Electronic Event–based Surveillance for Monitoring Dengue, Latin America

Dengue, a potentially fatal disease, is spreading around the world. An estimated 2.5 billion people in tropical and subtropical regions are at risk. Early detection of outbreaks is crucial to prevention and control of dengue virus and other viruses. Case reporting may often take weeks or months. Therefore, researchers explored whether electronic sources of real-time information (such as Internet news outlets, health expert mailing lists, social media sites, and queries to online search engines) might be faster, and they were. Although information from unofficial sources should be interpreted with caution, when used in conjunction with traditional case reporting, real-time electronic surveillance can help public health authorities allocate resources in time to avert full-blown epidemics.

every 2 years as a reference for health care providers who advise international travelers on health risks. The Yellow Book classifies regions of the world into dengue risk areas and areas with no known dengue risk on the basis of whether dengue is considered to be endemic in these areas.
This classification relies on expert-reviewed reports and peer-reviewed publications, as well as communications with subject-matter authorities. Areas are drawn at the scale of first-level administrative units (subnational regions such as state or province). For our study, we used only areas that had been labeled as no known risk but were adjacent to and contiguous with risk areas in the 2010 Yellow Book.
Because the Yellow Book is intended as information for clinicians and travelers and does not provide detailed information on the precise criteria that are used in making its classifications, we are limited in our ability to make inferences regarding the exact meanings of the 2 classifications. However, to our knowledge, the Yellow Book provides the most geographically comprehensive and most frequently updated dengue risk map currently available, which motivated its use as our reference map for this study.

Spatial Modeling
We fit a bivariate Gaussian mixture model to our dataset of HealthMap alerts. This is a statistical model of a probability density function made up of a weighted sum of Gaussian densities. While a common application of mixture modeling is cluster detection, mixture models are also used for density estimation. In a spatial context, this enables estimation of the underlying and unobservable continuous density of a set of observed points. Our candidate spatial models included the covariates latitude and longitude represented by Mollweide equal area-projected xand y-coordinates. Therefore, the computed models are weighted sums of component bivariate Gaussian distributions that represent the probability distribution of HealthMap observations in the study area. We fit bivariate Gaussian mixture models with 1, 2, …, 10 components and diagonal or spherical covariance functions and then selected the best fit model according to the Bayesian information criterion. We ensured that the best-fit model did not fall at the minimum or maximum number of components considered. Using the best-fit model, we then estimated the mean probability density for each first-level administrative unit contiguous with 2010 Yellow Book dengue-positive areas. This value that can be thought of as a model-based estimate of the intensity of HealthMap alerts for each subnational area.

Statistical Methods
Of included administrative areas, we identified all that had been changed from an area with no known dengue risk in the 2010 Yellow Book edition to dengue risk area in the 2012 Yellow Book edition. We hypothesized that these were areas into which dengue had spread between editions of the Yellow Book, and that the areas with the strongest HealthMap alert density would correspond to these areas of recent expansion. To test this hypothesis, we plotted receiver operating characteristic curves to evaluate the sensitivity and specificity of a range of threshold HealthMap alert density values for predicting the occurrence of new dengue risk areas in the 2012 Yellow Book and selected optimally predictive density threshold cutoffs by using the Youden statistic (4). To avoid over fitting, we performed receiver operating characteristics analysis with 5-fold cross validation.