Study On The Interpretability Of Predicting Mortality Rate Of ICU Patients Based On Deep Learning

With the development of medical and health services,the electronic medical record system has made great progress.In the 11 th Five Year Plan for the construction of health information,the state proposed to promote the electronic medical record and other health information projects.Relying on the electronic medical record system,the hospital can get a lot of valuable data to study it.The research on the mortality of EMR can be traced back to 1991.Nowadays,deep learning algorithm has shown its excellent performance in many fields.Therefore,more and more researches on EMR are using deep learning algorithm.However,the general deep learning algorithm can not explain the basis and principle of model classification well,and people have low trust in the model.Based on this problem,this paper uses an interpretable deep learning model to classify ICU patients and discuss the results.In this paper,we extract the physiological data of ICU patients from mimic Ⅲ medical database,convert the text data of patients into picture data,use the migration learning method,combine the deep neural network based on VGG16-Grad-CAM,use the method of VGG16-Grad-CAM to predict the survival and death of patients,and analyze the results.At the same time,compared with the nearest neighbor algorithm,support vector machine algorithm,multi-layer perceptron algorithm,random forest algorithm and CNN,the classification performance of the model is discussed.Finally,the model VGG16-Grad-CAM gets the results of precision = 0.89,recall = 0.89 and F1 score = 0.88,which is better than the nearest neighbor algorithm,multi-layer perceptron algorithm,random forest algorithm and CNN.Through the grad cam method,this paper explains the classification results of the model,and gets the conclusion that the model pays more attention to the five characteristics of blood pressure,heart rate,K +,hematocrit and WBC in the classification process,and the model pays more attention to the physiological data of patients in the early stage of entering ICU and more than 40 hours.