Leveraging deep learning and social heterogeneity to detect early warning signals of disease outbreaks

ZitaoHe

University of Waterloo

"Leveraging deep learning and social heterogeneity to detect early warning signals of disease outbreaks"

Identifying early warning signals (EWS) of shifts in vaccinating behaviors can be helpful in predicting disease outbreaks. Evolutionary game theory has been used to model individual vaccination decisions, while bifurcation theory has identified statistical EWS, such as increasing variance and lag-1 autocorrelation, near critical transitions. However, these conventional methods often struggle with noisy data. In this study, we improve coupled behavior-disease models by incorporating population heterogeneity, distinguishing between social media users and non-users, and examining the role of homophily in shaping disease dynamics. We develop deep learning classifiers, including Long Short-Term Memory (LSTM) and Residual Neural Networks (ResNet), trained on simulated data from a stochastic coupled model with Lévy noise that captures the heavy-tailed fluctuations characteristic of real-world systems. Our results show that these models outperform traditional statistical indicators in both sensitivity and specificity while offering clearer interpretability on empirical data. This approach provides a robust framework for detecting EWS and improving outbreak prediction, highlighting the power of deep learning in real-time public health monitoring.
Additional authors:

---