Dengue fever, a vector-borne disease, is a major public health challenge. Accurate prediction methods that can better reflect the complexity of the outbreak are essential for dengue prediction and vector control. In this study, we introduce an adapted Spatio-Temporal Graph Convolutional Network (STGCN), originally developed for traffic forecasting, to predict weekly dengue cases in nine countries in South and Central America from 2014 to 2022. In this approach, we use environmental and socio-economic data in addition to climate data and historical dengue case information to capture complex transmission dynamics. We evaluate the STGCN against a Random Forest (RF) model using the same predictors. The evaluation results show that the STGCN model effectively captures outbreak dynamics and short-term trends. This was especially evident in cases where early transmission patterns are critical. In most of the countries analyzed, STGCN outperformed the baseline random forest model, especially in short-term forecasts, and achieved lower forecast errors in most settings. Forecasting performance varied across regions, with R-2 values ranging from 0.78 to 0.98 and RRMSE between 0.14 and 0.43 in short-term forecasts. The strength of the STGCN algorithm lies in its ability to capture spatio-temporal dependencies and handle heterogeneous data sources. This has been particularly valuable in areas with a high dengue burden. Although performance of the model varied slightly across countries, our overall findings highlight the robustness and adaptability of STGCN as a graph-based deep learning framework for dengue surveillance and early detection of its outbreaks.
