手足口病发病预测4种时间序列预测模型比较

Four time series prediction models for incidence prediction of hand,foot and mouth disease:a comparative study

目的比较季节性差分自回归移动平均模型(SARIMA)、温特线性与季节指数平滑模型、Census X12季节分解模型和线性组合预测模型4种时间序列预测模型对手足口病发病的预测性能,为手足口病的预测提供方法支撑。方法收集中国疾病预防控制中心2008年1月—2019年12月发布的中国手足口病月度发病人数和《中国统计年鉴—2020》发布的年末常住人口数据,据此测算出2008年1月—2019年12月中国手足口病的月度发病率数据;以2008年1月—2018年12月中国手足口病月度发病率数据作为样本建模数据分别构建SARIMA模型、温特线性与季节指数平滑模型、Census X12季节分解模型和线性组合预测模型,以2019年1—12月中国手足口病月度发病率数据作为样本外评估预测数据评价4个模型的预测效果。结果SARIMA模型、温特线性与季节指数平滑模型、Census X12季节分解模型和线性组合预测模型的平均绝对误差(MAD)分别为10.311、14.433、8.424和9.334,预测误差的方差(MSE)分别为30.757、112.847、12.007和18.847,平均相对误差的绝对值(MAPE)分别为1.725%、2.415%、1.409%和1.562%;拟合效果最好的时间序列预测模型为Census X12季节分解模型,其次为线性组合预测模型,再次为SARIMA模型,温特线性与季节指数平滑模型的拟合效果最差。结论Census X12季节分解模型能较好地预测全国手足口病的发病情况,可为今后手足口病的预防控制工作提供决策性依据。

Objective To compare the performance of seasonal autoregressive integrated moving average(SARIMA model,Winter linear and seasonal exponential smoothing model,Census X12 seasonal decomposition model,and linear combination prediction model in incidence prediction of hand,foot and mouth disease(HFMD)for facilitating incidence prediction of the disease.Methods Data on reported monthly HFMD incidence from January 2008 through December 2019 were collected from the dataset published by China Center for Disease Control and Prevention and those on annual population during the same period were collected from China Statistical Yearbook 2020 for calculating the monthly incidence of HFMD in China in the period;then monthly incidence of HFMD in China from January 2008 to December 2018 was used as sample modeling data to construct SARIMA model,Winter linear and seasonal exponential smoothing model,Census X12 seasonal decomposition model and linear combination prediction model;finally the monthly incidence of HFMD in China from January through December 2019 was used as the out-of-sample evaluation prediction data to evaluate prediction efficacy of the four models.Results The mean absolute deviation(MAD),mean square error(MSE),mean absolute percentage error(MAPE)were 10.311,30.757,1.725%for SARIMA model,14.433,112.847,2.415%for Winter linear and seasonal exponential smoothing model,8.424,12.007,1.409%for Census X12 seasonal decomposition model,and 9.334,18.847,1.562%for linear combination prediction model,respectively.The optimum model established was Census X12 seasonal decomposition model,followed by linear combination prediction model,SARIMA model,and Winter linear and seasonal exponential smoothing model.Conclusion The established Census X12 seasonal decomposition model could well predict the incidence of HFMD in China and the utilization of the model could facilitate developing strategies on HFMD prevention and control.