摘要
目的研究长短记忆神经网络模型(long-short term memory,LSTM)预测未来每周流行性感冒(流感)暴发趋势的可行性,以提高预测精度和工作效率。方法选取2007-2017年深圳市宝安区每周流感发病数,考虑到时间序列数据的自相关性和高频特征,基于深度学习思想构建长短记忆神经网络模型对流感暴发趋势进行预测,并使用5种方法对LSTM模型预测效果进行评估。结果深圳市宝安区流感报告周频率具有随机波动性和周期性特征,通过LSTM神经网络模型能够准确学习时间序列特征并用于外推预测。仿真结果表明,相比ARIMA模型、BP神经网络、小波神经网络(WNN)、广义回归神经网络(GRNN)和动态自回归神经网络(NARX),LSTM神经网络的拟合度更接近实际值,预测精度较高。结论在数据量大和非平稳及周期特征的情况下,深度学习的LSTM神经网络对流感的预测效果更好。
Objective To study the feasibility of long-short term memory(LSTM)network model predicting weekly influenza outbreak trends in the future to improve predictive accuracy and work effective.Methods To select weekly influenza cases from 2007 to 2017,the auto-correlation and high frequency characteristics of time series data was taken into account and based on deep learning constructing the LSTM network model to predict influenza outbreak trends.5 methods were used to evaluate the effects of model prediction.Result The weekly influenza report numbers demostrated characteristics of random volatility and periodicity.Simulation result showed that the fitting of LSTM network model was closer to the actual values and prediction accuracy is higher compared to ARIMA model,BP network model,WNN model,GRNN model and NARX model.Conclusion In the case of large amount of data and non-stationary,the deep learning LSTM network model has a better predictive effect on influenza,and can provide important scientific bases for disease prevention and control.
作者
陈亿雄
李苑
刘小明
李淑珍
CHEN Yi-xiong;Li Yuan;Liu Xiao-ming;Li Shu-zhen(Bao'an District Center for Disease Control and Prevention,Guangdong shenzhen 518101,China)
出处
《江苏预防医学》
CAS
2019年第6期622-625,共4页
Jiangsu Journal of Preventive Medicine
基金
深圳市科技创新委员会(JCYJ20160427155352873)
深圳市宝安区科技创新局(2016CX226)
