Research And Implementation Of Survival Prediction And Reliability Evaluation Method Based On Data Augmentation

In recent years,there have been numerous studies on deep learning for survival prediction in Electronic Health Records(EHR).Due to the time-series nature of EHR,many studies use traditional recurrent neural networks for model building,and attention mechanisms are often used in research,The research can assist doctors in conducting reasonable clinical analysis,and doctors can use the analysis results as a basis to make effective clinical decisions.However,in the task of survival prediction based on time series EHR data,existing deep learning models face some challenges:(1)In actual medical datasets,there are often many missing values,which will seriously reduce the data quality and eventually interfere with model prediction effect.Besides,most existing methods do not consider the confidence of the estimated values in survival prediction modeling after estimating missing values,and misestimating missing variables may lead to modeling difficulties and performance degradation.(2)Most methods only focus on the extraction of global time series features,while ignoring some local specific information is also important for predicting patient mortality.This will result in incomplete model consideration,making it difficult to accurately predict patient mortality.(3)Since existing deep learning models only output classification probabilities,they are prone to overconfidence and difficult to generate valid uncertainty scores,thus affecting the reliability of the model.This will lead to models producing unreliable predictions that may misjudge patient health.Therefore,to address these challenges,this thesis conducts an in-depth study of the above issues based on time series EHR data:1.A Survival Prediction Model with Missing Value Imputation for Multivariate Time Series(MTSSP)is proposed.The model proposes a gated recurrent unit that can receive a mask matrix and time interval as supplementary information.This unit can impute missing values according to the absence and time of missing values,thereby enhancing the representational ability of EHR data.And in the downstream survival prediction task,based on the imputation completed data set,this thesis uses a one-dimensional dilated convolutional neural network and a bidirectional recurrent neural network to mine the time series EHR from both local and global aspects,respectively,to jointly capture the regularity of patient visits,thereby improving the performance of the survival prediction model.2.A reliability assessment mechanism based on VAE and MC Dropout model is proposed.Based on the first research content,the mechanism constructs a recurrent network cell structure UN-Cell based on Variational Auto-Encoder(VAE),which utilizes VAE to generate the mean and variance of each missing variable while recursively in time.The uncertainty score is generated inside the unit according to the variance of each variable,and the provision of the uncertainty score can bring more accurate results to the missing value completion task.Furthermore,in order to achieve model reliability,the model needs to provide an uncertainty score for the prediction.This thesis uses the Monte-Carlo Dropout method to provide uncertainty scores for the model prediction results.To realize the reliability evaluation of both the imputed value and the model output value.3.Based on the trained model,this paper designs and implements a prediction system that can predict patient mortality based on historical medical data of patients,which can provide auxiliary decision support for medical staff.Finally,this thesis conducts model validation on two public datasets,MIMIC-Ⅲ and MIMIC-Ⅳ.The experimental results show that the model proposed in this thesis can effectively improve the accuracy of the survival prediction model.