Forecasting infectious disease outbreak risks from vaccine sentiments on social media: A data-driven dynamical...

Early warning signals are vital in predicting critical transitions in complex dynamical systems. For behavioral epidemiology systems in particular, this includes shifts in vaccine sentiments that may precede disease outbreaks. Conventional statistical indicators, such as variance and lag-1 autocorrelation, often struggle in noisy environments and may fail in real-world scenarios. In this study, we leveraged universal signals of critical slowing down to train deep learning classifiers, specifically using long short-term memory (LSTM) and residual neural network (ResNet) architectures, for detecting early warning signals in disease-related social media time series. These classifiers were trained on simulated data from a stochastic coupled behavior-disease model with additive Lévy noise, a non-Gaussian noise that better reflects the heavy-tailed nature of real-world fluctuations. Our results show that these classifiers consistently outperform conventional indicators in both sensitivity and specificity on theoretical data while delivering quantitatively clear results that are easier to interpret on empirical data. Integrating deep learning with real-time social media monitoring offers a powerful tool for preventing disease outbreaks through proactive public health interventions.