Background: Patient-provider interactions could inform care quality and communication but are rarely leveraged because collecting and analyzing them is both time-consuming and methodologically complex. The growing availability of large language models (LLMs) makes these analyses more feasible, though their accuracy remains uncertain. Objectives: Assess an LLM's ability to analyze patient-provider interactions. Design: Compare a human's and an LLM's codings of clinical encounter transcripts. Setting/Subjects: Two hundred and thirty-six potential symptom discussions from transcripts of clinical encounters with 92 patients living with cancer in the mid-Atlantic United States. Transcripts were analyzed by GPT4DFCI in our hospital's Health Insurance Portability and Accountability Act compliant infrastructure instance of GPT-4 (OpenAI). Measurements: Human and an LLM-coded transcripts to determine whether a patient's reported symptom(s) were discussed, who initiated the discussion, and any resulting recommendation. We calculated Cohen's κ to assess interrater agreement between the LLM and human and qualitatively classified disagreements about recommendations. Results: Interrater reliability indicated "strong" and "moderate" agreement levels across measures: Agreement was strongest for whether the symptom was discussed (k = 0.89), followed by who initiated the discussion (k = 0.82), and the recommendation provided (k = 0.78). The human and LLM disagreed on the presence and/or content of the recommendation in 16% of potential discussions, which we categorized into nine types of disagreements. Conclusions: Our results suggest that LLMs' abilities to analyze clinical encounters are equivalent to humans. Thus, using LLMs as a research tool may make it more feasible to analyze patient-provider interactions, which could have broader implications for assessing and improving care quality, care inequities, and provider communication.
Fenton et al. (Wed,) studied this question.