Discovering health topics in social media using topic models

Area of Science:

• Computational Social Science
• Public Health Informatics
• Natural Language Processing

Background:

• Social media platforms like Twitter generate vast amounts of user-generated health-related content.
• Understanding public health discussions on social media is crucial for public health surveillance and intervention.
• Previous methods often require significant human supervision or historical data for topic modeling.

Purpose of the Study:

• To develop and evaluate a topic modeling framework for discovering health topics discussed on Twitter.
• To characterize the variety of health information shared by users on social media.
• To assess the ability of a general-purpose model to identify diverse health topics with minimal supervision.

Main Methods:

• Developed the Ailment Topic Aspect Model (ATAM), a statistical topic model tailored for health discussions.
• Implemented a system to filter general Twitter data using health keywords and supervised classification.
• Applied ATAM to 144 million Twitter messages from 2011-2013 to infer health-related topics.

Main Results:

• ATAM identified 13 coherent clusters of health-related topics from the Twitter data.
• Discovered topics showed statistically significant correlations with external data, including seasonal influenza (r=0.689), allergies (r=0.810), exercise (r=0.534), and obesity (r=-0.631).
• The model achieved these correlations with minimal human supervision and no historical training data.

Conclusions:

• Automated topic discovery from social media is feasible and can yield statistically significant correlations with real-world health data.
• The ATAM framework demonstrates the potential of a single, general-purpose model to identify a wide range of health topics.
• This approach offers a scalable method for public health surveillance and understanding online health discourse.