Weakly supervised text classification on free-text comments in patient-reported outcome measures.

Area of Science:

• Computational Linguistics
• Health Informatics
• Machine Learning

Background:

• Patient-reported outcome measures (PROMs) capture valuable health-related quality of life (HRQoL) data through free-text comments.
• Manual analysis of these comments is labor-intensive and time-consuming.
• Existing machine learning methods often require extensive labeled data and post-analysis interpretation.

Purpose of the Study:

• To evaluate the effectiveness of five weakly supervised text classification (WSTC) techniques for analyzing PROMs comments.
• To identify health-related quality of life (HRQoL) themes in cancer patient data using WSTC.
• To assess the performance and interpretability of WSTC methods with limited labeled data.

Main Methods:

• A scoping review identified key HRQoL themes and keywords.
• Five keyword-based WSTC methods (anchored CorEx, BERTopic, Guided LDA, WeSTClass, X-Class) were applied to colorectal and prostate cancer PROMs datasets.
• Performance was evaluated by overall and theme-specific metrics, with domain expert review for interpretability.

Main Results:

• Six main HRQoL themes were identified: Comorbidities, Daily Life, Health Pathways and Services, Physical Function, Psychological and Emotional Function, and Social Function.
• Method performance varied, with anchored CorEx achieving weighted F1 scores of 0.57 (colorectal) and 0.61 (prostate).
• Individual themes reached F1 scores up to 0.92, and methods utilizing expert seed terms and extrapolating from limited data performed best.

Conclusions:

• Keyword-based WSTC methods show significant potential for analyzing PROMs comments, particularly when labeled data is scarce.
• The study highlights both the capabilities and limitations of WSTC in accurately classifying HRQoL themes from patient-generated text.
• WSTC offers a promising avenue for efficient and scalable analysis of qualitative patient feedback in healthcare research.