3种预测模型在北京市西城区痢疾发病数预测中的应用与比较  

The application and comparison of three forecasting models for predicting the monthly number of dysentery in Xicheng District in Beijing

作　　者：孙小宇[1] 初艳慧[1] 张震[1] 刘潇潇[1] 

机构地区：[1]北京市西城区疾病预防控制中心传染病预防控制科,北京100120

出　　处：《现代预防医学》2014年第19期3470-3474,共5页Modern Preventive Medicine

基　　金：北京市西城区卫生局青年科技人才(科技新星)培养项目(xwkx2012-08)

摘　　要：目的探讨ARIMA季节乘积模型、温特斯法模型及多层感知器神经网络模型在痢疾预测中的应用。方法采集中国疾病预防控制信息系统中2005年1月-2012年12月报告的现住址为西城区的痢疾发病数。选择2012年9月前月发病数据进行建模拟合,应用拟合模型对2012年10月-12月的发病情况进行预测验证。结果应用训练样本构建的ARIMA(1,0,0)(0,1,1)12模型、温特斯法模型及多层感知器模型对原始数据拟合的平均绝对百分误差(MAPE)分别为21.06%、18.70%及25.51%,对2012年10-12月预测的MAPE分别为15.43%、26.67%、47.17%,模型预测能力有强到弱依次为ARIMA(1,0,0)(0,1,1)12>温特斯法模型>MLP神经网络模型。结论由于传染病的影响因素较复杂,应用历史数据构建模型进行预测是可行的。但应注意结合数据变动趋势选择合适的模型。本研究中,ARIMA乘积模型和温斯特法模型均适合痢疾的预测,可以为早期预警提供依据。Objective To research the application of ARIMA product season model,Winters＇ Multiplicative model and multilayer perceptron（MLP） neural network in predicting the monthly number of dysentery.Methods Collected the number of dysentery patients who were living in Xicheng district from Jan.2005 to Dec.2012.Data from Jan.2005 to Sep.2012 were used as training data,and data from Oct.2012 to Dec.2012 were used as test data.Results The mean absolute percentage errors（MAPEs） of ARIMA（1,0,0）（0,1,1）12,Winter＇s Multiplicative model and MLP for matching the training data were respectively 21.06%,18.70%and 25.51%.Models were used to forecast the test data,the MAPEs were respectively 15.43%,26.67% and 47.17%.For the test data,the best model was ARIMA（1,0,0）（0,1,1）12,followed by Winter＇s Multiplicative model and MLP.Conclusion As factors influencing the incidence of infectious disease are complicated,it is feasible to forecast with models built with historical data.But it is important to choose the best model combining the trends and distributions of data.In our research,ARIMA product season model and Winters＇ Multiplicative model are suitable to predict the monthly number of dysentery,and can provide early warning information on infectious diseases.

关 键 词：ARIMA季节乘积模型 温特斯法模型 多层感知器神经网络模型 痢疾发病预测 

分 类 号：R195[医药卫生—卫生统计学]