Research On The Prediction Model Of Community-acquired Pneumonia Mortality Based On Long Short-term Memory

Community-acquired pneumonia(CAP)is a common cause of patients admitted to intensive care unit(ICU).The prevalence of CAP in developed countries is far less than that in developing countries.The chance of CAP patients admitted to hospital and death event is about 13%,while the mortality rate of severe CAP patients is about 35%,of which the30-day mortality rate is extremely high.Using the data of electronic health record(EHR)to fully and effectively predict and evaluate the patient’s in-hospital mortality risk is one of the urgent medical tasks to be solved,especially for those who have been admitted to ICU.However,in practice,in view of the time constraints and complex conditions,it is unrealistic to make a quick and effective diagnosis of CAP.Then it is very meaningful to establish an accurate and efficient predictive model to assist doctors in making decisions.The traditional death risk assessment method is not effective in predicting the 30-day death risk of CAP.There are also some models based on logistic regression,which are difficult to effectively deal with high-dimensional data and relatively complex data in EHR.Therefore,this thesis proposes a method based on long short-term memory(LSTM)to predict the 30-day mortality rate of CAP.At present,more and more EHR time series data sets are collected,which can be used in combination with LSTM to diagnose many health problems and predict prognosis.The main contents of this thesis:(1)Data set acquisition and preprocessing.Analyze and summarize the scoring system and literature that judge the condition of CAP and the risk of death from CAP to determine the risk indicators to be selected in this study.Obtain the data set from the target database according to the risk indicators,and then preprocess it to complete the data preparation work;(2)Construction of a prediction model for the mortality of CAP.Based on LSTM,the mortality prediction model of CAP was established.The structure and parameters of the model were studied and set up,and the over fitting problem was also considered.Then the model is trained and tested.From the scoring index of the model,the effect of LSTM model is better than the model without considering the time sequence of data,indicating that it can fully mine the useful information inherent in the time series data,so the model can complete the task of predicting the 30-day mortality of CAP;(3)CAP mortality prediction application system development.The application of the prediction model is expanded,and the 30-day mortality prediction system of CAP is developed with object-oriented design method and Java language.This thesis conducts related research on the LSTM-based CAP mortality prediction method,and the mortality risk of CAP was predicted,so as to help ICU doctors provide better treatment for CAP patients.