Page 106 - 南京医科大学学报自然科学版
P. 106

南京医科大学学报(自然科学版)                                  第42卷第2期
               ·252 ·                     Journal of Nanjing Medical University(Natural Sciences)   2022年2月


             ·公共卫生与预防医学·

              基于深度学习的猩红热流行趋势预测模型研究



              赵子平 ,许       可 ,吴 莹 ,时影影 ,沈文琪 ,彭志行 ,刘文东 ,鲍倡俊 ,周明浩                        3*
                                                                      2
                                                              1
                                      2
                     1,2
                                                      2
                                                                               2*
                              2
                                              2
               南京医科大学公共卫生学院,江苏             南京   211166;江苏省疾病预防控制中心传染病防制所,江苏                  南京   210009;江苏省
              1                                        2                                                3
              卫生健康委员会,江苏 南京 210008
             [摘    要] 目的:探究季节性差分自回归滑动平均(seasonal autoregressive integrated moving average,SARIMA)模型和长短时记
              忆神经网络(long⁃short term memory neural network,LSTM)模型在江苏省猩红热发病趋势预测中的应用,为疫情防控工作提供
              理论依据。方法:以江苏省2005—2017年猩红热的逐月发病数据与气象数据作为训练数据,拟合SARIMA模型和LSTM模型,
              以江苏省2018—2019猩红热的逐月发病数据与气象数据作为测试集,检验模型的预测精度。结果:江苏省猩红热流行具有明
              显的季节性特征,每年4—6月和11月—次年1月为高发时间段。最优SARIMA模型为SARIMA(3,1,2)(1,1,1) 12,最优LSTM
              的结构为以过去3个月的发病数据结合气象公因子作为模型输入,当前周期发病数为对应的期望输出,建立4层LSTM,每层包
              含32个长短时记忆神经元,以及1层全连接层。两种模型测试集的平均绝对误差百分比分别为35.97%、16.94%;均方根误差
              分别为227.85、152.46。提示LSTM神经网络的拟合效果和前瞻性预测精度优于SARIMA模型。结论:LSTM模型对江苏省猩红
              热发病趋势拟合和预测效果较好,可以用于流行趋势研判和风险评估,为优化和调整猩红热监测、防控策略和措施提供依据。
             [关键词] LSTM神经网络模型;SARIMA模型;猩红热;预测
             [中图分类号] R515.1                   [文献标志码] A                        [文章编号] 1007⁃4368(2022)02⁃252⁃07
              doi:10.7655/NYDXBNS20220218


              Studying prediction model of scarlet fever epidemics based on deep learning

                                                                                                        2
                                                            2
                                                                                         3
                                   2
                          1
                                             2
                                                                          2
              ZHAO Ziping ,XU Ke ,WU Ying ,SHI Yingying ,SHEN Wenqi ,PENG Zhihang ,LIU Wendong ,BAO
                      2*
              Changjun ,ZHOU Minghao 3*
              1 School of Public Health,Nanjing Medical University,Nanjing 211166;Jiangsu Provincial Center for Disease
                                                                             2
                                                3
              Control and Prevention,Nanjing 210009;Jiangsu Provincial Health Commission,Nanjing 210008,China
             [Abstract] Objective:This study aims to compare the performance of seasonal autoregressive integrated moving average(SARIMA)
              model and long⁃short term memory(LSTM)model in predicting the epidemics of scarlet fever in Jiangsu Province,China,which may
              provide decision basis for prevention and control of scarlet fever. Methods:SARIMA model and LSTM model were fitted with data of
              monthly scarlet fever cases as well as meteorological variables during 2005 ⁃ 2017 in Jiangsu Province,China. Performance of the
              models were evaluated with data during 2018 ⁃ 2019. Results:The epidemics of scarlet fever in Jiangsu Province showed obvious
              seasonality,higherincidences were observed from April to June and from November to next January. SARIMA(3,1,2)(1,1,1) 12 had
              the best perfor mance of all the alternative SARIMA models. The optimal LSTM model had 4 LSTM layers and 1 full connected layer,
              and each LSTM layer contains 32 memory cells. The mean absolute percentage error(MAPE)of SARIMA model and LSTM model in
              testingset were 22.47% and 16.94%,respectively,and the root mean squared error(RMSE)were 227.85 and 152.46,respectively.
              Conclusion:LSTM model performed well in predicting the incidence of scarlet fever in Jiangsu province. This model can be used to
              investigate the prevalence trends and assess the epidemic risk of scarlet fever,thus providing basis for optimizing and adjusting
              monitoring,prevention and control strategies and measures of this disease.
             [Key words] deep learning;SARIMA model;scarlet fever;prediction
                                                                         [J Nanjing Med Univ,2022,42(02):252⁃257,263]



             [基金项目] 江苏省卫生健康委科研项目(Z2019006)
              ∗
              通信作者(Corresponding author),E⁃mail:zmh@jscdc.cn;bao2000.cn@163.com
   101   102   103   104   105   106   107   108   109   110   111