Analyzing Narratives of Patient Experiences: A BERT Topic Modeling Approach

Abstract

Due to healthcare systems increased focus on healthcare quality and patient- centered care, the patients’ perspective of delivered healthcare, has become an important part of healthcare service evaluations. Patient experiences can be used to improve the quality of care, as they reveal important information about health care encounters. An increasing number of organizations systematically collect and analyze patient experience data. The aim of our study was to identify major topics in narratives of patients’ healthcare related experiences and analyze the reactions of readers of patient experiences. 1663 blogs and 298806 textual comments were extracted on non-solicited patient experiences from a Hungarian online forum during a 10-year period. Topic modeling with state-of-the-art BERT embeddings were used to analyze the data and extract meaningful patterns and concepts. Sentiment analysis was utilized to categorize the emotional valence of the narrative writings. The huBERT and HIL-SBERT models identified 326 and 200 topics in terms of patient experiences and 508 and 728 topics regarding the reactions to these experiences without human supervision. Conceptually similar topics were integrated into major categories with manual analysis. 94.4% of the experiences and 77.5% of comments were classified as negative, reflecting the same annual tendency over the decade. Our study uses a data-driven approach for extracting patterns of healthcare related patient opinions, in Hungary. Topic modeling, based on BERT embeddings, could provide useful information on patient perceptions and perspectives, that could improve healthcare quality and safety.