Building Models For Risk Assessment Of Dengue Fever In The Pearl River Delta,China:A Study Based On The Fine Spatial Scale

Background: Dengue fever has been the major mosquito-borne infectious disease in Guangdong province since re-emerging at 1978.The heavy burden of dengue fever on health and economy calls for the accuracy tools to assess the risk of dengue fever.Method: The cases of dengue fever during 2007 to 2015 were collected and use to calculated the frequency,duration,intensity of dengue fever in each month and sub-county.After collect the monthly mean temperature,monthly precipitation,vegetation,water area,built-up area,road density,Gross Domestic Product(GDP)and population density information of each sub-counties by analyzing the Satellite Data,the random forest regression model(RFR)of frequency duration and intensity of dengue fever were built respectively.Based on the dataset between February of 2012 to September of 2014,the method of recursive feature elimination was applied to the primary feature selection of random forest,and the 5-fold cross-validation were used to find the parameter of models to build the RFRs of frequency,duration and intensity of dengue fever.And the dataset between October of 2014 to March 2015 to verified the accuracy of models.To a further comparison of accuracy of RFR,the support vector regression models(SVR),generalized addictive models(GAM)and generalized linear models(GLM),which were generally used to explore the risk of dengue fever in previous studies,were also built and the mean square error(MSE)were calculated.Result: We found the positive association between incidence and frequency,duration,intensity of dengue fever while the temporal-spatial distribution between those indexes show different characteristics.After built models with selected parameters,we found the pseudo R-squared of the models of frequency,duration and intensity were all above 90%,and the cross validation shows that the MSE of those models were 0.0019,1.4246 and1.8811.Moreover,we found the MSEs of RFR in test set were close to SVR but much less than tradition parameter models like GAM and GLM.After analyzing the distribution of errors of RFR,we found that the models show poor performance at 2014 and in the sub-county with severe dengue fever out-break,and the inclusion of dataset during severe outbreak may help to improve the accuracy.Conclusion: The RFR models based on the meteorological,environment and socioeconomic information had great performance on the accuracy.The index combined the frequency,duration and intensity of dengue fever can help the government and community to prevent and control the outbreak of dengue fever.