Advancing health coaching: A comparative study of large language model and health coaches.

Objective: Recent advances in large language models (LLM) offer opportunities to automate health coaching. With zero-shot learning ability, LLMs could revolutionize health coaching by providing better accessibility, scalability, and customization. The aim of this study is to compare the quality of responses to clients' sleep-related questions provided by health coaches and an LLM.

Design, Setting, And Participants: From a de-identified dataset of coaching conversations from a pilot randomized controlled trial, we extracted 100 question-answer pairs comprising client questions and corresponding health coach responses. These questions were entered into a retrieval-augmented generation (RAG)-enabled open-source LLM (LLaMa-2-7b-chat) to generate LLM responses. Out of 100 question-answer pairs, 90 were taken out and assigned to three groups of evaluators: experts, lay-users, and GPT-4. Each group conducted two evaluation tasks: (Task 1) a single-response quality assessment spanning five criteria-accuracy, readability, helpfulness, empathy, and likelihood of harm-rated on a five-point Likert scale, and (Task 2) a pairwise comparison to choose the superior response between pairs. A suite of inferential statistical methods, including the paired and independent sample t-tests, Pearson correlation, and chi-square tests, were utilized to answer the study objective. Recognizing potential biases in human judgment, the remaining 10 question-answer pairs were used to assess inter-evaluator reliability among the human evaluators, quantified using the interclass correlation coefficient and percentage agreement metrics.

Results: Upon exclusion of incomplete data, the analysis included 178 single-response evaluations (Task 1) and 83 pairwise comparisons (Task 2). Expert and GPT-4 assessments revealed no discernible disparities in health coach and LLM responses across the five metrics. Contrarily, lay-users deemed LLM responses significantly more helpful than that of human coaches (p < 0.05). LLM responses were preferred in the majority (62.25 %, n = 155) of the aggregate 249 assessments, with all three evaluator groups favoring LLM over health coach inputs. While GPT-4 rated both health coach and LLM responses significantly higher than experts in terms of readability, helpfulness, and empathy, its ratings on accuracy and likelihood of harm aligned with those of experts. Response length positively correlated with accuracy and empathy scores, but negatively affected readability across all evaluator groups. Expert and lay-user evaluators demonstrated moderate to high inter-evaluator reliability.

Conclusion: Our study showed encouraging findings by demonstrating that RAG-enabled LLM has comparable performance to health coaches in the domain tested. Serving as an initial step towards the creation of more sophisticated, adaptive, round-the-clock automated health coaching systems, our findings call for more extensive evaluation which could assist in the development of the model that could in the future lead to potential clinical implementation.