An Analysis of a Twitter Corpus for Training a Medication Intake Classifier

Area of Science:

• Computational linguistics
• Pharmacovigilance
• Health informatics

Background:

• Social media platforms are increasingly recognized as valuable data sources for medication-related research.
• Traditional observational studies of medications often rely on data sources other than social media.
• Advancing the use of social media data requires robust methods for analyzing medication information.

Purpose of the Study:

• To evaluate the performance of a machine learning classifier trained on a general corpus of medication-related tweets for detecting specific medication types.
• To identify medication classes where a general classifier performs suboptimally, necessitating specialized approaches.
• To inform the development of more accurate social media-based pharmacovigilance tools.

Main Methods:

• Analysis of an annotated corpus of 27,941 tweets related to medication intake.
• Training and evaluation of a baseline machine learning classifier on a diverse set of medication-related tweets.
• Comparative performance assessment of the classifier across different medication categories, with a focus on nervous system medications.

Main Results:

• The baseline classifier demonstrated significantly better performance for most medication types compared to its overall performance.
• A notable decline in classifier performance was observed for nervous system medications.
• This suggests that medication type significantly influences the effectiveness of general machine learning models.

Conclusions:

• A general machine learning model trained on social media data is useful for observational studies of most medication types.
• Studying nervous system medications using social media data may require developing classifiers trained exclusively on this specific medication class.
• Future work will explore data-level approaches for specialized medication classification.