Study On The Predictive Value Of Acute Severe Pancreatitis Complicated With ARDS Based On Abdominal CT Radiomics

Objective: Develop 3 machine learning models,namely,SVM,LR DNN to predict the occurrence of SAP concurrent ARDS and evaluate the predictive efficacy of the four models by collecting abdominal CT plain images of patients with SAP and extracting imaging histological features.Methods: The all 118 Patients diagnosed with severe acute pancreatitis during hospitalization at the Affiliated Hospital of Guilin Medical University between April 2017 and August 2022 were included in the retrospective study.Basic patient information and abdominal CT scan images were collected and118 patients were randomly divided into 94 patients in the training group and 24 patients.Image histology features were extracted from the pancreas as the ROI in the abdominal CT scan,and the top features were ranked in order of importance using the random forest(RF)algorithm.The top features were sequentially input into the model,and the corresponding image histological features were selected at the maximum AUC of each machine learning model,and 3 machine learning models were established: LR model,SVM model,and DNN model.Predictive performance was assessed by the ROC and the AUC.Result: After ranking the importance of the radiomics features with random forest algorithm,the logistic regression machine learning model reached the maximum AUC when selecting the first four radiomics features,achieving poor prediction performance with AUCs of 0.56 and 0.69 in the validation and training groups.Support vector machine learning models reached the maximum AUC when the first seven radiomics features were selected,with AUCs of 0.60 and 0.72 for the validation and training groups.The deep neural network learning model reached its maximum AUC when the first 10 radiomics features were selected,with AUCs of 0.97 and 0.79 for the validation and training groups.Conclusion: 1.The imageomics features extracted from abdominal CT can be used to build a prediction model to predict the occurrence of SAP complicated with ARDS,which can provide a basis for SAP patients to develop personalized treatment plans.2.LR,SVM and DNN machine learning models established in this study can be used to predict SAP concurrent ARDS.DNN model has the best prediction effect among the three models.