Towards an Early Warning System for Climate-Sensitive Infectious Diseases: Spatio-Temporal Modeling and Deep Learning for Dengue Forecasting in Brazil
This dissertation develops integrated spatio-temporal forecasting approaches that combine deep learning, climate data, spatial dependencies, and human mobility to improve dengue prediction and support early-warning systems in Brazil.
Overview
Climate-sensitive infectious diseases such as dengue pose major challenges for public health surveillance due to their strong dependence on environmental variability, spatial heterogeneity, and human movement. Accurate forecasts are essential for outbreak preparedness, resource allocation, and timely intervention. This dissertation develops integrated spatio-temporal forecasting approaches by combining deep learning, climate information, spatial dependencies, and multimodal human mobility data to enhance dengue prediction and its operational usefulness in Brazil.
This research makes four key contributions. First, it provides a comprehensive comparison of statistical, machine learning, and deep learning models for dengue forecasting using data from Rio de Janeiro, establishing the strengths and limitations of existing approaches. Second, it extends forecasting nationwide by integrating SHAP-selected lagged climate variables and spatial effects into LSTM frameworks, demonstrating substantial improvements in predictive accuracy across Brazil’s 27 states. Third, it incorporates flight and road mobility data into deep learning models to quantify how population movement structures outbreak synchronicity and shapes the spatial spread of dengue. Fourth, it develops a nationwide importation-risk model that combines deep learning–based incidence forecasts with a composite multimodal mobility matrix to quantify directional infection pressure among Brazil’s 5,570 municipalities, identifying source–sink regions and dominant transmission corridors.
Across these contributions, the results demonstrate that deep learning models enriched with climate, spatial, and mobility information substantially improve dengue forecasting and provide a more realistic representation of inter-regional transmission pathways. By advancing spatio-temporal modeling for climate-sensitive diseases, this dissertation bridges the gap between methodological innovation and practical early-warning capabilities, offering scalable tools for epidemic preparedness and public health decision-making in Brazil.
Presenters
Brief Biography
Xiang Chen is a Ph.D. candidate in Statistics in the Geospatial Statistics and Health Surveillance (GeoHealth) Group at King Abdullah University of Science and Technology (KAUST), supervised by Prof. Paula Moraga.
He received his B.Eng. and M.S. degrees in Computer Science from Harbin Institute of Technology (HIT), where he worked on automated machine learning and data-driven modeling methods. During his doctoral studies, he has developed advanced statistical and deep learning frameworks for dengue forecasting across Brazil, incorporating climate variability, spatial dependence, and mobility patterns.
His work has been published in journals including BMC Public Health, Tropical Medicine and Health, and Infectious Disease Modelling. His research aims to bridge statistical methodology and real-world health surveillance to support early warning systems and data-driven policy planning.
