Predicting zip code-level vaccine hesitancy in US Metropolitan Areas using machine learning models on public tweets.

Area of Science:

• Public Health
• Computational Social Science
• Machine Learning

Background:

• Despite increased COVID-19 vaccine uptake, significant vaccine hesitancy persists in specific US populations.
• Traditional surveys for vaccine hesitancy are costly and lack real-time data.
• Social media offers a potential avenue for real-time aggregate-level vaccine hesitancy signals.

Purpose of the Study:

• To investigate the feasibility of using machine learning models with publicly available social media data to detect aggregate-level vaccine hesitancy.
• To rigorously evaluate and compare established machine learning models against non-adaptive baselines.
• To assess the performance of machine learning models in identifying geographic and demographic clusters of vaccine hesitancy.

Main Methods:

• Utilized publicly available Twitter data collected over one year.
• Applied established machine learning models, focusing on evaluation rather than novel algorithm development.
• Incorporated socioeconomic and other features from publicly accessible sources for model training.

Main Results:

• The best-performing machine learning models significantly outperformed non-learning baselines in detecting vaccine hesitancy signals.
• Demonstrated the feasibility of using social media data for real-time, aggregate-level monitoring of vaccine hesitancy.
• Showcased that these models can be implemented using open-source tools and software.

Conclusions:

• Machine learning models analyzing social media data provide a viable and effective method for monitoring COVID-19 vaccine hesitancy.
• This approach offers a scalable, cost-efficient, and timely alternative to traditional survey methods.
• The study validates the use of computational methods for understanding public health trends and informing targeted interventions.