Social Media Can Warn of Declining Vaccine Support Before Outbreaks

By Pooja Toshniwal PahariaReviewed by Lauren HardakerNov 26 2025

A new study reveals that tracking real-time pro-vaccine discussions on Twitter can identify hidden behavioral tipping points long before measles outbreaks resurface, offering a powerful tool for public health teams trying to stay ahead of outbreaks.

Image credit: vectorfusionart/Shutterstock.com

A recent study published in Mathematical Biosciences and Engineering introduces a novel approach to forecasting outbreaks by combining deep learning (DL) with real-time social media monitoring. 

By tracking changes in vaccine sentiment on platforms like Twitter, the researchers highlight that patterns in pro-vaccine engagement can serve as an early warning signal (EWS) for reduced vaccine uptake. This work bridges simulated and real-world data, underscoring the importance of behavioral signals in disease prevention.

Early Warnings Hidden In Vaccin Discussions

Maintaining high vaccination coverage is crucial for preventing the resurgence of infectious diseases, yet shifts in public sentiment can cause uptake to decline long before outbreaks occur. Traditional surveillance often misses these early changes because vaccination data are infrequent and early warning indicators perform poorly in noisy, real-world settings.

In contrast, social media is a rapid reflection of evolving vaccine attitudes, with online discussions frequently revealing early declines in support for vaccination. Advancements in mathematical modeling and DL make it possible to use these data streams for detecting subtle behavioral shifts in a simulated setting before immunization levels fall below herd-level immunity thresholds, in theory.

The Behavior–Disease Framework 

In the present study, researchers developed a behavior-informed epidemiological model that applies DL to detect EWS by analyzing time series of pro-vaccine social media activity. The approach combines simulated outputs from stochastic behavior–disease modeling with Twitter data obtained before major measles surges. The model includes user and non-user groups and incorporates additive Lévy noise to capture heavy-tailed fluctuations.

To create training datasets, the team generated 10,000 positive (transition) and 10,000 neutral samples using simulations. A separate balanced set of 500 positive and 500 neutral samples across five noise amplitudes was used for ROC evaluation. Each sample contained 500 data points, equivalent to approximately 1.5 years of daily observations. The researchers calculated statistical features across the pre-transition period and within partial windows up to 100, 200, and 400 time points before the behavioral shift. The simulations enabled the DL models to capture signatures of critical slowing down before sentiment declines.

The team developed two classifiers: a Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) and a ResNet-based classifier. The CNN–LSTM model comprises a one-dimensional convolutional layer supported by dropout regularization, max pooling, two stacked LSTM layers, and a dense output layer. The ResNet model has three residual blocks. Steps included batch normalization, Rectified Linear Unit (ReLU) activation, dropout, and a global averaging task applied across feature maps to generate a sigmoid output. Both classifiers were optimized over 300 epochs, with a loss function suited for binary classification, with training parameters set to a 0.0005 learning rate and 1,024-sample batches.

For empirical validation, researchers collected English-language tweets referencing “measles” or “MMR” (2011–2019). After cleaning and geocoding, tweets were labeled as neutral, pro-, or anti-vaccine using COVID-TwitterBERT. The classifiers analyzed the smoothed time series of pro-vaccine posts, and Gaussian processing was used to filter the prediction trajectories.

The team evaluated model performance using four major North American measles outbreaks (California, New York, Washington, and British Columbia) as positive cases and four low-incidence countries with mandatory vaccination (Singapore, Mexico, Argentina, and Brazil) as neutral controls. They compared the results with traditional EWS, such as variance and lag-1 autocorrelation.

AI Outperforms Classic Warning Tools 

Across all analyses, the DL classifiers, LSTM and ResNet, significantly outperformed both traditional EWS methods. On simulated pre-transition data, both models demonstrated near-perfect discrimination, with area under the curve (AUC) values approaching 1.0. LSTM generally delivered the strongest performance but showed reduced accuracy at the longest (400-point) window for negative samples, whereas ResNet produced more stable predictions across cut-offs.

CNN–LSTM modeling achieved 86.3 % precision and 99 % recall for positive cases, 99 % precision and 84 % recall for negative cases, and an overall accuracy of 92 %. For positive cases, ResNet achieved 96 % precision and 76 % recall. For negative cases, it achieved 80 % precision and 97 % recall. Overall classification was 87 %. Both exceeded the sensitivity and specificity of variance- and autocorrelation-based EWS.

Both DL models, when applied to real-world measles outbreaks in California, New York, Washington, and British Columbia, showed rising transition probabilities consistent with emerging behavioral shifts. ResNet was more sensitive to early empirical signals, while LSTM responded faster in simulated settings and better filtered short-lived spikes that could trigger false alarms.

In data-sparse settings, such as British Columbia, ResNet maintained low transition probabilities. In contrast, LSTM produced temporary increases driven by short-lived bursts of discussion that did not correspond to sustained changes in sentiment. In low-incidence countries with mandatory vaccination (Singapore, Mexico, Argentina, Brazil), both classifiers predicted persistently low transition probabilities, aside from brief, non-sustained increases during the 2015 Disneyland outbreak.

Overall, the findings suggest that DL models may provide earlier and more reliable warnings of declining vaccine sentiment than conventional EWS indicators, highlighting strong potential for real-time detection of behavioral tipping points before outbreaks occur. By integrating simulated behavior–disease dynamics with high-frequency social media data, this framework enhances the anticipation of vaccine uptake collapse.

Challenges, such as limited multilingual data, potential overfitting to synthetic patterns, and short empirical time series, remain; however, advances in transfer learning can improve performance, supporting earlier, targeted public health interventions to prevent disease resurgence.

Journal Reference

Zitao He and Chris T. Bauch. (2025). Forecasting infectious disease outbreak risks from vaccine sentiments on social media: A data-driven dynamical systems approach. Mathematical Biosciences and Engineering, 22(10):2761–2779. DOI: 10.3934/mbe.2025101. https://aimspress.com/aimspress-data/mbe/2025/10/PDF/mbe-22-10-101.pdf