Improving topic modeling performance on social media through semantic relationships within biomedical terminology

Area of Science:

• Computational linguistics
• Health informatics
• Machine learning

Background:

• Topic modeling is used for social media analysis in healthcare.
• Social media data's unstructured nature poses challenges for traditional topic modeling.
• Existing methods struggle to extract meaningful health insights from noisy online text.

Purpose of the Study:

• To develop a novel semantic-type-based topic modeling pipeline.
• To enhance the discovery of self-reported health-related topics from social media.
• To overcome limitations of traditional topic modeling for analyzing messy online health discussions.

Main Methods:

• Developed a pipeline integrating semantic type information and Systematized Medical Nomenclature for Medicine (SNOMED) expressions.
• Applied traditional topic modeling to restricted medical concepts within social media texts.
• Used statin-related social media data for illustration and validation.

Main Results:

• The new approach yielded more novel, distinguishable, and meaningful health topics compared to traditional methods, based on expert evaluation.
• Validated a newly identified topic using electronic health records.
• Statin users showed a higher prevalence of depression or anxiety than non-users in clinical data.

Conclusions:

• The semantic-type-based topic modeling pipeline effectively extracts themes from noisy social media data.
• This method offers deeper insights for healthcare research by improving topic discovery.
• The findings highlight potential mental health impacts associated with statin use.