Research On Deep Learning Based Imputation Model For Clinical Time Series

Intensive Care Unit(ICU)have been the focus of the medical field for a long time.Evaluating the severity level of patients accurately and taking necessary measures play an important role in hospitalization.More and more researchers have been devoted themselves to the research of mortality prediction of ICU patients.However,in the medical diagnosis process,missing examinations,forgetting to fill in,etc.,often results in the missing value problem in the time series of medical records.Missing values in the time series,if not properly filled,can seriously impair the performance of the mortality prediction model.How to effectively fill in missing values and evaluate mortality has become an important issue in the field of medical big data.It is obvious that the traditional imputation model and classification model based pipeline pattern cannot meet the requirement if we take filling in missing values and predicting the category label at the same time.In recent years,deep learning has achieved great success in the field of computer vision,natural language processing,and speech recognition.With the rapid growth of medical data,deep learning is widely used in the medical field.Researchers have gradually proposed a series of model based on recurrent neural network for missing values imputation and mortality prediction.While these methods have some common disadvantages,first,they didn't mine the characteristics of the missing patterns deeply to extract effective information,second,they didn't combine different imputation module with mortality effectively prediction to perform a joint learning framework.In summary,aiming at solving the problem of missing values imputation and mortality prediction for medical health time series data,this paper first summarizes and reviews the existing methods,and then proposes two novel deep models based on recurrent neural networks.The main characteristics are as follows:1.The paper introduces four missing pattern representations into the deep learning model: masking,time interval,burstiness,and cumulative missing rate.Thus,we make it possible to use the missing pattern to predict mortality while capturing the long-term dependencies in the time series.2.We design a novel bidirectional multi-task recurrent neural network framework which takes missing values imputation and mortality prediction in the same time series in ICU as main task and divides the former into four sub-tasks.It can make joint learning by using the correlation of main tasks,sub-tasks and main tasks with sub-tasks.The sub-tasks can improve the overall performance of the framework by learning sequence data through the forward method and the backward method.Our model can perform missing value imputation and mortality prediction simultaneously through comprehensive information of main tasks and sub-tasks.3.We propose a loss function for multi-task which combines main tasks and related subtasks efficiently.Then we learn sub-tasks including imputation task based on time information,imputation task based on feature information,sequential imputation task,reverse-order imputation task of imputation problem with mortality prediction problem jointly.4.We evaluated our model on the ICU real dataset MIMIC III and extracted four kinds of disease data according to the ICD-9 classification rule.The experimental results show that our two models have achieved best performance on both missing value imputation and mortality prediction task.Our model can process time series data with multiple irregular outliers while the imputation quality can be effectively guaranteed.Therefore,our model is suitable for ICU applications which contain irregular samples,missing data and noise data,and can fit the medical data perfectly.