Exploring Personalized Health Support through Data-Driven, Theory-Guided LLMs: A Case Study in Sleep Health

Research Background and Issues

• Issues and Challenges:
  The authors point out that although sleep monitoring devices are widely available, users still struggle to translate data into actionable recommendations for improving sleep health. Current interventions can provide data-driven suggestions but often lack adaptability to real-life constraints and individual-specific contexts. Many existing template-based recommendations fail to flexibly address users' dynamic needs, resulting in limited engagement and effectiveness in behavior change.

• Importance of the Issue:
  Sleep is a critical factor for human health, influencing physical, cognitive, and emotional resilience. However, sustained health interventions aimed at optimizing sleep remain challenging due to difficulties in transforming complex sleep data into personalized and motivational behavioral guidance.

• Research Motivation and Related Work:
  The authors conducted a systematic analysis of existing research and technologies, including systems that provide sleep data insights using sleep monitoring devices (e.g., Fitbit and Oura Ring) and studies leveraging health behavior conversational technologies to promote user health interventions. However, these approaches suffer from poor flexibility, insufficient personalization, and a lack of timely recommendations, particularly in the domain of sleep health, where dynamic adaptive interventions guided by theoretical frameworks are lacking.

Solution

• Proposed Solution:
  The authors developed an intelligent health-support chatbot named HEALTH GURU. This system integrates large language models (LLM), wearable device data, contextual information, and behavior change theories to deliver personalized and context-adaptive sleep interventions. Additionally, it enhances user motivation through conversational interactions.

• Innovations:

  • Utilizes a multi-agent system framework to combine LLM's natural language processing capabilities with specific sleep health intervention goals.
  • Introduces a Contextual Multi-Armed Bandit (MAB) model to dynamically adjust health recommendations, leveraging validated effective activities while exploring potential new ones.
  • Enhances the scientific validity and user adaptability of recommendations through guidance from behavior change theories (e.g., self-efficacy and habit formation theories).

• Implementation Steps:

  1. Data Integration: Aggregates physiological data, sleep patterns, and activity records from wearable devices (e.g., Oura Ring).
  2. Context-Based Recommendation Model: The MAB model dynamically generates personalized activity suggestions, such as recommending suitable sleep-enhancing activities (e.g., running or meditation) based on weather and time.
  3. Application of Behavior Change Theories: Uses LLM technology to transform action recommendations into theory-guided conversational content, such as helping users set goals, providing feedback, and offering emotional support.
  4. Multi-Agent Architecture: Agents within the system handle data analysis, recommendation generation, and user interaction, ensuring comprehensive data synthesis and coherent responses.

Research Outcomes

• Specific Results:

  1. HEALTH GURU significantly improved users' sleep duration (average daily increase of 22 minutes) and activity scores (from 71.37 to 75.41).
  2. The system demonstrated superior performance in personalization, contextual awareness, and recommendation relevance compared to baseline systems.
  3. Users engaged in longer conversations with HEALTH GURU and exhibited higher interaction persistence.

• Advantages Over Existing Methods:

  • Health interventions are highly tailored to users' individual conditions and environments, rather than relying on static template-based traditional methods.
  • Provides a more natural conversational experience and real-time personalized feedback, significantly enhancing user motivation and engagement.

• Experimental or Evaluation Results:

  • Compared to baseline systems, HEALTH GURU achieved a significant improvement in user recommendation adherence (from 36.71% to 47.94%).
  • Higher system engagement, with the proportion of active user days increasing from 0.31 to 0.39.
  • Users reported better understanding of the relationship between sleep and activities and perceived the health recommendations as more contextually relevant.

• Limitations and Future Directions:

  1. Data Accuracy: The monitoring accuracy of Oura Ring is subject to certain limitations, leading to potential discrepancies between data and users' actual experiences.
  2. Contextual Constraints: Currently, only basic contextual factors such as weather and time are considered; future work could integrate more variables (e.g., work schedules, health conditions).
  3. Long-Term Intervention Effects: Due to the short experimental duration (eight weeks), the sustained impact of long-term use on behavior change remains unverified.
  4. Sample Diversity: The experiment participants were predominantly young adults; future studies should test applicability across broader demographics.
  5. Technical Improvements: Enhancing data visualization and multimodal interactions (e.g., voice control) could lower user barriers to adoption.

By introducing a dynamically adaptive health intervention mechanism and a theory-guided AI framework, HEALTH GURU provides new design insights and practical references for future LLM-based health applications.