Prediction Of Postoperative Risk Of Death In Patients Undergoing Abdominal Surgery Based On Machine Learning

Objectives:As the amount of surgery increases year by year,perioperative safety of patients has become a focus of attention.According to the 2018 China Health Yearbook,the annual surgical volume in China is about 60 million,and perioperative mortality of surgical patients is as high as 2%,up to 12% in patients combined with basic diseases.Accurate preoperative risk assessment is critical for early identification of high-risk patients,early intervention,and supporting for clinical decision making to better protect patients' perioperative safety.However,preoperative evaluation methods widely used in clinical practice are mostly various rating scales based on traditional statistical methods,which are highly subjective,time-consuming and laborious to be evaluated,especially they are difficult to meet with continuous dynamic evaluation required by high-risk patients.Therefore,as the patients' conditions become more complex and the clinician's workload becomes heavier and heavier,a fast,accurate,and objective prediction method is urgently needed to evaluate and predict the perioperative risk of patients,to reduce perioperative mortality,and to protect patients' perioperative safety.Therefore,we combined with Chongqing Institute of Green and Intelligent Technology,Chinese Academy of Sciences in this study,and intended to retrospectively analyze preoperative risk factors for patients who died after undergoing abdominal surgery,and to use machine learning algorithms to establish a risk prediction model for predicting the risk of death in patients undergoing abdominal surgery.The study aimed to explore new methods for preoperative risk assessment.Methods:1.We retrieved 158 patients who died in the hospital after abdominal surgery under general anesthesia from January 2007 to December 2018 via using the hospital's big data research platform as positive cases.According to the age group and operations' name and types of positive cases',we randomly matched patients who did not die after abdominal surgery as negative control cases during the same surgical period based on 1: 3 according to statistical requirements.Then we collect preoperative clinical data of all cases,analyze and compare the indicators between the positive death group and the negative survival group,and incorporate the statistically significant indicators(P <0.05)in the univariate analysis into the multivariate logistic regression analysis to screen out risk factors related to death after surgery and a risk regression equation was established to draw receiver operating characteristic curves(ROC curves).2.We utilized six machine learning algorithms,including Adaptive Boosting(Ada Boost),Extreme Random Trees(ET),Gradient Boosting Decision Tree(GBDT),Linear Support Vector Machine(Linear SVM),Logistic Regression,and Random Forest(RF)to select features for inclusion indicators,based on inclusion of 90% important indicators(F?90),95% important indicators(F?95)and all non-zero importance indicators(F?100)to construct risk prediction models for postoperative death of patients undergoing abdominal surgery and evaluate the effect,calculate and compare the area under the ROC curve(AUC),F1 score,and Brier score,positive predictive value,sensitivity,and specificity.3.We calculated probability characteristics through the collection of six single traditional machine learning algorithms: Ada Boost,ET,GBDT,Linear SVM,LR,and RF,and then using LR(This section combineed multiple machine learning algorithms for integrated learning,named " Postoperative Mortality Prediction(PMP method)based on three strategies including 90% important indicators(F?90),95% important indicators(F?95),and all non-zero importance indicators(F?100)for constructing evaluation of postoperative mortality risk prediction models for patients after undergoing abdominal surgery,calculating and comparing the AUC,F1 score,Brier score,positive predictive value,sensitivity and specificity among different models.Result:1.Multivariate Logistic regression analysis showed that the independent risk factors associated with postoperative mortality in patients undergoing abdominal surgery included: American Society of Anesthesiologists(ASA)classification [Odds Ratio(OR): 2.117,95% Confidence Interval(CI): 1.219 ? 3.679,P = 0.008],preoperative hypertension(OR: 2.072,95% CI: 1.094 ? 3.926,P = 0.025),preoperative coronary heart disease(OR: 3.944,95% CI: 1.433 ? 10.857,P = 0.008),preoperative lung disease(OR: 2.678,95% CI: 1.214 ? 5.907,P = 0.015),Activated partial thromboplastin time(APTT)(OR: 1.070,95% CI: 1.028 ? 1.114,P = 0.001),Aspartate aminotransferase(AST)concentration(OR: 1.007,95% CI: 1.000 ? 1.014,P = 0.044),Total bilirubin(TBIL)concentration(OR: 1.003,95% CI: 1.000 ? 1.007,P = 0.045),blood Glucose concentration(OR: 1.109,95% CI: 1.006 ? 1.222,P = 0.037).2.The indicators which were selected by feature selection in machine learning,in the same strategy on different machine learning algorithms > 3 times,and included by the three strategies were: Hb concentration,TBIL,Fbg,APTT,preoperative pulmonary disease,preoperative hypertension,the WBC count,pulse,propagated,NEUT count,serum sodium concentration,AST,preoperative combined renal insufficiency,A/G,PLT count,the number of LYMPH,ALT,BMI,Cre,blood calcium concentration.3.Among the researches using single machine learning algorithms,RF comprehensive evaluation performance was the best in the prediction models based on F?90(AUC: 0.816,95%CI: 0.761?0.871);F1 score: 0.77;Brier score: 0.144),and GBDT comprehensive evaluation performed best in the prediction models based on F?95 and F?100(AUC: 0.854,95%CI: 0.803?0.904);F1 score: 0.81;Brier score: 0.131),(AUC: 0.833,95%CI: 0.780?0.886);F1 score: 0.80;Brier score: 0.144).4.Among the studies on death risk prediction calculated by PMP method,the comprehensive evaluation of the prediction models based on F?95 was the best(AUC: 0.842,95%CI: 0.791?0.894);F1 score: 0.82;Brier score: 0.142).The PMP's performance is higher than the prediction model established by single machine learning algorithm.Conclusion:1.Independent risk factors for postoperative mortality in patients undergoing abdominal surgery included ASA grade,preoperative hypertension,preoperative coronary heart disease,preoperative pulmonary disease,APTT,AST,TBIL,and blood glucose concentration.2."Hb concentration,TBIL,Fbg,APTT,preoperative pulmonary disease,preoperative hypertension,WBC count,pulse,ALB,NEUT count,serum sodium concentration,AST,preoperative renal insufficiency,A/G,PLT count,LYMPH count,ALT,BMI,Cre,blood calcium concentration" are effective indicators for predicting the risk of death after abdominal surgery.3.The prediction model of death risk after abdominal surgery based on machine learning algorithm has good clinical value;The predictive efficiency of the integrated algorithm is better than that of the single traditional machine learning algorithms.