两种模型对百日咳发病数预测效果的比较研究

Comparison of the effects of two prediction models on the number of pertussis cases

目的分析2004—2020年我国百日咳发病数据,了解百日咳发病特征及规律,比较径向基函数(Radial Basis Function,RBF)神经网络模型、自回归综合移动平均模型(autoregressive integrated moving average model,ARIMA)对百日咳发病数的预测效果,为防控百日咳提供依据。方法选取2004—2020年《中国卫生和计划生育统计年鉴》中的百日咳逐月发病数作为拟合数据,建立RBF神经网络模型和ARIMA模型,对2021年1—9月我国百日咳月发病数进行拟合预测,根据国家卫生健康委员会官方网站发布的2021年1—9月百日咳实际月发病数,比较两种模型的预测效果。结果ARIMA(2,1,0)(1,0,1)12对2021年1—9月发病数预测的平均相对误差为-34.98%,RBF神经网络模型预测百日咳发病数的平均相对误差为-24.75%。RBF神经网络模型在拟合效果数值上优于ARIMA(2,1,0)(1,0,1)12的拟合效果,但差异无统计学意义(Z=-0.839,P=0.402),尚不能认为RBF神经网络模型拟合效果优于ARIMA模型。结论RBF神经网络模型和ARIMA模型均能短期预测百日咳发病情况。

Objective To analyze the incidence data of pertussis in China from 2004 to 2020,and to know the characteristics of pertussis incidence.Radial Basis Function(RBF)neural network and autoregressive integrated moving average(ARIMA)were compared in forecasting incidence,to provide evidence for the prevention and control of pertussis.Methods The monthly incidence of pertussis in China from 2004 to 2020 was collected as the fitting data,and the RBF neural network and ARIMA were established.According to the actual monthly incidence of pertussis from January to September,2021 published on the Official website of National Health Commission,the prediction effect of the two methods was compared.Results The average rediction relative error of ARIMA(2,1,0)(1,0,11)12 for actual monthly incidence of pertussis from January to September,2021 was -34.98%,while that of RBF neural network model was -24.75%.The fitting effect of RBF neural network model was better than ARIMA(2,1,0)(1,0,1)12,with no statistically significant difference(Z=-0.839,P=0.402).So it could not be concluded that the fitting effect of RBF neural network is better than ARIMA.Conclusion Both RBF neural network and ARIMA can be used for short-term prediction of pertussis morbidity.