Narrating Fitness: Leveraging Large Language Models for Reflective Fitness Tracker Data Interpretation

Title of the Paper

Narrating Fitness: Leveraging Large Language Models for Reflective Fitness Tracker Data Interpretation

Paper Information

• Subject Area: Human-Computer Interaction (HCI), Personal Informatics, Generative AI
• Keywords: Personal Informatics, Generative AI, Fitness Tracker, Reflection, User Experience, Data Visualization, Storytelling

Research Background and Problem

• Identified Problem: Fitness trackers primarily generate and display quantitative data, while users often prefer to understand their health status in qualitative terms. This "data-perception" gap may hinder users' ability to effectively reflect on their health, limiting the potential of personal informatics tools. Moreover, most current tools are goal-oriented and lack a genuine focus on users' overall well-being.
• Significance: Engaging users and enabling them to better understand health information through data reflection can drive more meaningful health behavior changes. At the same time, generative AI (e.g., Large Language Models, LLMs) offers new possibilities for enhancing user reflection through textual feedback, but this direction has yet to be sufficiently explored through empirical research.
• Research Motivation: To explore how LLM-generated narratives can enhance users' attention to and reflection on their personal fitness data, thereby identifying new methods to improve the design of personal informatics systems.

Solution

• Proposed Method or Solution:

  • This paper utilizes LLMs (e.g., GPT-4) to generate qualitative health narratives and compares them with conventional graphical data representations.
  • It investigates the impact of three conditions—"text representation," "graphical representation," and "text + graphical combined representation"—on user reflection and engagement.
  • A research platform was designed, allowing users to upload fitness step data and receive narrative feedback generated by LLMs.

• Innovations:

  1. Proposes the use of generative AI to create dynamic textual feedback for personal fitness data, bridging the gap between quantitative data and users' qualitative perceptions.
  2. Emphasizes non-goal-oriented narrative design to reduce the impact of negative feedback while enhancing users' insights into their health data.
  3. Combines quantitative and qualitative approaches, providing the first empirical analysis of the applicability of AI-generated text in personal informatics.

• Implementation Steps and Technologies Used:

  1. Preliminary Study: Conducted interviews with 10 participants to understand their acceptance of AI-generated fitness narratives and to refine design guidelines for narrative generation.
  2. Experimental Platform: Developed an online platform using the Terra API, enabling participants to upload 7 days of step data and generate data representations under different conditions.
  3. Online Experiment:
     • Recruited 273 users and presented their step data under three conditions.
     • Collected user reflections and feedback on the data representations.
     • Quantified research metrics through surveys: Reflection Support (TSRI) and User Engagement Scale (UES-SF).

Research Findings

• Specific Findings:

  • Data Reflection (TSRI Scale):
    • Text representation enhanced users' reflective ability in the comparison dimension.
    • The text + graphical hybrid mode did not significantly improve reflection compared to text alone.
  • User Experience (UES-SF Scale):
    • Compared to pure graphical representation, text representation performed better in capturing users' attention (Focused Attention) and providing a sense of reward (Reward).
    • Users were more likely to perceive interfaces with textual explanations as more valuable.
  • User Feedback:
    • Textual narratives helped users identify step patterns and motivated goal reflection.
    • Most users preferred a combined text and graphical representation but emphasized the need for personalization and contextually relevant information in the text.
    • A minority of users questioned the accuracy and credibility of the generated text, highlighting the need for transparency.

• Comparison with Existing Solutions:

  • LLM-generated dynamic text is more flexible than traditional pre-written script-based text and better adapts to different users' data and contexts.
  • The positive impact on user attention and engagement is challenging to achieve with existing quantitative personal informatics systems.

• Limitations and Future Directions:

  • Limitations:
    1. Limited contextual information in data generation; current fitness trackers lack the capability to provide highly detailed background analysis.
    2. Inability to fully control biases in generative AI narratives, leading to abstraction and consistency challenges.
    3. Constraints in device types (e.g., specific tracker brands) and cultural backgrounds in the experiment, requiring broader validation for global users.
  • Future Directions:
    1. Explore more complex multi-data-source narrative generation (e.g., integrating heart rate, calorie consumption, etc.).
    2. Investigate methods to maintain novelty and personalization of textual content during long-term use.
    3. Design interactive Q&A features, enabling users to engage with generated text to enhance system usability and credibility.