GPTCoach: Towards LLM-Based Physical Activity Coaching

Research Background and Problem

• Problem or Challenge: Approximately one-quarter of the global population does not meet the World Health Organization's recommended physical activity standards, with this proportion rising to nearly half in the United States. While health coaches can provide personalized support through face-to-face interactions, they are costly and difficult to scale. On the other hand, mobile health applications (e.g., fitness trackers) are accessible to the general public but are limited in capturing qualitative contexts such as users' goals, values, and past experiences, resulting in insufficient personalization.
• Significance: Effectively promoting physical activity is critical for global health. Providing scalable and cost-effective solutions to enhance the efficiency of personalized behavioral interventions is of great importance.
• Research Motivation and Related Work: Recent advances in large language models (LLMs) present new opportunities to address the aforementioned challenges. LLMs' capabilities in contextual adaptation, natural language interaction, and reasoning have the potential to fill the gap in personalized support left by traditional mobile health applications. However, existing LLM applications are primarily focused on health-related Q&A and insight analysis, falling short of guiding open-ended, personalized health behavior dialogues.

Solution

• Method or Solution: A health coaching chatbot named GPTCoach is proposed, inspired by Stanford's "Active Choices" health guidance program. GPTCoach is capable of:

  1. Understanding users' health data, motivations, and barriers through natural language conversations.
  2. Providing personalized physical activity recommendations.
  3. Utilizing data from wearable devices for quantitative analysis and visualization to support goal setting.

• Innovations:

  1. Incorporates Motivational Interviewing (MI) strategies, distinguishing itself from traditional Q&A interaction modes by adopting a non-directive, supportive tone.
  2. Implements a unique prompt chaining mechanism, including dialogue state management, MI strategy prediction, and tool invocation, ensuring interactions remain focused on health goals and leverage users' contextual information.
  3. Dynamically integrates qualitative information (e.g., users' goals and values) with quantitative data (e.g., fitness data) to provide personalized support.

• Implementation Steps:

  1. Design principles were summarized based on interviews with 12 health experts and 10 health service recipients.
  2. A GPT-4-powered chatbot was developed, featuring Q&A interactions, health data visualization, and tool invocation.
  3. The system was tested in a laboratory setting, completing a full health coaching conversation using participants' historical health data.

Research Outcomes

• Specific Results:

  1. GPTCoach demonstrated strong performance in psychological safety, comfort, and supportive tone, with users rating the system highly for "personalization" and "actionability."
  2. GPTCoach successfully collected rich qualitative information from users and provided personalized activity recommendations.

• Comparison with Existing Solutions:

  1. GPTCoach outperformed non-enhanced GPT-4 in implementing motivational interviewing strategies (e.g., 93% of GPTCoach's responses were consistent or neutral in behavior).
  2. Users perceived GPTCoach's interaction style as gentler and more targeted compared to existing health applications (e.g., Apple Fitness).

• Experimental or Evaluation Results:

  • A laboratory evaluation involving 16 participants revealed:
    • 92% of participants found GPTCoach's recommendations "thoughtful," and 87% felt supported.
    • Comparative tests with "Vanilla GPT-4" showed that GPTCoach avoided unnecessary persuasion and significantly increased open-ended questioning.
  • Users also identified areas for improvement, such as shortening response length and more proactively utilizing health data.

• Limitations and Future Directions:

  1. Limitations:
     • The current experiment only evaluated single-session interactions, leaving the potential impact of multiple sessions or long-term use on health behaviors unexplored.
     • Limited data analysis capabilities, with some instances of ineffective use of sensor data.
     • Lack of socioeconomic diversity in the sample (e.g., participants were predominantly well-educated iPhone users).
  2. Future Directions:
     • Conduct long-term, multi-stage coaching evaluations.
     • Introduce real-time intervention mechanisms, such as Just-In-Time Adaptive Interventions (JITAIs).
     • Employ more advanced multimodal models or on-demand retrieval of external health information databases (e.g., Retrieval-Augmented Generation methods).
     • Address system biases, privacy protection, and robustness in the face of potential ethical risks.

Conclusion

• This study proposes an LLM-based health coaching solution that effectively integrates users' qualitative and quantitative data, opening new possibilities in the field of health behavior support. The findings indicate that GPTCoach outperforms current mainstream methods in supportive dialogue, motivation enhancement, and personalization. Future research should further explore safety, privacy protection, and long-term effectiveness evaluations.