Predicting The Severity Of Acute Gallstone Pancreatitis Based On Machine Learning And Radiomics

Background and Aim: The severity assessment of patients with acute gallstone pancreatitis(GSP)is of great significance for subsequent treatment.The aim of this study was to explore the value of a machine learning model based on CT imaging(including clinical laboratory examination,CT characteristics,and radiomics)for early prediction of GSP severity.Methods: The clinical and CT data of patients hospitalized for GSP in the First Hospital of Lanzhou University from January 2016 to January 2022 were retrospectively analyzed,and the data sets were randomly divided according to the ratio of 7 : 3.The collected patient data included clinical(age,gender,BMI),laboratory examination(white blood cell,neutrophil percentage,hemoglobin,calcium,total bilirubin,direct bilirubin,alkaline phosphatase,gamma-glutamyltransferase,etc.)and CT features(gallbladder wall thickening,number of gallstones,maximum gallbladder diameter,maximum stone diameter,bile duct diameter,pancreatic duct diameter,etc.)were collected.Mutual information(MI)method and random forest(RF)model were used for feature selecting,and a prediction model ML GSP was established.Shapley additive explanation(SHAP)is used to show the importance of each variable,and the overall relationship between GSP severity and each variable is described according to the model.Univariate and multivariate Logistic regression analyses were performed on clinical,laboratory and CT features to verify the accuracy of feature screening by machine learning.3D-slicer was used for image segmentation,standardization processing,region of interest(ROI)delineation and radiomics feature extraction.The Mann-Whitney U test and least absolute shrinkage and selection operator(LASSO)regression were utilized for screening and calculating the radiomics score to establish a prediction model.The model’s predictive performance was evaluated using receiver operating characteristic curve(ROC),area under ROC curve(AUC),accuracy,sensitivity,specificity,positive predictive value(PPV),and negative predictive value(NPV).To verify the calibration performance and clinical efficacy of the model,calibration curve,decision curve analysis(DCA),and clinical impact curve(CIC)were employed.In addition,we have built a Web-based ML GSP model calculator,which is freely and openly accessible at http://www.pan-chess.The study was conducted according to the guiding principles of the Helsinki Declaration and were registered with Clinicaltrial.gov(NCT05498961).Results: A total of 301 eligible patients,ranging in age from 12 to 87 years old,were included,including 161 males and 140 females.The data sets were randomly divided according to 7:3,among which the training cohort(n=210)and the verification cohort(n=91)had no statistical significance(P≥0.05),and the data sets were evenly divided.The CT features,clinical and laboratory tests were ranked using the MI method.The RF model then identified seven risk factors most associated with predicting GSP severity at the maximum AUC,including calcium,WBC,urea,combined acute cholecystitis(yes,no),gallbladder wall thickening(yes,no),gallstones(single,multiple)and hydrothorax(yes,no).The ML GSP model was established.Multivariate Logistic regression showed gallbladder wall thickening(OR=2.224;95% CI: 1.488-3.325;P <0.001),multiple gallstones(OR=1.830;95% CI: 1.317-2.543;P < 0.001),combined with acute cholecystitis(OR=2.701;95% CI: 1.931-3.776;P < 0.001),urea(OR=1.472;95% CI: 0.955-2.180;P = 0.053),calcium(OR=0.479;95%CI: 0.322-0.713;P < 0.001)with hydrothorax(OR=2.165;95% CI: 1.564-2.996;P< 0.001)was an independent risk factor for identifying the severity of GSP,which was basically consistent with the risk factors of RF method.In the training cohort,the AUC of ML GSP,BISAP and MCTSI models were 0.916(0.872-0.958),0.770(0.703-0.836)and 0.751(0.682-0.819),respectively.In the validation cohort,the AUC of ML GSP,BISAP and MCTSI models were 0.914(0.851 to 0.978),0.767(0.667 to 0.867)and 0.740(0.636 to 0.844),respectively.Calibration plots show that the ML GSP model has a good agreement between the predicted probability and the observed probability in the training and validation queues.The DCA curve shows that the ML GSP model provides a good net benefit within the probability range of greater than 15% threshold.The Mann-Whitney U test and LASSO regression reduced the number of radiomics from 1171 to 29,and established the radiomics score.The AUC,accuracy,sensitivity and specificity of Radiomics model constructed based on RF were 0.878(CI: 0.826-0.928),0.800,0.893 and 0.633 in the training cohort.In the validation cohort,the AUC was 0.841(CI: 0.757-0.926),the accuracy was 0.813,the sensitivity was 0.830,and the specificity was 0.763,showing a good classification effect.Calibration curves,and DCA showed good clinical benefits.The example prediction using the SHAP value method shows the better prediction performance of the model.Conclusion: The ML GSP model was established based on machine learning to determine the risk factors for predicting the severity of GSP.The model shows high diagnostic performance in the classification of mild and severe GSP,and has great guiding value for further treatment of patients.Establish the application of network prediction tool promotion model.In addition,the radiomics model also shows strong predictive performance,which can provide certain reference value for the follow-up clinical decision-making of patients.