Application And Research Of Electronic Health Records Based On Deep Learning

With the wide application of Electronic Health Records systems in the world and development of deep learning,EHRs' data mining,medical staff assisting,EHRs' accuracy improving and medical cost reducing by deep learning are hot search topic.In this thesis,we conduct research and trial of EHRs by deep learning and other methods.This thesis mainly discusses the following research contents:(1)This thesis proposes an Attention-Guided Sequential Feature Learning for Multitask(AGSFLM)model.AGSFLM mainly combines the attention mechanism and multitask learning to learn the common and private features of different tasks.In this paper,We conduct experiments on MIMIC-? dataset by this method.The experimental results show that this method can significantly improve the prediction accuracy,and has excellent results compared with the latest model.(2)In view of the fact that ordinary Regression methods can't estimate patient's length-of-stay time well,considering that patient's length-of-stay time can be regarded as an orderly label prediction problem,this thesis proposes an Ordinal Regression method to model patient's length-of-stay time.Based on deep learning,this thesis considers the sequence relationship of the patient's length-of-stay time and combines the Ordinal Regression method with deep learning.The experimental results show that the introduction of ordered regression to predict patient's length-of-stay time reduces the error and improves the accuracy of the model.(3)It is often for the medical staff in the ICU ward to judge whether the patient needs to be manually intervened in advance.In order to solve this issue,We extract the ventilation,vasoactive medication,and sedative medication tasks according to the MIMIC-? data set,furthermore,we introduce the completion strategy of data missing values.In addition,most of the traditional methods rely on prior knowledge of the medical staff.We predicts the above tasks in single task and multi-task through deep learning.The algorithm can achieve high accuracy without very professional medical background knowledge.(4)This thesis presents a Time Series-based Deep Mixture of Experts for Multi-task Learning(TSDMEML)model.The model combines a mixture of experts and multi-task learning in time series data.It consists of a global feature layer,a multi-layer mixture of experts and a multi-task mixture of experts model.The experimental results show that the TSDMEML model is better than the traditional hybrid expert model and multi-task model.