模拟指标与观测指标之间的距离在传染病预测模型效果评价中的应用  

Application of the distance between indices of simulation and observation in evaluating the effectiveness of infectious disease prediction models

作　　者：刘天[1] 童叶青[2] 杨瑞[2] 吴杨[2] 宋开发[1] 姚梦雷[1] 黄继贵[1] 赵婧[3] LIU Tian;TONG Yeqing;YANG Rui;WU Yang;SONG Kaifa;YAO Menglei;HUANG Jigui;ZHAO Jin(Department for Infectious Disease Control and Prevention,Jingzhou Center for Disease Control and Prevention,Jingzhou,Hubei 434000,China;不详)

机构地区：[1]荆州市疾病预防控制中心传染病防治所,湖北荆州434000 [2]湖北省疾病预防控制中心 [3]中国疾病预防控制中心卫生应急中心,传染病溯源预警与智能决策全国重点实验室

出　　处：《中国预防医学杂志》2025年第2期199-205,共7页Chinese Preventive Medicine

基　　金：中国疾病预防控制中心公共卫生应急反应机制的运行(102393220020010000017)。

摘　　要：目的介绍模拟指标与观测指标之间的距离(distance between indices of simulation and observation,DISO)在传染病预测模型效果评价中的应用,为传染病预测模型评价提供参考。方法以2005—2021年湖北省细菌性痢疾逐月发病率数据为例,分别建立季节自回归移动平均模型(seasonal autoregressive integrated moving average model,SARIMA)、指数平滑空间状态模型(exponential smoothing state space model,ETS)、TBATS模型(三角函数季节性、Box-Cox变换、ARMA误差、趋势和季节性成分组合模型)和自回归神经网络模型(neural network autoregression,NNETAR)以及上述模型的组合模型共5种模型,预测2022年1—12月湖北省细菌性痢疾发病率。选择平均绝对误差百分比(mean absolute percentage error,MAPE)、平均绝对误差(meanabsoluteerror,MAE)、均方误差根(rootmeansquareerror,RMSE)、平均误差率(mean error rate,MER)以及R2共5个评价指标计算拟合DISO值、预测DISO值和综合DISO值,利用DISO选择最优模型。结果SARIMA、ETS、TBATS、NNETAR和组合模型拟合的MAPE、MAE、RMSE、MER和R2的模型精度顺位分别为5、5、5、5、4;2、2、3、2、4;4、4、2、4、2;3、2、3、2、2;1、1、1、1、1。模型预测精度顺位存在较大差异。SARIMA、ETS、TBATS、NNETAR和组合模型拟合DISO值依次为1.00、0.68、0.74、0.62和0.00,预测DISO值依次为1.00、0.00、0.01、0.17和0.01,综合DISO值依次为1.41、0.68、0.74、0.65和0.01。拟合精度最高模型为组合模型,预测精度最高模型为ETS,拟合及预测综合精度最高的为组合模型。结论DISO可以用于传染病预测模型效果评价,值得推广应用。Objective To introduce the application of distance between indices of simulation and observation(DISO)in evaluating the effectiveness of infectious disease prediction models,providing a reference for model assessment.Method Taking the monthly incidence data of bacterial dysentery in Hubei Province from 2005 to 2021 as an example,five models,including the seasonal autoregressive integrated moving average model(SARIMA),exponential smoothing state space model(ETS),TBATS model(trigonometric seasonal,BoxCox transformation,ARMA error,trend,and seasonal component combined model),neural network autoregression model(NNETAR),and the combination of the above models were established to predict the incidence of bacterial dysentery in Hubei Province from January to December 2022.The fitted DISO value,predicted DISO value,and comprehensive DISO value were calculated using five evaluation indicators:mean absolute percentage error(MAPE),mean absolute error(MAE),root mean squared error(RMSE),mean error rate(MER),and R2.The DISO value was used to select the optimal model.Results The accuracy rankings of the SARIMA,ETS,TBATS,NNETAR,and combination models based on MAPE,MAE,RMSE,MER,and R2 are as follows:For MAPE,SARIMA,ETS,TBATS,and NNETAR models are all ranked 5,while the combination model is ranked 4.For MAE,SARIMA,ETS,and NNETAR models are ranked 2,TBATS is ranked 3,and the combination model is ranked 4.For RMSE,TBATS is ranked 2,SARIMA and NNETAR are ranked 4,and ETS and the combination model are both ranked 4.For MER,ETS and NNETAR models are ranked 2,SARIMA and TBATS are ranked 3 and 2,respectively,and the combination model is ranked 2.For R2,all models are ranked 1,indicating equal high performance in this metric.There were also significant differences in the rankings of model accuracy for the prediction.The fitted DISO values of the SARIMA,ETS,TBATS,NNETAR,and combined models were 1.00,0.68,0.74,0.62,and 0.00,respectively;the predicted DISO values were 1.00,0.00,0.01,0.17,and 0.01,respectively;and the comprehensive DISO value

关 键 词：DISO 传染病 预测 模型 精度 

分 类 号：R183[医药卫生—流行病学]