Global monitoring of COVID-19 vaccine hesitancy on social media with deep learning and natural language processing

Abstract

Background: COVID-19 vaccine hesitancy is a major global public health challenge. Discussions on social media can reflect public sentiment. Given the volume and velocity of data, machine learning approaches could effectively analyse the data in real-time, monitor topics of interest, and track trends in sentiment. Methods: Chinese and English posts on COVID-19 vaccines were extracted from Twitter and Weibo from September 2020 to May 2021. We compared Naïve Bayes, Support Vector Machine, Bidirectional Long Short Term Memory Model and Bidirectional Encoder Representations from Transformers Model (BERT) to evaluate relevance and vaccine sentiment. Hierarchical Dirichlet Process (HDP) was trained to model common tweet topics. Gross and by-topic trends in sentiment were summarised across 11 global locations. Results: We collected 3,740,935 posts on Weibo and 251,060,939 posts on Twitter. BERT outperformed other models (accuracy: 85.0%-94.9%). Hesitancy was most evident in China (19.8%), North America (10.8%; U.S.A and Canada), and Europe (12.1%; France), and lower in South America (7.0%; Argentina, Brazil, Chile, Peru, Colombia) and rest of Asia (4.9%; India and Thailand). Discussion on vaccine delivery was the most popular topic on both platforms. On Weibo, hesitancy was strongest over safety (22.0%), and hesitancy over trials increased over the study period (P=0.014). On Twitter, hesitancy over effectiveness (P=0.003) and safety (P=0.004) had the greatest declines over the observation period. Network analysis revealed relatively positive sentiments among academia and negative sentiments from media outlets and their influence on general users. Globally, hesitancy was positively associated with vaccine supply (P=0.004). Conclusions: Vaccine hesitancy has decreased notably on Twitter but increased mildly on Weibo over the study period. Doubts remained on many topics that varied by location. Network analysis revealed idea mixing patterns and may guide subsequent interventions. Social media data could provide dynamic insights for real-time tracking of vaccine hesitancy and guide future digital interventions.