ICU病人数据对LSTM和GRU预测模型效果及收敛速度的对比分析  

Comparing the Train Speeds and Effects of LSTM and GRU Related Predictive Models for Congestive Heart Failure ICU Patients Data with Missing Values

作　　者：蔡慧 罗佳伟 CAI Hui;LUO Jiawei(College of Mathematics of Sichuan University,Chengdu,Sichuan 610065;College of Clinical Medicine of Sichuan University,Chengdu,Sichuan 610041)

机构地区：[1]四川大学数学学院,四川成都610065 [2]四川大学临床医学院,四川成都610041

出　　处：《绵阳师范学院学报》2020年第11期1-10,共10页Journal of Mianyang Teachers＇ College

基　　金：国家自然科学基金资助项目(11670218).

摘　　要：随着生物医学和医疗保健的大数据增长以及电子病历记录(EHR)数据的大量数字化,对医学数据的准确分析将有助于早期疾病检测.但是,当医学数据的质量不完整时,分析准确性会降低.为了捕捉潜在复杂信息,提高准确率,越来越多人选择使用神经网络预测模型,其中RNN的变体如LSTM和GRU预测效果不错,但模型需要特别苛刻的条件,首先数据必须完整无缺失,其次时间步长固定,且模型本身不捕获时间信息.在本文中,我们在已有的变体中拓展了模型,共讨论了四种LSTM和GRU的变体模型,分别是LSTM-D、GRU-D、P'-LSTM-D和P'-GRU-D,它们都能直接处理带有缺失的数据.我们从大型公共重症监护医学数据库MIMIC III中提取了10315个充血性心力衰竭病人的数据,比较了四种模型的效果及收敛速度.发现GRU-D和LSTM-D两种模型的表现效果极佳,在以呼吸频率为临床结局的任务中,平均AUC能够到达0.96,另外GRU-D比LSTM-D的收敛速度要快一些.With the big data growth in bio-medicine and health care and the massive digitization of electronic health care record(EHR)data,accurate analysis of medical data benefits early disease detection.However,the analysis accuracy is reduced when the quality of medical data is incomplete.In order to capture potentially complex information and improve accuracy,more and more researchers choose to use neural network prediction models.RNNs'variants,such as LSTM and GRU,have good prediction results,but the model requires particularly harsh conditions.First,the data must be complete without missing values.Secondly,the time step is fixed,and the model itself does not capture time information.In this article,we extend the model from the existing variants of LSTM and GRU and discuss four variant models,namely LSTM-D,GRU-D,P'-LSTM-D,and P'-GRU-D.And both of them can deal with missing data directly.We extracted 10315 congestive heart failure patients'data from the large public intensive care medical database MIMIC III,and compared the effects and convergence speeds of the four models.We found that the GRU-D and LSTM-D models did very well.The average AUC can reach 0.96 in tasks with respiratory frequency as the clinical outcome.In addition,GRU-D has a faster convergence rate than LSTM-D.

关 键 词：大数据分析 神经网络 LSEM GRU 缺失预测模型 电子病历数据 

分 类 号：O235[理学—运筹学与控制论]