 Clinical application and preliminary validation of the predictive risk of infection after posterior lumbar interbody fusionObjective: The aim was to observe and verify the accuracy of the infection risk prediction model,and to measure the applicable value of the model in practical clinical application.Methods: This study conducted a statistical analysis of case-related data from patients with lumbar degenerative disease hospitalized with PLIF at the Third Hospital of Hebei Medical University between January 2016 and June 2019.The infection risk prediction model constructed in Part II was used to evaluate and screen the subjects included in the study.According to the progress of the case collection,relatively small,medium and large samples were established.Avoiding the algorithmic optimal potential brought about by the number of samples studied in this part,random forest(RF),support vector machine(SVM)and C5.0 classification algorithms in the original sample data with high accuracy was applied in this part of study to obtained the final model with the highest accuracy.Results: The first phase of this part of the study(January 2016 to October 2017)included 654 patients undergoing PLIF surgery,21 cases with actual infections,infection incidence of 3.2%.In the second phase(November 2017 to June 2018),867 cases were included in the study,and 28 cases were actually infected,with an infection rate of 3.2%.In the third stage(from July 2018 to June 2019),1,286 cases were included in the study,and 48 cases were actually infected,with an infection rate of about 3.7%.Applying random forest(RF),support vector machine(SVM),and C5.0 classification algorithm model to predict the results of the three samples respectively,can be found that the accuracy of all the three algorithms improves with the increasing sample size.The random forest algorithm is still the most accurate,with an accuracy of 84%,84.8%,and 87.5% in the three different data samples,followed by the support vector machine and finally the C5.0 decision tree.In this part of the experiment,the highest precision of random forest was 87.5%,which is close to the 90.6% precision obtained in the second part,within a reasonable error range.Conclusions: Random forest(RF),support vector machine(SVM)and C5.0 classification algorithm predict the results respectively,and the accuracy of the three algorithms is improved with the increase of sample size.The random forest algorithm is still the most accurate,with an accuracy of 84%,84.8%,and 87.5% in the three different data samples.Thus,the highest precision of the random forest is 87.5%,which is close to the 90.6% precision obtained in the second part,within a reasonable error range. |
PDF Full Text Request