Natural Language Processing Reveals Patient Reported Experiences

Patient-reported experiences play a significant role in healthcare performance monitoring and evaluation, and are critical to developing patient-centred care approaches. While quantitative approaches to uncovering patient experience through survey-based methods traditionally dominate healthcare analyses, qualitative analyses of experiences shared in free-text formats – such as those on social media and healthcare review websites – attract less attention despite their richness. This disparity may be largely attributed to the challenges of analysing vast corpora of unstructured text. This thesis presents several novel contributions to understanding patient-reported experiences using Natural Language Processing (NLP) techniques. We identify three key research gaps: the need for rapid insights in emerging health crises, the lack of a guiding NLP framework to uncover patient-reported experiences for healthcare professionals, and absence of contextually-aware fine-grained emotion modeling in healthcare settings. Our research employs topic modeling, sentiment analysis, and probabilistic modeling to address these gaps, summarising collective patient experience into themes that reveal patient perspectives on healthcare, explore how these themes relate to clinical outcomes and patient satisfaction, as well as using them to model patient emotions. Our first contribution addressed the need for rapid insights into patient experiences with COVID-19 during the initial stages of the novel public health crisis. Using NLP techniques, we revealed and explored timelines of symptom clusters, including the identification of then poorly recognised symptoms such as loss of smell and taste, which are now characteristic of COVID-19. This demonstrated the potential of NLP analysis of patient-data to provide rapid insights during emerging health crises. To reduce the technical gap for healthcare practitioners, we developed a comprehensive NLP framework for analysing patient-reported narratives. We applied this framework in a case study of prostate cancer experiences on Reddit, uncovering insights into sensitive and socially marginalising topics such as sexual dysfunction and incontinence. These findings were transformed into informative visualisations, enhancing understanding of these experiences and fostering community-building among affected individuals. Addressing the need for contextually-aware emotion modeling, we analysed narratives from the Australian healthcare review website Care Opinion, associating themes with labeled patient emotions. This revealed that aspects of patient experiences such as interactions with healthcare workers, rather than clinical outcomes, predominantly influence patient sentiments in experience narratives. We produced a landscape of patient emotions that provides insights into the relationships between intrinsic emotional states such as suicidal, that don’t reflect how the feeling arose, and extrinsically descriptive emotional states such as rejected. Building on this, we developed a fine-grained probabilistic emotion recommender system using topic modeling as a dimension reduction in a Naive Bayes framework. This system enables healthcare practitioners to augment free-text patient experiences with emotions, facilitating tailored patient-centred care interventions. To ensure accessibility and reproducibility, we developed this model into an online dashboard and R package. This thesis advances the methodology for analysing patient-reported experiences by providing a practical NLP framework and associated tools. It offers insights into diverse patient experiences, from COVID-19 symptoms to the emotional landscape of healthcare interactions. The developed techniques bridge the gap between complex NLP methods and their practical application in healthcare settings. By enabling the analysis of unstructured patient narratives at scale, this research contributes to a more comprehensive understanding of patient experiences. These contributions have the potential to inform patient-centered care approaches and healthcare policy, ultimately supporting more responsive and effective healthcare systems.

Author dc:creator

• Murray, Curtis William

Advisors dc:contributor.advisor

• Mitchell, Lewis
• Tuke, Jonathan
• Mackay, Mark (James Cook University)

• Natural Language Processing
• Patient-Reported Experiences
• Social Media
• Topic Modelling
• Sentiment Analysis

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/2440/145226

OAI identifier oai:identifier

oai:digital.library.adelaide.edu.au:2440/145226