百度关键词在清远市流行性感冒发病预测中的应用及效果评价

The application and effect evaluation of Baidu keywords in the prediction of influenza in Qingyuan

目的探讨百度关键词在清远市流感发病预测中的适用性。方法收集《中国疾病预防控制信息系统》清远市范围内2012年12月31日至2017年12月24日每周流感发病数,同时从百度指数官网收集同期流感相关关键词在清远市范围内的百度指数,利用流感实际发病数拟合自回归移动平均模型(ARMA),利用实际发病数和相关的流感百度指数拟合自回归分布滞后模型(ARDL),比较两个模型的优劣。结果共收集260周数据,流感发病11 895例,共收集15个流感相关百度指数,其中与实际发病数相关系数大于等于0.3的百度指数6个,与流感发病数呈单向格兰杰因果关系的百度指数有1个,为"流感"。利用前256周数据建立的ARMA模型预测未来4周流感发病数时,模型调整决定系数为0.80,其均方根误差(RMSE)=95.25,平均绝对百分误差(MAPE)=45.49,希尔不等式系数(TIC)=0.25。利用前256周数据建立的ARDL模型预测未来4周流感发病数时,模型调整决定系数为0.81,其均方根误差(RMSE)=79.07,平均绝对百分误差(MAPE)=39.67,希尔不等式系数(TIC)=0.19。ARDL模型的预测精度优于ARMA模型。结论利用百度指数建立的流感发病预测模型能够增加预测准确性,适用于流感发病预测。

Objective To explore the applicability of Baidu keywords in predicting influenza incidence in Qingyuan City. Methods To collect the weekly incidence data of influenza diseases from December 31, 2013 to December 24, 2017 in Qingyuan City, and gather the Baidu indexes related to influenza diseases of Qingyuan City in the same period. Autoregressive integrated moving average（ARMA） model was established using the actual incidence of influenza. Auto-regressive distributed lag（ARDL） model was builded using the actual incidence of influenza and related Baidu indexes, and the applicability of two models were compared. Results A total of 260 weeks＇ data of influenza with 11 895 cases were collected, and a total of 15 flu-related Baidu indexes were collected. Among the Baidu indexes, there were 6 Baidu indexes with correlation coefficients greater than or equal to 0.3 with the actual influenza incidence, and 1 with the one-way Granger causality relationship with the actual influenza incidence, which was ＂Influenza＂. The established ARMA model＇s adjusted coefficient of determination was 0.80 when predicting the next 4 weeks flu incidence, and the root mean square error（RMSE） was 95.25, and the mean absolute percentage error（MAPE） was 45.49, and the Theil＇s inequality coefficients（TIC） was 0.25. The established ARDL model＇s adjusted coefficient of determination was 0.81 when predicting the next 4 weeks flu incidence, and the RMSE was 79.07, and the MAPE was 39.67, and the TIC was 0.19. The ARDL model has better prediction accuracy than ARMA model.Conclusions ARDL prediction model established by using Baidu index can increase the accuracy of prediction and is suitable for the prediction of influenza incidence.