Probabilistic emotion and sentiment modelling of patient-reported experiences

Area of Science:

• Natural Language Processing
• Healthcare Informatics
• Computational Linguistics

Background:

• Traditional patient surveys are limited in scope and costly.
• Unstructured online patient feedback offers rich insights but is challenging to analyze.
• Developing methods to interpret patient emotions from text is crucial for healthcare quality improvement.

Purpose of the Study:

• To present a methodology for interpretable probabilistic modeling of patient emotions from patient-reported experiences.
• To uncover key themes and their relationships with emotions in patient feedback.
• To develop a context-specific probabilistic emotion recommender system.

Main Methods:

• Metadata-network topic modelling was applied to 13,380 patient experiences from Care Opinion (2012-2022).
• A Naïve Bayes classifier was used to predict multi-label emotions and binary sentiments, with topics as predictors.
• Performance was assessed using nDCG, Q-measure, and F1 score, achieving 0.921.

Main Results:

• Patient emotions (positivity/negativity) are strongly linked to patient-caregiver interactions, not clinical outcomes.
• Positive patient experiences correlate with educational engagement.
• Negative experiences like dismissal and rejection are associated with suicidality and depression.

Conclusions:

• The proposed methodology offers a cost-effective, timely, and transparent approach to analyzing unstructured patient feedback.
• This method can augment traditional patient experience collection and provide emotional insights.
• The R package and dashboard facilitate accessibility for research and clinical practice, enabling tailored patient care.