基于机器学习整合气象、互联网和监测数据的流感预测模型构建与评价  

Construction and evaluation of influenza prediction models based on the integration of meteorological,internet,and surveillance data using machine learning

作　　者：于树悦 梁焱榆 陈立凌[3] 庞媛媛[3] 汤景云 朱杰 YU Shuyue;LIANG Yanyu;CHEN Liling;PANG Yuanyuan;TANG Jingyun;ZHU Jie(Department of Information Statistics,Suzhou Health and Family Planning Statistics Information Center,Suzhou 215000,China;School of Public Health,Suzhou University,Suzhou 215123,China;Department of Acute Infections Disease Prevention and Control,Suzhou Center for Disease Control and Prevention,Suzhou 215100,China)

机构地区：[1]苏州市卫生健康信息中心信息统计科,江苏苏州215000 [2]苏州大学公共卫生学院,江苏苏州215123 [3]苏州市疾病预防控制中心急传科,江苏苏州215100

出　　处：《现代医学》2024年第11期1695-1702,共8页Modern Medical Journal

基　　金：苏州市重点扶持学科项目(SZFCXK202146);江苏省预防医学科研课题(Ym2023019);苏州市科技发展计划重点项目(2022SS14);苏州市科技计划(SKY2022095)。

摘　　要：目的:探索气象因素、互联网数据在苏州市流感例预测方面的作用,并基于机器学习方法构建苏州市流感预测模型。方法:收集苏州市2012年1月1日至2019年12月29日气象资料、流感监测资料以及互联网流感关键词搜索资料,采用交叉相关分析方法检验初筛气象因素、流感关键词与流感病例在前、后4周时间范围内的时滞相互关系。根据时滞相关性分析,对关键词进一步过滤,基于确定的不同类型关键词、流感样病例百分比数据以及流感病原阳性检出率数据,构建流感预测模型,并选用SARIMA、LSTM、Att-LSTM 3种方法进行模型的训练和构建。以均方误差(MSE)、平均绝对误差(MAE)和均方根误差(RMSE)3个指标对各模型进行评价。结果:SARIMA模型的拟合度较低,Att-LSTM模型拟合度较高,其MSE、RMSE、MAE值分别为0.055、0.235、0.184。结论:基于气象、互联网和监测数据构建的Att-LSTM模型可以有效提高模型预测精度,其结果可为苏州地区实现更加精准的流感防控提供科学依据。Objective:To Explore the role of meteorological factors and internet data in predicting influenza cases in Suzhou City,and construct influenza prediction models for Suzhou City based on the method of machine learning.Methods:We collected meteorological data,influenza surveillance data,and internet influenza keyword search data in Suzhou from January 1,2012,to December 29,2019.Then,we used cross-correlation analysis to test the lagged relationships between the preliminary screened meteorological factors,influenza keywords,and influenza cases within a 4-week time frame.Based on the lagged correlation analysis,researchers further filtered the keywords.Utilizing the determined different types of keywords,influenza-like illness consultation rate and positive rate,we constructed influenza prediction models by using SARIMA,LSTM,and Att-LSTM methods.Finally,we evaluated each model using the Mean Squared Error(MSE),Mean Absolute Error(MAE),and Root Mean Squared Error(RMSE)metrics.Results:The fitting of the SARIMA model was relatively low,whereas the Att-LSTM model showed a high fitting degree,with its MSE,RMSE,and MAE values respectively being 0.055,0.235 and 0.184.Conclusion:The Att-LSTM model,constructed based on meteorological,internet,and surveillance data,can significantly enhance predictive accuracy.The results will provide a scientific basis for more precise influenza prevention and control efforts in Suzhou.

关 键 词：机器学习 流感 气象 搜索指数 预测模型 

分 类 号：R183.3[医药卫生—流行病学]