Fig 1.
Spatiotemporal dynamics of dengue cases and dengue search index (DSI) during 2011–2014 in Guangdong province, China.
(A) Geographical distribution of dengue cases in Guangdong province, China in 2014. (B) Time series of dengue cases and DSI in Guangzhou city. (C) Time series of dengue cases and DSI in Foshan city. (D) Time series of dengue cases and DSI in Zhongshan city. (E) Time series of dengue cases and DSI in Zhuhai city. (F) Time series of dengue cases and DSI in Shenzhen city.
Fig 2.
Comparison of prediction performance and goodness of fit of the models considered involving the support vector regression (SVR) model, step-down linear regression model, gradient boosted regression tree model (GBM), negative binomial regression model (NBM), least absolute shrinkage and selection operator (LASSO) linear regression algorithm and generalized additive model (GAM) using the root-mean-square error (RMSE) and R-squared statistic, respectively.
(A) Data corresponding to the period between the 41st to 53rd weeks (the last 12 weeks) in 2014 was used to assess the models using the RMSE. (B) Data corresponding to the period between the 35th to 46th weeks which covers the outbreak in dengue incidence in 2014 was used to assess the models using the RMSE. (C) Data corresponding to the period between the 41st to 53rd weeks (the last 12 weeks) in 2014 was used to assess the models using the R-squared. (D) Data corresponding to the period between the 35th to 46th weeks which covers the outbreak in dengue incidence in 2014 was used to assess the models using the R-squared. The RMSE and R-squared values were standardized according to the specific city in Guangdong province.
Fig 3.
Observations and model predictions of dengue case counts in Foshan city, China, 2014.
(A) Model forecasts using the SVR algorithm for the dengue epidemic period between the 41st to 53rd weeks (the last 12 weeks) in 2014. The black lines represent observed values, the blue dashed lines denote model-based fitted values, the red dashed lines correspond to model-based predicted values, and the pink contours represent the corresponding 95% prediction intervals. The observations and predictions of dengue case counts are expressed as a log-scale. (B) Residuals of the SVR model for the last 12 weeks forecasts were assessed using the autocorrelation function (ACF) plot. (C) Model forecasts using the SVR algorithm for the period between the 35th to 46th weeks which covers the outbreak in dengue incidence in 2014. (D) Residuals of the SVR model for the outbreak period forecasts were assessed using the ACF plot.
Table 1.
Comparison of model performance and goodness-of-fit for support vector regression (SVR) model, step-down linear regression model (Linear), gradient boosted regression tree model (GBM), negative binomial regression model (NBM), least absolute shrinkage and selection operator (LASSO) linear regression algorithm and generalized additive model (GAM) by the means of root-mean-square error (RMSE) and R-squared, respectively.
Two prediction periods were considered: 1) data corresponding to the period between the 41st to 53rd weeks (the last 12 weeks) in 2014 was used to validate the models; 2) data corresponding to the period between the 35th to 46th weeks which covers the outbreak in dengue incidence in 2014 was used to validate the models. Results are presented for five cities with a high risk of dengue infection, and the other cities in Guangdong province.
Fig 4.
Observations and model predictions (1-week-ahead predictions) for the dengue outbreak in Guangdong, 2014.
(A) Observations and model predictions of dengue case counts were only shown for five cities with a high risk of dengue infection in Guangdong province. In each panel, the blue points represent observed case counts, the red dashed lines denote model-based predicted values. Dynamic forecasts of dengue epidemics are presented in Video Files 1–5, respectively. (B) The actual dengue incidence map and that from the SVR model-based 1-week-ahead predictions in Guangdong, 2014. Incidence is expressed as the number of case counts per 100,000 people.
Fig 5.
Temporal dynamics of dengue infection and dengue search index (DSI) during 2011–2014 in mainland China.
(A) Time series of dengue cases and DSI in Guangdong province. (B) Time series of dengue cases and DSI in Yunnan province. (C) Time series of dengue cases and DSI in Guangxi province. (D) Time series of dengue cases and DSI in Hunan province. (E) Time series of dengue cases and DSI in Fujian province. (F) Time series of dengue cases and DSI in Zhejiang province. Blue lines represent time series of dengue case counts, and red lines represent time series of DSI, respectively.
Fig 6.
Comparison of prediction performance of the models including the support vector regression (SVR) model, step-down linear regression model, gradient boosted regression tree model (GBM), negative binomial regression model (NBM), least absolute shrinkage and selection operator (LASSO) linear regression algorithm and generalized additive model (GAM) by the means of root-mean-square error (RMSE).
Left panel: comparison of 1-month-ahead predictions for the 2014 outbreak in Yunnan, Guangxi, Hunan, Fujian and Zhejiang which pose a high risk of dengue infection. In each panel, the RMSE values of different forecast windows for each model are summarized and presented as box plots. Right panel: the actual trend of 2014 dengue epidemics in Yunnan, Guangxi, Hunan, Fujian and Zhejiang are shown. In each panel, the blue lines represent observed case counts, and the red lines denote model-based predicted values.