Construction And External Validation Of Safe Discharge Criteria After Radical Gastrectomy

Research backgroundIn recent years,with the promotion of Enhanced Recovery After Surgery(ERAS),length of postoperative hospital stay falls,but the unplanned readmission rates has been significantly increased.This study intends to construct a model for predicting safe discharge through retrospective data analysis,and carry out prospective external validation to establish objective criteria for safe discharge after radical gastric cancer surgery to reduce the rate of readmission after discharge.PART I: Construction of safe discharge criteria after radical gastrectomy Background and purposeThe discharge criteria mentioned in the current ERAS study are different.The discharge criteria proposed by scholars in Europe and the United States,Japan and South Korea or domestic scholars are based on the individual experience of the operator,lack of evidence-based medical evidence,and no objective indicators.Therefore,the establishment of objective and effective discharge criteria is an urgent need for the promotion and application of ERAS.There are more studies on the high-risk factors of postoperative complications of gastric cancer,and the current high-risk factors found include obesity,smoking,chronic diseases,nutritional status,tumor location and stage,and so on.Although these highrisk factors are helpful to guide the management of gastric cancer perioperative period,how to use these high-risk factors to judge the complications after discharge has not been reported.MethodThis study retrospectively collected 735 cases of gastric cancer surgery in our hospital between March 2016 and March 2017.Exclusion criteria included history of gastric surgery,emergency surgery,failure to perform surgery or patient refusal to operate,ASA grade > III,combined with other malignant tumors or palliative surgery due to peritoneal dissemination or distant metastases,length of postoperative hospital stay < 5 days,complications of grade ≧III within 5 days after operation,failure to meet the design requirements of complete data collection and refusal of informed consent signed by patients.The remaining 496 cases were modeled and grouped into complication group and no complication group.Complications included gastrointestinal fistula,chylous fistula,abdominal infection,wound opening,deep venous thrombosis,gastroparesis and intestinal obstruction,and so on.Then,the independent influence factors of postoperative complications of gastric cancer were analyzed by logical regression group,and the predictive model of safe discharge after radical gastrectomy was constructed.The area under ROC curve(AUC),the optimal cutoff point of the model,the sensitivity,specificity,positive prediction rate,negative prediction rate and theirs 95% confidence interval were evaluated.Calibration curve,ROC curve,clinical decision curve and nomogram of prediction model were drawn.ResultThis study included 496 modeling cases,among 118(23.8%)developed complications after 5 days after surgery.The independent risk factors for complications analysized by logical regression included sex,autonomous activity time,maximum body temperature,and total intake on day 4 after surgery,neutrophil proportion ≥75%and pain score ≥4 points on day 5 after surgery,and whether to defecate in 5 day after surgery.A predictive model of risk of complications after 5 days of radical gastrectomy and the model nomogram were constructed based on the above 7 indexes.When the nomogram score was lower than 110,that is,the probability of the model predicting complications was lower than 0.196,the probability of no complications after the 5th day of surgery in this patient was 0.957(95% CI:0.928-0.976).The model predicted 322low-risk cases(64.9%),of which 308 had no complications.The comprehensive test of model coefficient showed that the fitting degree of the model is good(p <0.001)and the model had high stability(p =0.540).Model consistency assessment showed Kappa value was 0.598(95%CI:0.524-0.673,p =0.621),and Mc Nemar test value was0.113(95%CI:0.078-0.148,p <0.0001).The area under the ROC curve(AUC)was0.918(95%CI:0.891-0.941),sensitivity was 0.881(95%CI:0.809-0.934),specificity was0.815(95%CI:0.772-0.853),and accuracy was 0.831(95%CI:0.795-0.863).The clinical decision curve showed that the clinical benefit of the predictive model was significant.ConclusionIn this study,a logical regression was used to successfully construct a risk prediction model for the risk of complications after day 5 of radical gastrectomy.Using the 7 indexes including sex,autonomous activity time,maximum body temperature,and total intake on day 4 after surgery,neutrophil proportion ≥75% and pain score ≥4 points on day 5 after surgery,and whether to defecate in 5 day after surgery,cases of possible complications after 5 day of surgery can be screened out,the negative prediction rate was 95.7%,the positive prediction rate was 59.8%.Whether the prediction model can be used as a safe discharge criteria on the 5th day after radical gastrectomy needs further external case validation.PART II: External validation of safe discharge criteria after radical gastrectomy Background and purposeThe discharge criteria is the independent risk factor of the readmission rate after discharge.When will the patient be discharged safely after radical gastrectomy,and there are no complications after discharge,there is still a lack of objective scoring criteria.The first part of the study successfully constructed a predictive model of postoperative complications after the fifth day of radical gastrectomy.The aim of this study is to evaluate the performance of the predictive models previously constructed through prospective external validation.MethodThis study prospectively collected 526 cases of gastric cancer surgery admitted in our hospital between September 2018 and March 2019.After excluding the substandard cases,the remaining 245 cases were taken as the validation case set.Based on the prediction model established in the previous stage,the probability of complications after five days of operation of validation cases were calculated.Then the validation cases were divided into low-risk group and high-risk group,and the accuracy of the prediction of the two groups were calculated separately,and the quality indicators of the model were evaluated.ResultThis study included 245 cases as validation set,of which 46(18.5%)had complications after 5 days after operation.A logical regression prediction model based on the above seven indicators was used for prospective external validation of validation cases.The results showed that the model predicted 119 cases of low-risk cases(48.6%),of which 109 had no complications,the negative prediction rate was 0.916(95%CI:0.851-0.959),and 126 cases of high-risk cases(51.4%),of which 36 had complications and the positive prediction rate was 0.286(95% CI:0.209-0.373).Model consistency assessment showed Kappa value was 0.198(95%CI:0.105-0.291,p=0.258)and Mc Nemar test value was 0.327(95% CI:0.258-0.395,p<0.0001).The area under the ROC curve(AUC)was 0.719(95%CI:0.645-0.794),sensitivity was 0.783(95%CI:0.636-0.891),specificity was 0.548(95% CI:0.476-0.618),and accuracy was 0.592(95%CI:0.527-0.654).The clinical decision curve showed that the clinical benefit of the predictive model was significant when the predict risk of postoperative complications below 0.36.ConclusionIn this study,the prospective external validation for the predictive model for the risk of complications after 5 days of radical gastrectomy constructed by logical regression method was performed.The results showed that the predictive accuracy of the model for no complications after the fifth day of surgery was 91.6%,and the positive predictive accuracy was 28.6%.The risk of complications in the predictive model can be used as a safe discharge criteria for the 5th day after radical gastrectomy.PART III: Construction and External Validation of Safe Discharge Criteria afterradical gastrectomy with Machine Learning XGBoost Algorithm Background and purposeMachine learning is a class of computer algorithms developed for accurate prediction.Comparative studies of predictive ability between machine learning and logical regression shows that in most cases machine learning has better ability to construct predictive models.Although the risk prediction model of complications after the 5th day of radical gastrectomy was successfully constructed by logical regression method in the previous study of this subject,it is not clear whether the machine learning technology can be used to construct a simpler and more optimized prediction model.The purpose of this study is to build a model by using the machine learning XGBoost algorithm,then compare with the logical regression model to find which is better.MethodIn this study,496 enrolled cases collected in the first part of this subject were taken as the modeling case set,and 245 enrolled cases collected in the second part of this subject were taken as the validation case set.Based on the seven independent risk factors of postoperative complications of radical gastrectomy screened in the first part of this subject,the machine learning XGBoost algorithm was used to reconstruct the prediction model of complications after the 5th day of radical gastrectomy,evaluate the quality indicators,and compared with the logical regression model.ResultIn this study,the machine learning XGBoost algorithm is used to analyze the importance of the seven predictive indexes of the logical regression prediction model,and the results showed that the three of these indicators could replace the above seven indexes.The three indicators were time of autonomous activity,total intake and maximum body temperature on the fourth day after radical gastrectomy.When the probability of predicting complications in the machine learning model was lower than 0.268,the patient was a low-risk case,and the probability of no complications after the 5th day of surgery was 0.961(95%CI:0.936-0.979).The MLmodel predicted 364 low-risk cases(73.4%),of which 350 had no complications,and132 had been predicted for high-risk cases(26.6%),of which 104 had complications and a positive prediction rate was 0.788(95%CI:0.708-0.854).Prospective external validation showed that the model predicted 143 low-risk cases(58.4%),of which 125 had no complications,with a negative predictive rate of 0.874(95% CI:0.808-0.924),and a high-risk case of 102(41.6%),of which 28 had complications and a positive predictive rate of 0.275(95%CI:0.191-0.372).In the modeling case set,the AUC of machine learning prediction model was significantly higher than the logical regression model(0.949 vs 0.918,p =0.004),and the predictive ability was improved(NRI=0.111,p=0.0001),with more significant clinical benefits.In the validation case set,the AUC of machine learning prediction model was lower than the logical regression model(0.646 vs 0.719,p=0.051),and the predictive ability decreased(NRI=-0.140,p=0.093),but the clinical benefit of high-risk cases was more pronounced.ConclusionIn this study,the machine learning XGBoost algorithm was used to successfully construct the risk prediction model of complications after the 5th day of radical gastrectomy.Using the highest body temperature,total intake and autonomous activity time of the fourth day after operation,the cases of possible complications after the fifth day of operation could be screened out,and the negative prediction rate was 96.1%,and the accuracy rate of positive prediction was 78.8%.The prediction model constructed by machine learning XGBoost algorithm was better than the model constructed by logical regression.