Heterogeneity and homophily in coupled behavior-disease dynamics: from model structure to early warnings

Abstract

Understanding how human behavior and infectious disease dynamics interact is essential for anticipating and mitigating outbreaks. While coupled behavior-disease models have provided valuable insights into the feedback between disease transmission and vaccination behavior, many assume homogeneous populations and neglect the influence of social structure in shaping individual vaccination strategies. Traditional surveillance systems often lack timely data on vaccination behavior, making it difficult to monitor changes in public vaccine sentiment. Moreover, existing statistical methods for detecting early warning signals of critical transitions rely on assumptions that do not always hold in real-world settings. This thesis addresses these limitations by incorporating population heterogeneity and homophily into a coupled behavior-disease model, and by using the resulting simulations to support the training of data-driven models for forecasting outbreak risks from high-frequency social media data. Specifically, we develop a coupled behavior-disease model that distinguishes social media users from non-users, capturing indirect heterogeneity in how individuals access vaccine-related information. The model demonstrates that homophily slows the spread of pro-vaccine strategies, pushing the population closer to tipping points. It also suggests that early vaccine-related online discussions may offer predictive signals of future outbreaks. Building on these findings, we generate synthetic time series with heavy-tailed noise to mimic real-world social media data. These model-generated data are used to train deep learning classifiers, under CNN-LSTM and ResNet architectures, to detect early warning signals in social media data. These classifiers outperform conventional statistical indicators, such as variance and lag-1 autocorrelation, in both sensitivity and specificity. Finally, we extend the modeling framework to a generalized multi-group vaccination game, considering direct heterogeneity in levels of vaccine support. Simulations reveal that homophily contributes to the persistence of opinion polarization in the population, regardless of the presence of diseases. Together, these studies highlight the need to account for heterogeneity in modeling vaccination behavior and that homophily can have various effects depending on the states of the system. We also show that combining mechanistic models and data-driven techniques can help detect emerging risks of disease outbreaks, informing more proactive public health policies.