| Objectives:A recent report from the China Health Yearbook indicated that the number of inpatient surgeries within Chinese medical institutions escalated to approximately 81.031 million in2021.This significant number brings into focus the increasing importance of perioperative complications,which have come to constitute a pertinent societal issue.Perioperative Neurocognitive Disorders(PND),an umbrella term suggested by experts in November 2018 to describe cognitive impairments detected either preoperatively or postoperatively,are among the most prevalent perioperative complications in elderly patients.Published in renowned anesthesia-related journals,these disorders are noted to prolong the length of hospitalization,affect the quality of life,and increase postoperative mortality,so that have garnered substantial attention in the medical community.The incidence rate of PND fluctuates according to both the nature of the research and the type of surgery performed.For instance,in the first few weeks after non-cardiac surgery it can be as high as 53%.Currently,the pathogenesis of PND remain elusive,and early-stage interventions are markedly insufficient.Early accurate screening,implementing intervention measures or individualized anesthesia management plan as early as possible and providing special perioperative nursing measures,which are all crucial for the perioperative management and prevention of PND.However,rudimentary models based on traditional statistical methods are unable to satisfy the requisites for ongoing dynamic assessments of high-risk PND patients.Consequently,the dual aim of this study was firstly to develop an initial model based on machine learning algorithm using single-center hospital data to predict the risk of PND in the non-cardiac surgery.This is aimed at establishing the viability of applying machine learning to PND prediction.Secondly,this study also developed an online tool for PND prediction based on machine learning algorithm using multi-center hospital data to help clinical medical staff visually assess risks.Methods:1.We retrieved patients with PND after the non-cardiac surgery under general anesthesia as the PND group on the medical big data platform of our hospital.According to the type of surgery and age,the patients who did not have PND and were cured and discharged at the same time period were randomly matched as the Non-PND group with the ratio of 1: 3.Then,all the screened patients were randomly divided into training dataset and test dataset according to the ratio of 7: 3.All preoperative clinical data of patients were collected,and data pre-processing was performed.Then we employed a univariate analysis to discern differences between the PND and non-PND group,as well as between the training and testing datesets.The least absolute shrinkage and selection operator(LASSO)was used for feature selection,and three machine learning algorithms,namely,logistic regression(LR),support vector machine(SVM)and decision tree were used to establish predictive models of PND for the non-cardiac surgery.The model performance was evaluated by sensitivity,specificity,F1 score,and the area under the receiver operating characteristic(ROC)curve(AUC).2.A big data platform was composed of 49,768 surgical patients from three representative teaching hospitals in China(Southwest Hospital of Third Military Medical University,Xuan Wu Hospital of Capital Medical University,and West China Hospital of Sichuan University).Through the big data platform,the patients who underwent surgical treatment in general anesthesia and were diagnosed as PND after discharge were selected as the positive group.On the same big data platform,according to the type of surgery,age,and data source of the positive group,the patients who did not have PND and were cured and discharged in the same time period were randomly matched according to the ratio of 1: 3,as the negative group.Then the statistical differences were compared between two groups.At the same time,all patients data were divided into training dataset,test datasetand external validation dataset.After data preprocessing and feature selection,we inputted all features and features with statistical significance between two groups(P<0.05)to form three strategies based on the least absolute shrinkage and selection operator(LASSO),and then constructed predicting models respectively based on SVM,the generalized linear model(GLM),artificial neural network(ANN)and naive bayes(NB)using three strategies.The model performance was evaluated by AUC,the area under the precision recall curve(PRAUC),brier score,sensitivity,specificity,F1 score and decision curve analysis(DCA)curve.Finally,the best model was determined to provide a model basis for the next step of developing online visualization tool.3.Based on the model with the best performance,an online visualization tool,the web version of the dynamic nomogram,was developed using the Dynnom package in the R software and Shiny App.Result:1.368 patients enrolled in the study,including 92 patients in the PND group,276 patients in the Non-PND group.The training and testing datesets included 259 and 109 patients,respectively.The result of features selected by LASSO algorithm included age,aCCI,nutritional score,types of surgery,ASA status,emergency,Glu and PT.The prediction models established with these features based on LR,SVM and decision tree all had good prediction ability,and the SVM comprehensive model possessed the best performance.In the training dataset,SVM model the AUC of was 0.987(95% CI: 0.970-1.000),sensitivity and specificity were 98.5% and 99.0%,and F1 score was 0.977.In the test dataset,the AUC of was 0.957(95% CI: 0.905-1.000),sensitivity and specificity were 92.6% and 98.8%,and F1 score was0.943.2.Based on the medical big data platform of multi-center hospitals,1051 patients were included in the final analysis(PND group: 242,non-PND group: 809).Based on the performance of various predictive models,results of DCA curves,and the number of features analyzed,the GLM model with Strategy 3 demonstrated the most optimal performance and greater clinical applicability.In external validation group,the ROCAUC and PRAUC of GLM model with Strategy 3 were 0.874(95% CI,0.833-0.915)and 0.716(95% CI0.613-0.819),of which sensitivity and specificity were 72.6%(95% CI 61.4%-81.5%)and 84.4%(95% CI 79.3%-88.4%),and brier score was 0.131.3.Finally,we developed an online tool(https://pnd-predictive-model-dynnom.shinyapp s.io/Dyn Nomapp/)based on GLM model with Strategy 3,which included 15 general variables could early visual assessment the risk of PND.Conclusion:1.Our study found that age,aCCI,nutritional score,type of surgery,ASA status,emergency,Glu and PT were all important risk factors for PND in non-cardiac surgery.At the same time,a prediction model for PND in the non-cardiac surgery was successfully established based on machine learning using single-center hospital data,which SVM model had high prediction efficiency and the practical value of clinical application,representing the feasibility and efficiency of using machine learning for predicting PND.2.Among the prediction models constructed using multi-center hospital data based on three input strategies and four machine learning algorithms,the GLM model based on strategy 3 had the best performance and good predictive ability for PND evaluation,providing a good model foundation for the next step in building online visualization tool.3.We developed a simple and rapid online visualization tool based on multi-center hospital datasets for screening high-risk patients of PND to assiet medical staff in early screening,guiding perioperative management strategies and improving the prognosis of patients. |
PDF Full Text Request