Predicting Dengue Outbreak Dynamics from Climate and Socio-Economical Data

Dengue fever represents a significant public health threat, particularly in tropical and subtropical regions. The incidence of this mosquito-borne viral infection has increased dramatically in recent decades, posing a considerable challenge to healthcare systems. Accurate and early prediction of dengue outbreaks is crucial for effective public health intervention, yet remains a challenge due to the complex spread factors. This paper addresses this challenge by developing and comparing two machine learning models for forecasting dengue incidence in several South American countries. We constructed a comprehensive dataset integrating heterogeneous data sources, including climatic, socio-economic, and historical epidemiological records. We propose and evaluate two distinct approaches: a stacking ensemble model composed of CatBoost regressors and a feed-forward Neural Network. The models were evaluated on an unseen test set composed by up to 14 months into the future. Our results demonstrate that the Neural Network generally achieves superior performance in terms of error metrics, whereas the ensemble model, while effective in identifying key predictive features such as temporal and environmental variables, showed limitations in accurately capturing the peak magnitude of outbreaks. Our findings highlights the potential of integrating heterogeneous datasets with advanced machine learning techniques for early dengue epidemic detection.