Leveraging natural language processing to understand financial toxicity in cancer care: A thematic and sentiment analysis of online patient narratives.

Abstract

e13602 Background: Financial toxicity associated with cancer treatment significantly impacts patient outcomes and quality of life. Traditional research methods, such as surveys and interviews, may not fully capture the nuanced experiences of patients. Online platforms provide a valuable source of unstructured data where patients and caregivers share their challenges. This study evaluates the effectiveness of Natural Language Processing (NLP) techniques in analyzing these narratives to identify key themes and emotional responses related to financial toxicity. Methods: Data were collected from the "Cancer" and "Health Insurance" subreddits. Regular expressions were applied to filter posts related to cancer and financial burden, resulting in 217 patient and caregiver narratives. Narratives underwent cleaning and tokenization using the Natural Language Toolkit. Annotators reviewed the dataset to exclude unrelated sentences before further processing. A total of 411 sentences were prepared for analysis. Sentiment analysis was conducted using a pretrained RoBERTa-based model, and results were manually verified. Annotators classified sentences, compared outputs with model predictions, and refined sentiment classifications where discrepancies were noted. Emotion analysis of Negative sentiment sentences was conducted using a pretrained DistilRoBERTa-based model to identify prevalent emotional responses. Results: Of the 411 analyzed sentences, 134 (32.60%) were classified as Negative, 255 (62.04%) as Neutral, and 22 (5.36%) as Positive. Annotators identified and corrected 19 misclassified sentences that the model misclassified as Neutral but were found to contain mixed sentiments relevant to financial toxicity. Thematic categorization of Negative sentences revealed key challenges, with Insurance Barriers being the most prevalent theme (56.9%), followed by Employment and Disability Challenges (15.6%), Cost of Treatment (12.6%), Systemic Issues to Access Care (10.8%), and Other (4.2%). Emotion analysis of Negative sentences revealed sadness (62.87%) as the predominant emotion, particularly in discussions around Insurance Barriers and Employment Challenges, followed by fear (12.57%). Conclusions: This study demonstrates the potential of NLP techniques to enhance qualitative research by efficiently analyzing unstructured data. By identifying key financial toxicity themes and associated emotional burdens, this research highlights areas requiring targeted interventions. These findings provide a scalable approach to understanding financial toxicity across diverse patient populations, aiding healthcare providers in developing equitable, patient-centered solutions. Future advancements, such as fine-tuning NLP models, can improve the accuracy and applicability of these methods in oncology research.