Prediction Of Early Mortality Risk In Patients With Acute Pancreatitis Based On Few Shot Learning

Acute pancreatitis is an acute abdominal inflammatory disease,and the mortality rate of patients is still as high as 30%,especially early death is the most common.On the one hand,early mortality risk prediction for patients with acute pancreatitis provides an important basis for clinicians to make reasonable decisions,helps doctors make early diagnosis quickly,and reduces patient mortality.On the other hand,it can carry out risk stratification for patients,give more rapid diagnosis and treatment to severe patients,and rationally allocate limited medical resources.However,in clinical practice,problems such as imbalanced patient data and small sample size have brought challenges to predicting the early mortality risk of acute pancreatitis.In this background,it is of great practical significance and clinical value to study the method for predicting the early mortality risk of acute pancreatitis in the case of a small sample.In this paper,on the basis of preprocessing patient data and screening effective indicators,a method for predicting early mortality risk in acute pancreatitis patients based on few shot learning is proposed.The main research contents are divided into the following parts:Firstly,the data of patients with acute pancreatitis used in this paper are retrieved from the MIMIC-IV database,and the patient data are filled with missing values,normalized,and feature dimensionality reduction.Combined with chi-square test,analysis of variance and XGBoost method,the important characteristics for judging the early mortality risk of patients are screened out.Secondly,considering both data augmentation and small sample models,a method for early mortality risk prediction in patients with acute pancreatitis is proposed.Using the Smote Tomek sampling method to generate new samples,while solving the problem of data category imbalance,data augmentation is performed on small sample data.A graph neural network based few-shot model is then used for mortality risk prediction,which is learned from a limited number of training examples by aggregating similar patient information by instead using a Gaussian function as an attention kernel function.Compared with 7 classical machine learning algorithms such as random forest and neural network,the method proposed in this paper has better prediction results,where the sensitivity is 0.9130,the specificity is0.9637,the F1 score is 0.8235,and the AUC is 0.9384.Finally,on the basis of few shot learning,a nomogram is established for prediction of patient mortality risk score.Using conventional and dynamic nomograms not only facilitates doctors to explain the disease to patients simply and intuitively,but also helps patients predict risks online.Compared with the commonly used clinical SIRS,SOFA and SAPSII scores,the nomogram score prediction model has a C-index of 0.961,which has a better prediction effect,and is also superior to other scores in terms of clinical decision-making benefits,and has high practical value.