[论文解读] GPTCoach: Towards LLM-Based Physical Activity Coaching

Matthew Jörke, Shardul Sapkota+5 more|arXiv (Cornell University)|May 9, 2024

Digital Mental Health Interventions被引用 8

一句话总结

GPTCoach 是一个基于 LLM 的聊天机器人，实施循证健康教练计划，使用动机性访谈策略，并能够查询可穿戴健康数据以支持身体活动行为改变；以技术探针的方式进行评估，共有 16 名参与者。

ABSTRACT

Mobile health applications show promise for scalable physical activity promotion but are often insufficiently personalized. In contrast, health coaching offers highly personalized support but can be prohibitively expensive and inaccessible. This study draws inspiration from health coaching to explore how large language models (LLMs) might address personalization challenges in mobile health. We conduct formative interviews with 12 health professionals and 10 potential coaching recipients to develop design principles for an LLM-based health coach. We then built GPTCoach, a chatbot that implements the onboarding conversation from an evidence-based coaching program, uses conversational strategies from motivational interviewing, and incorporates wearable data to create personalized physical activity plans. In a lab study with 16 participants using three months of historical data, we find promising evidence that GPTCoach gathers rich qualitative information to offer personalized support, with users feeling comfortable sharing concerns. We conclude with implications for future research on LLM-based physical activity support.

研究动机与目标

• 识别健康专家如何进行教练以克服身体活动障碍，以及 LLMs 能为这些策略贡献哪些内容。
• 评估自我跟踪数据如何用于促进活动，以及 LLMs 如何利用此类数据进行教练。
• 设计一个以既定的教练计划和 MI 技术为基础的促进性、非规范性 AI 教练。
• 评估 GPTCoach 对教练原则的遵循情况及其在真实对话中的数据使用。

提出的方法

• 对 12 名健康专家和 10 名非专家进行形成性访谈，以提取面向 LLM 健康教练的设计考虑。
• 将 GPTCoach 发展为与经验证的健康教练计划和动机性访谈技术保持一致的入职对话。
• 实现一个数据与提示管线，使用工具调用通过 HealthKit 获取可穿戴数据并在 UI 中可视化。
• 采用提示链以确保遵循教练计划、MI 策略和适当的数据使用。
• 原型设计并以 16 名参与者进行试点测试，以评估 MI 行为、教练遵循情况和数据利用情况。

实验结果

研究问题

• RQ1 RQ1：健康专家使用哪些教练策略，哪些可以被 LLM 采用以克服身体活动的障碍？
• RQ2 RQ2：健康专家如何使用自我跟踪数据，LLMs 又如何利用这些数据来促进活动？
• RQ3 RQ3：可以在集成个人数据的同时，LLM 基于教练是否能够维持促进性、非评判性的教练风格？
• RQ4 RQ4：在 LLM 中提示链接（prompt chaining）多大程度上可以强制遵循结构化教练计划？
• RQ5 RQ5：在数据使用和个性化方面，使用 LLM 进行健康教练有哪些风险和局限？

主要发现

• MI-consistent or neutral behaviors occurred 84% of the time in automated MI coding.
• Participants reported feeling supported and comfortable sharing concerns with the chatbot.
• Prompt chaining helped GPTCoach adhere to the coaching program and initiate appropriate tool calls.
• Data use by GPTCoach was more variable, with some conversations leveraging data for motivation and others not proactively integrating data.
• Compared to vanilla GPT-4, GPTCoach showed greater alignment with MI principles, asking more open questions and giving less unsolicited advice.
• Participants believed AI could augment data analysis for goal setting and accountability, but privacy and personalization challenges remained.