Vaccine Adverse Event Mining of Twitter Conversations: 2-Phase Classification Study

Area of Science:

• Pharmacovigilance and Public Health Surveillance
• Computational Linguistics and Natural Language Processing
• Digital Epidemiology

Background:

• Traditional adverse event following immunization (AEFI) reporting systems face inherent delays in detecting safety signals.
• Early detection of AEFIs is crucial for timely public health interventions.
• Social media presents a promising, albeit largely untapped, resource for real-time safety monitoring.

Purpose of the Study:

• To evaluate the effectiveness of social media monitoring for early vaccine safety signal detection.
• To extract vaccine adverse event mentions (VAEMs) from Twitter using natural language processing (NLP).
• To document and identify optimal NLP techniques for VAEM extraction and social media surveillance.

Main Methods:

• Development of the VAEM-Mine method, integrating topic modeling and classification.
• Extraction of VAEMs from a large Twitter stream of vaccine-related posts.
• Identification of VAEM posts based on linguistic structure, not keyword matching.

Main Results:

• The VAEM-Mine method successfully isolated 8,992 VAEMs from 811,010 vaccine-related tweets.
• Achieved a high F1 score of 0.91 in the classification phase, indicating robust performance.
• Demonstrated the feasibility of large-scale social media data processing for pharmacovigilance.

Conclusions:

• Social media serves as a valuable complementary data source for detecting vaccine safety signals.
• A social media-based VAEM data stream can reveal emerging safety trends, mitigating limitations of passive reporting.
• This approach offers a more timely and comprehensive method for monitoring vaccine safety.