FINDING POTENTIALLY UNSAFE NUTRITIONAL SUPPLEMENTS FROM USER REVIEWS WITH TOPIC MODELING

Area of Science:

• Pharmacovigilance
• Computational Linguistics
• Public Health

Background:

• Dietary supplements, widely used and generally safe, can cause severe or fatal adverse reactions.
• Current regulatory surveillance for supplement safety is not always effective.
• User-generated online reviews offer a potential data source for monitoring.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) system for monitoring dietary supplements using online reviews.
• To identify and categorize supplements based on potential danger using adverse reaction data.
• To assess the system's effectiveness compared to human annotators.

Main Methods:

• Utilized a variation of Latent Dirichlet Allocation (LDA) topic modeling on Amazon.com reviews for nutritional supplements.
• Integrated an adverse reaction dictionary to identify and score potential product dangers.
• Developed a scoring mechanism to categorize products into "high," "average," and "low" potential danger levels.

Main Results:

• The NLP system successfully generated topics semantically capturing adverse reactions from user reviews.
• The system categorized products based on potential danger.
• The system's categorization agreed with human annotators 69.4% of the time.

Conclusions:

• The developed NLP system demonstrates promise as a viable, low-cost, active approach for dietary supplement monitoring.
• Leveraging user-generated data and NLP techniques can enhance public health surveillance for supplement safety.
• This proof-of-concept system offers a novel method for proactive identification of at-risk dietary supplements.