Probabilistic emotion and sentiment modelling of patient-reported experiences

Files

hdl_145051.pdf (1.89 MB)
  (Published version)

Date

2025

Authors

Murray, C.
Mitchell, L.
Tuke, J.
Mackay, M.

Editors

Advisors

Journal Title

Journal ISSN

Volume Title

Type:

Journal article

Citation

Artificial Intelligence in Medicine, 2025; 167:103178-1-103178-18

Statement of Responsibility

Curtis Murray, Lewis Mitchell, Jonathan Tuke, Mark Mackay

Conference Name

Abstract

Patient feedback is necessary to assess the extent to which healthcare delivery aligns with public needs and expectations. Surveys provide structured feedback that is readily analysed; however, they are costly, infrequent, and constrained by predefined questions, limiting a comprehensive understanding of patient experience. In contrast, the unstructured nature of online reviews and social-media posts can reveal unique insights into patient perspectives, yet that very lack of structure presents a challenge for analysis. In this study, we present a methodology for interpretable probabilistic modelling of patient emotions from patient-reported experiences. We employ metadata-network topic modelling to uncover key themes in 13,380 patient-reported experiences from Care Opinion (2012-2022) and reveal insightful relationships between these themes and labelled emotions. Our results show positivity and negativity relate most strongly to aspects of patient experience, such as patient-caregiver interactions, rather than clinical outcomes. Patient educational engagement exhibits strong positivity, whereas dismissal and rejection are linked to suicidality and depression. We develop a context-specific probabilistic emotion recommender system that predicts both multi-label emotions and binary sentiments with a Naïve Bayes classifier using topics as predictors. We assess performance with nDCG and Q-measure and achieve an F1 of 0.921, significantly outperforming standard sentiment lexicons. This methodology offers a cost-effective, timely, and transparent approach to harness unconstrained patient-reported feedback, with the potential to augment traditional patient-reported experience collection. Our R package and interactive dashboard make the approach readily accessible for future research and clinical practice applications, enabling hospitals to integrate emotional insights into surveys and tailor care to patient needs. Overall, this study provides a new avenue for understanding and improving patient experience and the quality of healthcare delivery.

School/Discipline

Dissertation Note

Provenance

Description

Available online 8 June 2025

Access Status

Rights

© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).

License

Call number

Persistent link to this record