| Objective:Gastrointestinal tract is a high occurrence area for malignant tumors,and radical surgical resection is the main treatment for gastrointestinal cancer.Anastomotic leakage is a serious and fatal complication during the postoperative period.A more accurate risk stratification of anastomotic leakage and the study for new predictive factors may help surgeons make more reasonable intraoperative decisions(such as temporary ileostomy in rectal cancer surgery)and improve postoperative management.The purpose of this study is to develop a prediction model for anastomotic leakage in rectal cancer based on artificial intelligence and analyze the feasibility of the model in assisting in temporary ileostomy,investigate the abdominal composition related to anastomotic leakage and develop a predictive model for anastomotic leakage of proximal and total gastrectomy.Methods:This study is a single center retrospective study,data from patients with rectal cancer who received anal preservation surgery and patients with gastric cancer who received total or proximal gastrectomy was collected.The data set was divided into training set and test set.The training set was used to train several artificial intelligence algorithms,such as:logical regression,support vector machine,random forest,gradient boosting,Naive Bayes,and eXtreme gradient boosting algorithms.The test set was used to validate the model(s).Over-sampling,down-sampling technology and cross validation was adopted to train high performance models.The operating characteristic curve,area under the curve,sensitivity,specificity,positive predictive value,negative predictive value and accuracy were used to assess the performance of the models.Feature importance analysis was used to evaluate the importance of each predictive factor.The comparison between the proportion of patients with high risk of anastomotic leakage predicted by artificial intelligence model and the actual temporary ileostomy rate represented the ability of the model to assist decisionmaking of temporary ileostomy.The study of relationship between abdominal composition and anastomotic leakage in rectal surgery was conducted using a case-control study(1:1).The abdominal composition related indicators were calculated based on the computer tomography images of the lower edge of the third lumbar spine,including abdominal circumference,subcutaneous fat thickness,subcutaneous fat area,skeletal muscle area,skeletal muscle index,abdominal anterior to posterior diameter and transverse diameter of abdominal cavity.Univariate and multivariate logistic regression was used to analyze the independent risk factors.Results:2240 cases of patients with rectal cancer and 1660 cases of patients with gastric cancer who received proximal or total gastrectomy were included in this study.The incidence rate of anastomotic leakage after rectal cancer surgery was 5.40%(121/2240),and the incidence rate of anastomotic leakage after proximal and total gastrectomy was 2.17%(36/1660).Among rectal cancer patients,five artificial intelligence models showed good predictive performance in the training set,among which the support vector machine model performed better and more explanatory.More importantly,the support vector machine model was able to make more reasonable decisions of temporary ileostomy compared to surgeons,and outperformed the reported scoring model or ASCRS guideline.In patients with gastric cancer,four artificial intelligence models showed good predictive ability,and the random forest model showed more advantages in predicting anastomotic leakage for proximal and total gastrectomy.In patients with same gender and similar body mass index,high visceral fat content and a narrower abdominal structure were associated with anastomotic leakage.Conclusion:The artificial intelligence model developed in this study could predict anastomotic leakage favorably in rectal cancer surgery,total and proximal gastrectomy.More importantly,the model for rectal cancer anastomotic leakage may be helpful in assisting the surgical decision-making of temporary ileostomy and which outperformed to the experience of surgeons.This is expected to provide a tool for the selective implementation of temporary ileostomy.In addition,this study also found that high visceral fat content and narrow abdominal structure may be related to the occurrence of anastomotic leakage after rectal cancer surgery.Part Ⅰ:Artificial Intelligence Assists Surgeons’ Decision-making of Temporary IleostomyObjective:Anastomotic leakage is a serious life-threatening complication after radical resection of rectal cancer.Temporary ileostomy is introduced to reduce the serious clinical outcomes followed by anastomotic leakage.However,at present,the intraoperative decision-making of temporary ileostomy still depends on the experience of surgeons,and there is no practical clinical guideline to follow,which leads to the rate of temporary ileostomy is far higher than the incidence of anastomotic leakage.Therefore,this part of the study aims to develop an artificial intelligence model for predicting anastomotic leakage of rectal cancer surgery and thus assisting in decision-making of temporary ileostomy.Methods:This part of the study is a single-center retrospective research,the patients who received rectal cancer surgery from July 2010 to December 2020 in our center were included.The preoperative and intraoperative variables reported as risk factors for anastomotic leakage by previous literature and which were available in our center were collected.The patients were divided into training set(2010-2016),test set 1(2017-2018)and test set 2(2019-2020)using a temporal sampling strategy,and the training set is used for training artificial intelligence algorithms,including linear support vector machine,logical regression,Naive Bayes,gradient boosting and random forest.Test set 1 and 2 were used for model validation.Area under the curve,sensitivity,specificity,positive predictive value,negative predictive value and accuracy were calculated to assess the performance of models.The contribution of the each included variable to the prediction results is analyzed by feature importance analysis.The ratio of patients with high risk of anastomotic leakage predicted by the model was compared with the actual temporary ileostomy rate to represent the ability of the model for decision-making.Results:A total of 2240 patients were included in the present study,including 1145 in training set,550 in test set 1 and 545 in test set 2.The overall incidence of anastomotic leakage was 5.4%,including 6.4%in training set,4.7%in test set 1 and 4.0%in test set 2,while the rates of temporary ileostomy in training set,test set 1 and test set 2 were 19.2%,31.8%and 31.2%,respectively.The five artificial intelligence models trained by the training set achieved satisfactory prediction performance,and the support vector machine model was finally selected for further analysis.The analysis of assisting the decision-making of temporary ileostomy demonstrated that the overall rate of temporary ileostomy would have be significantly reduced in test cohorts 1 and 2,and the implementation rate of temporary ileostomy in patients with anastomotic leakage would have be significantly increased(P<0.05).Feature importance analysis showed that the operation time and the distance from the tumor to the anal margin were the most important two variables.In order to address the application of the model in clinics and future investigations,a web app(https://alrisk.21cloudbox.com/)was developed to allow for real-time prediction.This will enable a surgeon to perform accurate intraoperative decision-making.Conclusions:The support vector machine model developed in this part performed favorable in predicting anastomotic leakage in patients who received radical resection of rectal cancer,and it may be helpful in adopting temporary ileostomy more reasonably.Part Ⅱ:Comprehensive Abdominal Composition Evaluation of Rectal Cancer Patients with Anastomotic LeakageObjective:BMI was reported to be related to the occurrence of anastomotic leakage,but patients with similar BMI may have different risks of anastomotic leakage.The abdominal composition was associated with several postoperative complications in surgical operations.However,it is still unclear whether the abdominal composition is different between patients with and without anastomotic leakage when those had similar BMI.Due to BMI cannot distinguish muscle,fat,and their distribution,therefore,this part of the study aims to investigate the differences in abdominal composition between patients with anastomotic leakage and those without anastomotic leakage through matching BMI.On the one hand,it may provide new insight for the occurrence of anastomotic leakage.On the other hand,it may provide new factors for the prediction of anastomotic leakage.Methods:This part of the study is a single-center case-control study,the data from patients who received rectal cancer surgery from January 2011 to September 2020 in the Department of Gastrointestinal Surgery were reviewed.Among them,patients with anastomotic leakage and complete data were defined as the anastomotic leakage group,and patients with the same sex and BMI±1kg/m2 matching from patients without anastomotic leakage were defined as the control group.Preoperative CT images(the lower edge of the third lumbar spine)were used to measure the abdominal composition related indicators,such as abdominal circumference,subcutaneous fat thickness,subcutaneous fat area,skeletal muscle area,skeletal muscle index,visceral fat area,anterior to posterior diameter of abdominal cavity and transverse diameter of abdominal cavity.Feature importance analysis was used to quantify the contribution of each abdominal composition related indicator to anastomotic leakage.Univariate and multivariate logistic regression analysis was used to analyze the independent factors associated with anastomotic leakage.Results:A total of 2065 anal-preservation operations were performed from 2011 to 2020 in our center and 107 cases developed anastomotic leakage(5.18%).Using a 1:1 matching strategy,156 cases were enrolled,including 78 cases of the anastomotic leakage group and 78 cases of the control group.The comparison between the control group and the anastomotic leakage group showed that the patients with anastomotic leakage had a longer operation time,lower preoperative hemoglobin level,lower preoperative albumin level,larger tumor size and later tumor stage(P<0.05).In addition,the comparison of abdominal composition showed that the patients with anastomotic leakage had a larger visceral fat area,smaller anterior to posterior diameter of abdominal cavity and smaller transverse diameter of abdominal cavity(P<0.05).Feature importance analysis indicated that visceral fat area,anterior to posterior diameter of abdominal cavity and transverse diameter of abdominal cavity were the three most important variables of predicting anastomotic leakage.Univariate and multivariate logistic regression analysis showed that preoperative albumin level,tumor size,tumor stage,visceral fat area,anterior to posterior diameter of abdominal cavity and transverse diameter of abdominal cavity were independent risk factors for anastomotic leakage(P<0.05).Conclusions:Compared with rectal cancer patients without anastomotic leakage matching with BMI and sex,patients with anastomotic leakage may have greater visceral fat content and narrower abdominal structure.Part Ⅲ:Application of Artificial Intelligence for Predicting Anastomotic Leakage in Patients with Gastric Adenocarcinoma Who Received Total or Proximal GastrectomyObjective:Gastric cancer is the most common malignant tumor in the upper digestive tract and radical resection is main treatment.Anastomotic leakage after gastrectomy is still a serious life-threatening complication.The incidence of anastomotic leakage is higher in the proximal or total gastrectomy than which in distal gastrectomy.This part of the study focuses on prediction of anastomotic leakage in proximal or total gastrectomy and which may be helpful for intraoperative decision-making and optimization of perioperative management.However,there is no high-performance model,therefore,this part of study aims to use artificial intelligence to develop a model for predicting anastomotic leakage in proximal and total gastrectomy.Methods:This part of the study is a single-center retrospective study,including patients who received proximal or total gastrectomy in our center from January 2011 to September 2020 were included.The preoperative and intraoperative variables of each patient were collected so as to train a predictive model for identifying the risk of anastomotic leakage during surgery.The dataset was randomly divided into training set and validation set using a ratio of 8:2.The training set was used to train four artificial intelligence algorithms,such as logical regression,random forest,support vector machine and eXtreme gradient boosting.The validation set was used to validate the performance of model(s).The area under the curve,sensitivity,specificity,positive predictive value,negative predictive value and accuracy were used to evaluate the ability of the model in the validation set.The importance of the each variable of the prediction is calculated by the feature importance analysis.Results:A total of 1660 patients with proximal or total gastrectomy were enrolled in the present part of study,including 1338 in the training set and 332 in the validation set.The overall incidence of anastomotic leakage was 2.17%,1.9%in the training set and 3.3%in the validation set.The random forest model and the eXtreme gradient boosting model performed with larger area under the curves(0.92 and 0.89),while the logistic regression model and the support vector machine model showed average performance(0.84 and 0.82).Further comparison showed that the random forest model has higher specificity(0.822,95%CI:0.775-0.862 vs 0.701,95%CI:0.647-0.750,P<0.001)and accuracy(0.822,95%CI:0.776-0.861 vs 0.708,95%CI:0.656-0.756,P<0.001)than the support vector machine model.Similarly,the random forest model indicated higher specificity(0.822,95%CI:0.775-0.862 vs 0.723,95%CI:0.670-0.770,P=0.003)and accuracy(0.822,95%CI:0.776-0.861 vs 0.729,95%CI:0.678-0.776,P=0.004)than the eXtreme gradient boosting model.The analysis of the importance of each variable in the random forest model found that hypertension,diabetes,BMI,smoking index,albumin,hemoglobin,tumor size,tumor obstruction,ASA score and operation time were the ten most important variables.Conclusion:The random forest model developed in the present part of study(https://gasal.21cloudbox.com)can predict anastomotic leakage of proximal or total gastrectomy in a favorable performance. |
PDF Full Text Request