Developing Embedded Taxonomy and Mining Patients' Interests From Web-Based Physician Reviews: Mixed-Methods Approach

Area of Science:

• Health Informatics
• Natural Language Processing
• Social Sciences

Background:

• Web-based physician reviews offer valuable insights into patient experiences.
• Existing research has primarily focused on review text, neglecting systematic topic structures.
• Developing a robust topic taxonomy is crucial for effectively mining physician review data.

Purpose of the Study:

• To create a hierarchical topic taxonomy for physician reviews.
• To uncover the underlying structure and dimensions within patient feedback.
• To demonstrate the application of this taxonomy for analyzing patient interests.

Main Methods:

• Utilized a mixed-methods approach combining literature review, data-driven topic discovery, and human annotation.
• Developed a topic taxonomy with 3 domains and 9 subtopics from over 122,000 physician reviews.
• Applied a classification algorithm (Labeled-Latent Dirichlet Allocation) for topic identification.

Main Results:

• The taxonomy encompasses physician-related (ethics, competence, communication), patient-related (profile, symptoms, diagnosis), and system-related (financing, operations) domains.
• Achieved an average F-measure of 0.816 with the classification algorithm.
• Identified significant differences in patient interests based on disease type (acute vs. chronic) and severity (mild vs. serious).

Conclusions:

• The developed mixed-methods approach provides a rigorous framework for creating physician review taxonomies.
• The Labeled-Latent Dirichlet Allocation algorithm effectively mines patient interests from reviews.
• Analysis revealed distinct patient concerns varying by disease characteristics and healthcare context.