| Objectives:Postoperative sepsis is a life-threatening organ dysfunction caused by dysregulated host response to infection,occurring after a surgical procedure or during the postoperative hospital stay,whose fatality rate is very high,and may affect the long-term outcome and bring high economic burden.Due to the long procedure time,bacterial translocation,and immune deficiency,abdominal surgery is one of the most susceptible surgical types to postoperative sepsis.Preoperative screening of high-risk patients,early prediction of mortality risk,and timely and reasonable perioperative management may improve outcome.However,the diagnosis and management of sepsis is badly complex but not effective at present.Therefore,new technologies are needed to achieve early screening diagnosis,treatment and prognosis prediction.In this study,we aimed at postoperative sepsis in patients undergoing abdominal surgery,developed a method to screening high-risk patients preoperatively based on machine learning algorithms with routine variables on a multicenter dataset from three representative academic hospitals in China,and to investigate the associated important variables.In addition,the Medical Information Mart for Intensive Care IV(MIMIC-IV)was used to establish a strategy of unfractionated heparin based on reinforcement learning.Meanwhile,models for predicting the risk of death were constructed based on machine learning algorithms on this public dataset.Methods:1.We retrieved patients who diagnosed with postoperative sepsis after abdominal surgery from three representative academic hospitals(Southwest Hospital of Army Medical University,West China Hospital of Sichuan University,Xuan Wu Hospital of Capital Medical University)in China during May 2014 to January 2020 as positive cases,then negative cases were randomly extracted from the whole dataset according to the age and surgical types of the positive cases with the ratio of 1:2.Relevant clinical variables were extracted and pre-processed,analyzed and compared between two groups.Moreover,the dataset was randomly split into training(70%)and test(30%)datasets according to whether diagnosed with postoperative sepsis or not.On the basis of the result of feature selection by Boruta algorithm,e Xtreme Gradient Boosting(XGboost),Multi-Layer Perceptron(MLP),Logistic Regression(LR)and K-Nearest Neighbor(KNN)were used to develop prediction model for the onset of postoperative sepsis.The area under the receiver operating characteristic curve(AUC),sensitivity,specificity,F1 score,accuracy,Positive Predictive Value(PPV)and Negative Predictive Value(NPV)were used for model evaluation.2.After obtaining the permission of MIMIC-IV,we collected patients who diagnosed with postoperative sepsis and received unfractionated heparin after abdominal surgery as positive cases,and those who did not received unfractionated heparin were taken as negative cases.Relevant clinical variables were extracted and pre-processed,analyzed and compared between two groups.Meanwhile,the reinforcement learning strategy of unfractionated heparin for sepsis after abdominal surgery was constructed based on Q-learning algorithm for positive cases,and the following three methods were used to evaluate the strategy: 1)comparing the strategy value of reinforcement learning and clinician;2)the coincidence of clinician strategy and reinforcement learning strategy;3)the relationship between relative benefit and survival rate.3.Patients who diagnosed with postoperative sepsis after abdominal surgery and died within 90 days of admission in MIMIC-IV were collected as positive cases,and those who survived were regarded as negative cases.Relevant clinical variables were extracted and preprocessed,analyzed and compared between two groups.Moreover,the dataset was randomly split into training(70%)and test(30%)datasets according to whether died or not in 90 days.On the basis of the result of feature selection by LASSO algorithm,LR,Random Forest(RF),Gradient Boosting Decision Tree(GBDT),Adaptive Boosting(Ada Boost)and Support Vector Machine(SVM)were used for model fitting to predict the mortality risk.The AUC,sensitivity,specificity,PPV,NPV,accuracy and F1 score were used for model evaluation.Result:1.A total of 648 patients undergoing abdominal surgery from the three hospitals were finally analyzed(sepsis:212,non-sepsis:436).Models based on XGboost,MLP,KNN and LR all performed well,while XGboost was the best among them: In the training dataset,yielded the AUC of 0.958(95%CI: 0.937～0.980),accuracy 95.8%,sensitivity 93.3%,specificity 98.4%,PPV 96.5%,NPV 96.8%,and F1 score 0.949;In the test dataset,yielded the AUC of 0.839(95%CI: 0.783～0.896),accuracy 83.9%,sensitivity 81.0%,specificity 86.9%,PPV 75.0%,NPV 90.4%,and F1 score 0.779.Albumin,laparoscopic procedure,Neutrophil-Lymphocyte ratio,bilirubin,creatinine and platelet were the top features associated with postoperative sepsis.2.A total of 986 patients were diagnosed with postoperative sepsis in MIMIC-IV,of which560 patients received unfractionated heparin after surgery.Compared with patients who did not receive unfractionated heparin after surgery,those in the unfractionated heparin group had longer length of stay(13.10 days vs.8.72 days,P<0.001),but lower 90-day mortality(22.5%vs.29.3%,P=0.015).And there was no significant difference in general characteristics,blood product infusion,coagulopathy,initial fluid resuscitation and vasopressors between two groups.Moreover,the reinforcement learning strategy of unfractionated heparin for sepsis demonstrated that the strategy value was generally higher than that of clinician,and the higher the coincidence of clinician strategy and AI strategy,or the greater the relative benefit in the treatment process,the higher the survival rate of patients.3.A total of 986 patients were diagnosed with postoperative sepsis in MIMIC-IV,of which251 patients died within 90 days of admission.In the test dataset,the AUCs of LR,GBDT,RF,SVM and Ada Boost models were 0.852(95%CI: 0.799～0.905),0.903(95%CI: 0.856～0.950),0.921(95%CI: 0.888～0.955),0.940(95% CI: 0.909～0.972)and 0.906(95% CI: 0.868～0.944),sensitivities were 66.2%,78.7%,62.7%,76.0% and 66.7%,and specificities were 92.0%,91.4%,94.6%,95.5% and 90.9%,PPVs were 74.6%,75.6%,79.7%,85.1% and 71.4%,NPVs were 88.5%,92.6%,88.1%,92.1% and 88.9%,F1 scores were 0.702,0.771,0.702,0.803 and0.690 and accuracies were 79.1%,85.0%,78.6%,85.7% and 78.8%,respectively.Overall models based on these four ensemble learning algorithms manifested better than the traditional LR,and SVM was the best among them.Conclusion:1.XGboost is a new and potential generalizable model to preoperatively screening the high-risk patients of sepsis after abdominal surgery with the advantages of representative training cohort and routine variables.Meanwhile,low albumin,laparoscopy surgery,high neutrophil-lymphocyte ratio,high bilirubin,high creatinine and low platelet were considered as important indicators,suggesting doctors should pay more attention on them preoperatively.2.The reinforcement learning model optimal unfractionated heparin management for patients with sepsis after abdominal surgery,proved to be able to achieve greater benefits than clinician,but still needing improvement.Meanwhile,patients with postoperative sepsis received unfractionated heparin have lower 90-day mortality without increasing risk of blood product transfusion or coagulopathy was discovered in this study.3.Based on ensemble learning of GBDT,RF,SVM and Ada Boost to build mortality prediction model for sepsis after abdominal surgery was practicable,with the performances all superior to the traditional LR model,while SVM model performed most satisfyingly. |
PDF Full Text Request