差分自回归移动平均与广义回归神经网络组合模型在丙型肝炎月发病率中的预测应用

Application of ARIMA-GRNN Combination Model in Predicting Monthly Incidence of Hepatitis C

目的探讨差分自回归移动平均(ARIMA)与广义回归神经网络(GRNN)组合模型在丙型肝炎月发病率中预测建模效果及应用前景,为疫情预测提供依据。方法 2015年5月—2016年5月,选取山东省疾病预防控制中心法定传染病直报系统2004—2014年丙型肝炎月度发病率数据及山东省统计局发布的同期人口资料。对2004—2014年山东省丙型肝炎月发病率数据构建ARIMA模型,验证拟合精度并外推预测;将ARIMA模型拟合值作为GRNN模型的输入,实际值作为GRNN模型的输出,对样本进行训练和预测。比较单纯ARIMA模型和ARIMA-GRNN组合模型在丙型肝炎月发病率中的预测效果。结果 2004—2014年山东省丙型肝炎年均发病率为17.28/10万,并随着时间的推移呈上升趋势(Z=29.05,P<0.01)。ARIMA(1,2,1)模型预测2014年山东省丙型肝炎发病率与实际发病率基本一致,落在95%置信区间内,拟合效果较好。以ARIMA(1,2,1)模型拟合值作为GRNN模型的输入,丙型肝炎月发病率实际值作为GRNN模型的输出,取最优光滑因子0.12训练模型,ARIMA-GRNN组合模型预测的拟合值与实际值基本吻合。ARIMA模型和ARIMA-GRNN组合模型的平均误差率(MER)分别为16.87%、15.30%;决定系数(R^2)分别为0.53、0.60;平均绝对误差(MAE)分别为0.17、0.09;平均绝对百分误差(MAPE)分别为1.18、0.35。结论 ARIMA-GRNN组合模型对山东省丙型肝炎月发病率拟合及预测效果优于单纯ARIMA模型,具有较高的拟合精度,有较为广阔的应用前景,对于疫情预测工作有一定的实用性意义。

Objective To explore the predictive modeling effects and application prospects of ARIMA-GRNN combination model in the monthly incidence of hepatitis C,and to provide basis for the epidemic prediction.Methods From May 2015 to May 2016,the 2004—2014 monthly data on the incidence of hepatitis C were selected from direct reporting system of legal infectious diseases in Shandong Provincial Center for Disease Control and Prevention,and the population at the same period released by Shandong provincial Bureau of Statistics were also chosen in the study.ARIMA fitted model of the monthly incidence data of hepatitis C in Shandong province from 2004 to 2014 was constructed,and the fitting precision was verified and extrapolated; the fitted value of ARIMA model was taken as the input of GRNN model,and the actual value of monthly incidence of hepatitis C as the output,and the samples were trained and predicted.The effects of ARIMA model and ARIMA-GRNN combination model on predicting the monthly incidence of hepatitis C were compared.Results The annual average incidence of hepatitis C in Shandong province from 2004 to 2014 was 17.28 /100 000,and showed an increasing trend as time went on（ Z= 29.05,P〈0.01）.By the use of ARIMA（ 1,2,1） model,the predictive incidence of hepatitis C in Shandong province in2014 was basically the same as the actual incidence,which falls within the 95% confidence interval with good fitting effects.The fitted value of ARIMA（ 1,2,1） model was taken as the input of GRNN model,and the actual value of monthly incidence of hepatitis C as the output,the training model with an optimal smoothing factor of 0.12 was selected,and the fitted value of ARIMA-GRNN combination model basically agreed with the actual value.The mean error rate（ MER） of ARIMA model and ARIMA-GRNN combination model were 16.87% and 15.30% respectively; their determination coefficients（ R^2） were 0.53 and0.60 respectively; their mean absolute errors（ MAE） were 0.17 and 0.09 respectively; and the mean absolute percent errors（ MAPE） were 1.18 and 0.35 respectively.Conclusion The fitting and predictive effects of ARIMA-GRNN combination model on the monthly incidence of hepatitis C in Shandong province is better than those of simple ARIMA model,and has a high fitting precision and a promising application prospects.It is of certain practical significance in the epidemic prediction.