Development And Validation Artificial Intelligence Predictive Models For Complications Of Acute Pancreatitis

Acute pancreatitis(AP)is one of the common critical illness in digestive system,and its incidence has been increasing in recent years.Severe acute pancreatitis(SAP)is a serious disease with high mortality.Persistent organ failure and infected pancreatic necrosis are the two most important factors that determine the prognosis of AP.Compared with SAP,moderately severe acute pancreatitis(MSAP)has a better prognosis.If we can identify SAP and predict the risk of severe complications as soon as possible,it will be helpful to formulate a better treatment strategy and improve the prognosis.The clinicians could strengthen the monitoring of highrisk patients,control systemic inflammatory response syndrome(SIRS)and protect organs such as heart,lung and kidney to prevent the occurrence of serious complications.At present,the most common parameter of evaluating the severity and prognosis of AP is C-reactive protein(CRP).The scoring systems includes Acute Physiology and Chronic Health Evaluation Ⅱ(APACHE Ⅱ)score,Ranson score,bedside index for severity in acute pancreatitis(BISAP)and CT severity index(CTSI)and so on.These evaluation parameters or systems are helpful to predict the severity,complications or prognosis of AP,however,they still have some limitations,such as poor timeliness and accuracy,tedious operation,and so on.Therefore,clinicians need a more timely,accurate and simple model to predict the complications and prognosis of AP.In recent years,some researchers have gradually realized that coagulation dysfunction plays an important role in the pathophysiology of AP.Activation of coagulation system and inflammation are two mutually reinforcing processes.Therefore,these researchers have proposed that coagulation parameters can predict the severity,complications or prognosis of AP.For example,D-dimer had high sensitivity and specificity for predicting SAP.However,the efficacy of coagulation parameters for predicting SAP or complications needs to be demonstrated by high quality clinical studies.Compared to coagulogram,thromboelastogram(TEG)which is an accurate and rapid coagulation monitoring tool,can reflect the whole coagulation state.The R value,K value and MA value in TEG reflect the activity of coagulation factors,the level of fibrinogen,and the amount and function of platelets,respectively.TEG is a better tool than coagulogram in qualitative and quantitative assessment of coagulation status,however,there are few studies about TEG for predicting the complications of AP.Because of the machine learning algorithms’ powerful capabilities of processing and analysing data in artificial intelligence,it is possible to identify features which are closely related to outcome indicators from a large number of features,then obtain stable and efficient prediction models.Therefore,this study screens the clinical parameters of AP patients comprehensively,makes use of machine learning algorithms to establish models for predicting AP complications,and obtains clinical parameters which are closely related to AP complications,so as to explore the role of coagulation parameters in predictive models.First,we conducted a meta-analysis about the difference in coagulation parameters between mild acute pancreatitis(MAP)and SAP patients to determine whether coagulation parameters were related to the severity of AP,meanwhile,we could get evidence that coagulation parameters can help to predict the complications of AP.The meta-analysis provided basis for next research.Then,we comprehensively and retrospectively analysed the 263 MSAP and SAP patients’ clinical data including blood routine,coagulation(coagulogram and TEG)and inflammation parameters at admission,and developed logistic regression analysis(LRA),support vector machine(SVM)and artificial neural network(ANN)models to predict multiple organ failure(MOF),analyzed the predictive efficacy of the three models and compared them with the APACHE Ⅱ and BISAP scores.Finally,we established ANN model to predict intra-abdominal infection,and compared it with LRA to obtain the best model.Material and Methods:Part I The difference of coagulation parameters betweenMAP and SAP: a meta-analysisIn the study,three internationally recognized databases including Pubmed/Medline,Embase and Cochrane Controlled Trial Center were retrieved for clinical studies about coagulation parameters in AP patients.The difference of coagulation parameters between MAP and SAP were analysed by meta-analysis.Part Ⅱ Development and validation of machine-learning models in artificial intelligencefor predicting MOF in MSAP and SAPWe retrospectively analyzed the MSAP and SAP patients admitted to the three affiliated hospitals(Daping hospital,Southwest hospital and Xinqiao hospital)of Army Military University from July 1,2014 to June 30,2017.The diagnostic criteria of AP refered to revised Atlanta classification in 2012.We collected clinical data at admission,those were general data including gender,age,height,weight,body mass index,history of hypertension and diabetes,etiology,and clinical blood biochemical parameters including blood routine,coagulogram,thromboelastogram,inflammation parameters,APACHE Ⅱscore and BISAP score.Modified Marshall score was used to determine whether MOF happened at 48 hours after admission.Firstly,according to the presence or absence of MOF,patients were divided into MOF group and non-MOF group.Univariate analysisby SPSS 23.0 software was used to screen out the different parameters between the two groups.Then,the program SVM,LRA and ANN were completed by MATLAB 2014 software.The different parameters were selected as the characteristics in machine learning,and the area under the ROC curve(AUC)value was obtained by five fold cross-validation.The SVM,LRA and ANN models for predicting MOF were the best combination of features that had the highest AUC values,and the predictive efficacy of these three models,APACHE Ⅱ score and BISAP score were compared.Part Ⅲ Prediction of intra-abdominal infection in MSAP and SAP by ANN modelin artificial intelligenceThe collected clinical data were as same as the part Ⅲ.The diagnostic criteria of intra-abdominal infection refered to thecriteria of infected pancreatic necrosis,i.e.the presence of bubbles on enhanced CT or positive bacterial or fungal cultures in ascites.We also determined whether systemic inflammatory response syndrome happened at admission.The statistical analysis and establishment of ANN and LRA model were completed by SPSS 23.0 software.Firstly,MSAP and SAP patients were divided into two groups: intra-abdominal infection group and non-intra-abdominal infection group.Univariate analysis was carried out to screen out the different parameters between the two groups,and these parameters were enteredinto the backward logistic regression analysis.Then,we used LRA to assess which parameters were independent predictors of intra-abdominal infection and got the logistic regression equation.Meanwhile,we established ANN model.The input layer were the different parameters which were screened by univariate analysis.The output layer were the group with intra-abdominal infection and that non-intra-abdominal infection.The whole group of MSAP and SAP were randomly divided into training set(70%)and validation set(30%)to establish ANN model.Finally,we compared the prediction efficacy of ANN and LRA models.Results:Part I The difference of coagulation parameters between MAP and SAP: a meta-analysis1.Thirteen studies were included in the meta-analysis.There were eleven prospective cohort studies and two case-control studies.A total of 1175 patients including 684 MAP patients and 491 SAP patients were enrolled.The study population came from China,Poland,Italy,Sweden,the United Kingdom,India,Portugal and Serbia.2.The study compared the coagulation parameters between MAP and SAP patients within 24 hours after admission.Six studies compared theprothrombin time between 333 MAP patientsand 199 SAP patients.Five studies compared theactivated partial thromboplastin time between 258 MAP patientsand 164 SAP patients.Four studies compared the international standardized ratio between 203 MAP patientsand 180 SAP patients.Nine studies compared fibrinogen level between 475 MAP patients and 308 SAP patients.Seven studies compared D-dimer level between 407 MAP patients and 324 SAP patients.3.The meta-analysis showed that the plasma prothrombin time of SAP patients was 1.76 seconds longer than that of MAP patients,the fibrinogen level was 0.82 g/L higher than that of MAP patients,and the D-dimer level was 1.37 mg/L higher than that of MAP patients.There was no significant difference in activated partial thromboplastin time and international standardized ratio between the two groups.4.The changes of plasma prothrombin time,fibrinogen and D-dimer levels in SAP patients were significantly higher than those in MAP,suggesting that the activation of the coagulation system and fibrinolysis system in SAP was more significant than those in MAP.The coagulation parameters may be as predictorsof AP severity..Part Ⅱ Development and validation of machine-learning models in artificial intelligence for predicting MOF in MSAP and SAP1.The accuracy of SVM,LRA and ANN artificial intelligence models for predicting MOF in MSAP and SAP patients was higher,not significantly differentfrom that of APACHE Ⅱ score.SVM was more accurate than BISAP score.The AUC values of SVM,LRA,ANN models,APACHE Ⅱ score and BISAP score were 0.840(0.783-0.896),0.832(0.773-0.890),0.834(0.777-0.890),0.814(0.759-0.869)and 0.774(0.716-0.833),respectively.2.K value in thrombelastogram,HCT,IL-6 and creatinine were the common predictors in these three models,indicating that coagulation and inflammation parameters played an important role in predicting MOF.3.SVM,LRA and ANN artificial intelligence models had good prospects for predicting MOF in SAP.We recommended ANN model because it required the least parameters.Part Ⅲ Prediction of intra-abdominal infection in MSAP and SAP by ANN model in artificial intelligence1.ANN model had a better specificity(89.44% vs 77.46%,P < 0.05),false positive rate(10.56% vs 22.54%,P < 0.05),positive predictive value(86.73% vs 72.65%,P < 0.05),negative predictive value(84.67% vs 75.34%,P < 0.05),correct rate(85.55% vs 74.14%,P < 0.05)and AUC value(0.923 vs 0.802,P < 0.05)than LRA for predicting intra-abdominal infection in SAP.ANN was a prospective tool for predicting intra-abdominal infection,however,we needed further prospective trials to validate its value.2.The accuracy of single index such as D-dimer,MA value in the thromboelastogram or interleukin-6 for predicting intra-abdominal infection was lower.Combining coagulation parameters with inflammation parameters were more valuable for predicting intra-abdominal infection.Conclusions:1.The meta-analysis showed that the levels of prothrombin time,fibrinogen and D-dimer of SAP patients were higher than those of MAP at admission,which indicated that the activation degree of coagulation and fibrinolysis system of SAP patients was higher than that of MAP,and exogenous coagulation pathway played a significant role in activation of coagulation system in AP.2.SVM,LRA and ANN artificial intelligence models had no significant difference in predicting MOF in SAP patients compared with APACHE Ⅱ scores.Coagulation and inflammation parameters played important role in predicting MOF.The ANN model was more practical because it required the least parameters.3.ANN model was more accurate than LRA for predicting intra-abdominal infection in SAP.It was a promising tool for predicting intra-abdominal infection in SAP.In summary,the study summarized the published data to clarify the association between coagulation dysfunction and the severity of AP by meta-analysis and got the evidence that coagulation parameters were helpful to predict the complications of AP.Then,based on the clinical data of MSAP and SAP patients with complete coagulation data who were admitted to local hospitals in recent three years,three more convenient and efficient model were established bymachine learning method in artificial intelligence to predict MOF and intra-abdominal infection.And we verified their effectiveness.Finally,we confirmed that these three machine learning models of SVM,LRA and ANN could predict MOF accurately,and the ANN predicted intra-abdominal infection more accurately than LRA.However,SVM,LRA and ANN models were more convenient than traditional scoring system.Meanwhile,we also confirmed that coagulation parameters in TEG and coagulogram had great value for predicting MOF and intra-abdominal infection.In particular,TEG parameters were important factors of predictive model,suggesting that coagulation dysfunction was not only one of the manifestations of SAP but also a key parameter for predicting complications of SAP.We recommended that all MSAP and SAP patients should received TEG and coagulogram examination promptly and dynamically to evaluate coagulation status,guide anticoagulation therapy and use the ANN model to predict the risk of MOF and intra-abdominal infections,so as to formulate an optimal treatment strategy at an early stage and prevent the disease aggravating.There were some limits in the study.Firstly,although we had estimated the sample size before carrying on the retrospective analysis and the sample size of this study met the requirements,the number of patients who were enrolled in the study was less generally.If the sample size could be increased,the conclusion would be more convincing.Secondly,there were no prospective studytoverify the predictive efficiency of machine learning models.And we are preparing to carry out a multi-center prospective cohort study for validating the performance of predictive models,adjusting parameters to make the model stable,and assessing the actual clinical value of the predictive model.Thirdly,although this study found that coagulation parameters were helpful for predicting complications in AP,we did not research the mechanism why coagulation parameters were valuable for prediction in depth.In future work,wehope to establish animal models for exploring the mechanism why coagulation dysfunction could influencethe complications of AP.