Purpose: Aedes aegypti is a mosquito responsible for transmitting mainly dengue, zika, and chikungunya. In low- and middle-income countries, controlling the spread of this mosquito poses a major public health challenge. Currently, Aedes aegypti control policies are extremely important for lowering the risk of potential arbovirus outbreaks. One of the effective strategies for combating the burden of mosquito-borne arboviruses are the pre-emptive predictions and forecasts for future outbreaks. In this sense, we, therefore, apply machine learning using a spatiotemporal approach to build distribution maps of Aedes aegypti breeding sites in Recife.
Methods: We obtained data from Aedes aegypti breeding sites and climatic factors in Recife City during 2013-2014. From the information of the breeding sites, bimonthly spatial distribution maps of the breeding sites were generated using the Inverse Distance Interpolation (IDW). We generated monthly spatial distribution maps of climatic variables (temperature, rainfall, and wind speed) using the same method. From the generated distribution maps, several models were evaluated, among them: support vector regressor, multilayer perceptron, random forest, and linear regression. The model performance was evaluated according to Pearson's correlation coefficient and percentage root relative squared error (RRSE%) metrics.
Results: Among the evaluated regressors, the 3-degree polynomial-kernel support vector regressor showed superior performance compared to the other regressors evaluated. For this regressor, the correlation coefficient was on average 0.9875 (and standard deviation of 0.01) while the RRSE% metric was on average 14.60%.
Conclusion: Machine learning proved to be a promising tool in predicting the Aedes aegypti breeding sites’ spatial distribution in the city of Recife. The spatiotemporal predictions pointed out that neighborhoods with low income and lack of water supply are presented with elevated concentrations of mosquito breeding sites. The findings of this work can support health authorities in decision-making linked to policies for reducing the burden of mosquito infestation, while at the same time allowing authorities to optimize their limited resources in low-resource settings.