Disease Nowcasting Using Integrated and Adaptive Statistical Models

Overview

Timely disease surveillance is essential for monitoring trends, detecting outbreaks, and informing public health action, yet reporting delays often prevent surveillance systems from reflecting the true current burden of disease. This dissertation develops statistical nowcasting methods to address reporting delays and produce more timely estimates of disease activity.

It proposes a series of Bayesian nowcasting frameworks that improve real-time prediction under delayed and incomplete reporting. First, it shows that digital behavioral signals, such as Google search trends, can complement traditional surveillance data and substantially improve dengue nowcasts. Second, it introduces a flexible Bayesian modeling framework for reporting delays that captures time-varying and non-stationary reporting processes through stochastic delay dynamics, allowing the model to adapt to both gradual and abrupt changes in surveillance systems while providing full uncertainty quantification. Third, it extends this framework by imposing structure on latent disease intensity, incorporating seasonal patterns, autoregressive dependence, and external covariates to strengthen predictive performance, especially when recent reported data are sparse or delays are severe.

These methods are validated through simulation studies and real-world applications, including dengue surveillance in Brazil, severe acute respiratory infections in Brazil, and HUS-O104 cases in Germany. In addition, this dissertation presents Dengue Tracker, a state-level nowcasting and visualization platform for Brazil that demonstrated the practical value of real-time statistical surveillance during the 2024 dengue outbreak.

Overall, this dissertation provides methodological and applied contributions toward more reliable, timely, and decision-relevant disease surveillance, with future extensions aimed at capturing spatial heterogeneity in real-time epidemic monitoring.