Mohamad Al Bannoud, Jorge Luiz Mendes, Rodrigo Seiti Sacay, Tiago Dias Martins, Iara Rocha Antunes Pereira Bresolin, Joana Bratz Lourenço
Dengue remains one of the most critical vector-borne diseases globally, with Brazil consistently reporting the highest incidence in the Americas. The 2024 epidemic was the largest in the country's history, with São Paulo as its epicenter. This study developed and evaluated a machine learning (ML) framework to forecast dengue under extreme epidemic conditions, focusing on enhancing model robustness when testing data exceed the range of historical observations. A novel virtual data augmentation strategy was introduced to expose models to synthetic large-scale outbreaks and to assess the predictive value of urban incident variables such as flooding and flooded streets. Epidemiological, climatic, and urban incident data from 2014 to 2022 were used for model training, and 2023-mid-2025 for testing. Fourteen ML algorithms were systematically evaluated under multiple preprocessing configurations, including feature scaling, lagged predictors, feature combinations, data augmentation, and feature reduction. Forecasts targeted probable, confirmed, severe, and fatal dengue cases at a one-week horizon. The algorithms AdaBoost and CatBoost achieved the highest predictive accuracy. Virtual data augmentation substantially improved robustness when test values exceeded training maxima, reducing RMSE by up to 80% for probable and confirmed cases, 42% for severe cases, and 53% for deaths compared with models trained on unaugmented data. Flooding-related indicators, especially flooded-street case counts, were among the most influential predictors. The proposed ML framework accurately forecasted dengue incidence and severity during the unprecedented 2024 outbreak. The virtual augmentation method effectively managed distributional shifts and extreme epidemic magnitudes, providing a transferable foundation for early-warning systems and public-health preparedness in urban settings.