Prediction Of Infected Pancreatic Necrosis Based On Decision Tree Model

Research background and purpose: Necrotizing pancreatitis(NP)occurs in about 15-20% of patients with acute pancreatitis.The incidence rate of NP secondary infection is related to the increase of mortality.Few studies provide a simple and practical prediction for the occurrence of Infected Pancreatic Necrosis(IPN).Therefore,it is necessary to determine the best predictors of IPN,which is the key to improve the survival rate of patients.Data mining refers to the process of finding some hidden information in a large number of data through a certain rule,and these information has a special connection.Decision tree is a very important research method in data mining technology,which is a common classification method.In essence,it is the process of data classification through some rules.In machine learning,decision tree is equivalent to a prediction model,which is a graphic method directly using probability analysis.Classification and Regression tree(CART)is a classical algorithm commonly used in decision trees,which has been widely used in many fields.Literature research shows that CART algorithm has not been used in IPN decision-making model.The purpose of this study is to build a decision tree model to predict IPN through CART algorithm,and verify the performance of the model.Materials and Methods: This study retrospectively analyzed the clinical data of necrotizing pancreatitis(NP)patients from January 2012 to December 2018 in the database of general surgery center of Western Theater general hospital.224 necrotizing pancreatitis patients were randomly divided into training samples(n = 156)and test samples(n = 68)according to the ratio of 7:3.The data of patients' laboratory is standardized,and the decision tree model is constructed by using the sklearn toolkit in Python and cart algorithm.Training samples are used to build the model,test samples are used to verify the built model,and hierarchical 10 fold cross validation is used to evaluate the prediction ability of the model.AUC(area under receiver operating characteristic)was constructed to evaluate the decision tree model.Result:1.224 cases of necrotizing pancreatitis were included in this study.The prevalence of IPN was 10.7% in the whole study population,including 129 males(57.6%),95 females(42.4%)and a median age of 49 years(41.25-61).Among all the patients,85(37.9%)had biliary diseases,55(24.6%)had hyperlipidemia,34(15.2%)had alcoholic diseases,and 50(22.3%)had other causes,such as post ERCP,abnormal anatomy and idiopathic pancreatitis.2.By using Python programming language and cart algorithm as decision tree growth rule.By comparing the accuracy of different depth of decision tree,the accuracy of the model is the highest when the depth is 5 layers.Finally,in the decision tree model established by training samples,the number of nodes is 17,the number of terminal nodes is 9,and the depth of decision tree is 5.Eight important variables,such as serum amylase(AMY),CO2 CP,serum cholinesterase(Ch E),plasma osmolality,platelet distribution width(PDW),BE,serum Cystatin C and serum Ig G4,were screened.Their cut-off values were 14.12mmol/L?32.67mmol/L?1.69 U/m L?275.42Osm/kg H2O?25.02%?2.09mmol/L?1.48mg/L?0.15g/L,respectively.According to the possibility of developing into IPN,the terminal nodes of decision tree model are reconstructed into two groups.There are five groups in high-risk group and four groups in low-risk group.3.When the decision tree model is applied to test samples,the accuracy of the decision tree is 88.3%(95% CI,0.79-0.95),the sensitivity and specificity are 42.9%(95% CI,0.12-0.80)and 93.4%(95% CI,0.83-0.98),respectively.The false positive rate is 6.5%(95% CI,0.02-0.17)and the false negative rate is 57.1%(95% CI,0.20-0.88).The positive predictive value is 42.9%(95% CI,0.12-0.80)and the negative predictive value is 93.4%(95% CI,0.83-0.98).The 10 fold cross validation method was used to evaluate the prediction ability of the model,with an average accuracy of 88.4%.According to ROC curve analysis,AUC is 0.69(95% CI,0.46 – 0.91).Conclusion: In this study,we found that serum amylase(AMY),CO2 CP,serum cholinesterase(Ch E),plasma osmolality,platelet distribution width(PDW),BE,serum Cystatin C and serum Ig G4 can be used as important node indicators to predict IPN,and further reconstruct and comb the terminal nodes of the decision tree model according to IPN risk assessment value,and get the high-risk factors of five groups of IPN.Through a series of model validation and evaluation,it is confirmed that these five groups of high-risk predictors can effectively predict the occurrence of IPN.The results of this study through the decision tree model found that there is a potential interaction between the single variable indicators,and obtained the multi variable binary prediction factors,high clinical practicability,which is conducive to the accurate prediction and evaluation of IPN,has a certain guiding significance for the formulation of clinical treatment plan,and also provides a new direction for the in-depth study of IPN.