Twitter conversations predict the daily confirmed COVID-19 cases

Area of Science:

• Computational Social Science
• Epidemiology
• Data Science

Background:

• The COVID-19 pandemic spurred increased social media engagement, particularly on platforms like Twitter.
• Socially generated conversations offer valuable insights for situational awareness during crises.
• Accurate early forecasting of COVID-19 cases is crucial for resource allocation and public health management.

Purpose of the Study:

• To develop a methodology for incorporating public discourse from social media into COVID-19 forecasting models.
• To specifically target forecasting for the steep-hill phase of pandemic waves.
• To enhance the prediction accuracy of epidemiological models by leveraging social media data.

Main Methods:

• A sentiment-involved topic-based latent variables search methodology was proposed.
• The methodology was applied to Australian COVID-19 daily case data and associated Twitter conversations.
• Latent social media variables were identified for their predictive power.

Main Results:

• Latent social media variables were found to Granger-cause daily COVID-19 confirmed cases.
• These variables demonstrated additional prediction capability beyond traditional forecasting models.
• Inclusion of social media variables led to 48.83%-51.38% RMSE improvement over baseline models.
• A large-scale, geotagged global tweets dataset (MegaGeoCOV) was released.

Conclusions:

• Public discourse on social media contains predictive signals for COVID-19 case numbers.
• Sentiment and topic analysis of social media data can significantly enhance epidemiological forecasting.
• The released MegaGeoCOV dataset can facilitate further research into pandemic-related conversational dynamics.