The Use of Natural Language Processing to Interpret Unstructured Patient Feedback on Health Services: Scoping Review.

Background: Unstructured patient feedback (UPF) allows patients to freely express their experiences without the constraints of predefined questions. The proliferation of online health care rating websites has created a vast source of UPF. Natural language processing (NLP) techniques, particularly sentiment analysis and topic modeling, are increasingly being used to analyze UPF in health care settings; however, the scope and clinical relevance of these technologies are unclear. Objective: This scoping review investigates how NLP techniques are being used to interpret UPF, with a focus on the health care settings in which this is used, the purposes for using these technologies, and any impacts reported on clinical practice. Methods: Searches of the MEDLINE, Embase, CINAHL, Cochrane Database of Reviews, and Google Scholar were conducted in February 2024. No date limits were applied. Eligibility criteria included English-language studies that used NLP techniques on UPF that pertained to an identifiable health care setting or providers. Studies were excluded if human actors solely performed coding or if NLP was applied to structured feedback or non-patient-generated content. Data were extracted and narratively synthesized regarding health care settings, NLP methods, and clinical applications. Results: From 4017 records, 52 studies met inclusion criteria. NLP was most commonly applied to UPF from secondary care settings (n=33) with fewer in primary (n=10) or community (n=5) care. Three NLP techniques were identified in the included studies: sentiment analysis (n=32), topic modeling (n=15), and text classification (n=7). Sentiment analysis was applied to explore associations between patient sentiment and health care provider characteristics, track emotional responses over time, and identify areas for improvement in health care delivery. Topic modeling, primarily using latent Dirichlet allocation algorithm, was used to uncover latent themes in patient feedback, compare patient experiences across different health care settings, and track changes in patient concerns over time. Text classification was used to categorize patient feedback into predefined topics. The association between NLP-derived insights and traditional health care quality metrics was limited, with few studies describing concrete clinical impacts resulting from their analyses. Conclusions: NLP has been applied to UPF across a number of contexts, primarily to identify features of health services or professionals that support good patient experience. The growth of research publications demonstrates an academic interest in these technologies, but there is little evidence these approaches are being used in clinical settings. Future research is required to assess how NLP may capture the nuance of health care interactions, align with existing quality metrics, and how it may be used to influence clinician behavior.