Voice Assistants for Health Self-Management: Designing for and with Older Adults

Research Background and Issues

What problems or challenges did the authors identify?

• Older adults face two major health management issues: insufficient awareness of health information and difficulties in adhering to medical regimens. Specific challenges include difficulties in understanding After Visit Summaries (AVS) from doctors and creating and adhering to medication reminders.
• Key issues with current voice assistant technologies: 1) lack of user-specific contextual support, and 2) designs that do not cater to the needs of older adults, making them difficult to use.
• Medication management technologies, such as mobile health apps and wearable devices, are often overly complex, not user-friendly for older adults, and lack personalization or adaptation to their existing habits, creating additional barriers to use.

Why is this issue important?

• The global aging trend is putting pressure on healthcare resources. Promoting self-health management (self-care) among older adults is crucial not only for their independent living but also for alleviating the burden on healthcare systems.
• Failures in health management among older adults can lead to medication misuse and missed appointments, resulting in severe health consequences and resource wastage.
• Effectively addressing health management issues for older adults can improve their quality of life and reduce dependence on caregiving services.

Research Motivation and Related Work

• The World Health Organization's mAgeing initiative emphasizes leveraging digital technologies to enhance the health management capabilities of older adults.
• Voice assistants, due to their voice-based interaction convenience, are particularly suitable for users with physical or cognitive impairments, making them a potential tool for improving self-health management.
• The authors identified specific reasons why existing voice assistants fail to meet the needs of older adults, such as communication breakdowns and non-intuitive reminder functionalities.

Solution

What methods or solutions did the authors propose?

• A personalized voice assistant powered by large language models (LLMs, such as GPT-4) was proposed to process After Visit Summaries (AVS) from doctors, provide information synthesis, and assist in setting personalized medication reminders.
• A five-stage development process was designed, ranging from user research to prototype design and system validation, ensuring the technology effectively addresses the challenges in health management faced by older adults.

What are the innovative aspects of this solution?

1. Integration of LLMs’ contextual understanding capabilities into the voice assistant, enabling smoother conversations based on actual health data.
2. Support for personalized interactions: generating customized medication reminders based on users’ health data and personal habits, with flexibility for adjustments.
3. Creation of user profiles: allowing the voice assistant to dynamically set reminders based on users’ daily routines (e.g., sleep schedules, medication box usage) rather than fixed schedules.
4. Implementation of personalized health-related Q&A based on user context, addressing the lack of contextual support in existing voice assistants.

What are the implementation steps? What key technologies were used?

1. Phase 1: User Interviews
   Conducted interviews with 17 older adults to identify key health management challenges such as difficulties with medication management and understanding medical information.

2. Phase 2: Prototype Development
   Developed an initial high-fidelity prototype by integrating GPT-4 into existing voice assistants (e.g., Amazon Alexa) to enable AVS summarization and medication reminder setup.

3. Phase 3: Co-Design Workshops
   Collaborated with 10 older adults in co-design workshops to gather feedback and suggestions for improving the voice assistant’s functionality.

4. Phase 4: Prototype Refinement
   Refined the system based on user feedback, adding features such as grouped medication reminders and adjustable reminder frequencies.

5. Phase 5: In-Home Validation Study
   Conducted real-world usability testing with 5 participants to evaluate the system’s ability to process health information and create medication reminders.

Research Outcomes

What specific results were achieved?

• The system successfully processed After Visit Summaries (AVS), providing users with clear health information synthesis and improving their health awareness.
• Simplified the process of creating reminders, enabling the generation of personalized medication reminders based on user habits, with the flexibility for modifications.
• Achieved a system usability score of over 85, significantly surpassing the "good" threshold of 70.

What advantages does it have compared to existing solutions?

• Personalization and contextual support: The system generates reminders based on health data and user habits, significantly improving usability compared to existing voice assistants that rely on fixed settings.
• Support for autonomy and control: Users can modify reminder settings at any time, ensuring they maintain control over their health management.

What were the experimental or evaluation results?

• Usability Results:
  • All participants successfully completed AVS information synthesis, with 100% of health-related questions answered.
  • A total of 112 reminders were set, achieving an accuracy rate of over 98%.
  • The system demonstrated high user satisfaction, with 4 out of 5 participants expressing willingness to continue using it.

Limitations and Future Directions

• Research Limitations:

  • The sample size was small, and participants were concentrated in a specific community, limiting the generalizability to diverse populations (e.g., older adults with mild cognitive impairments).
  • The system was tested using simulated AVS data; future work should evaluate its ability to process real medical documents and unstructured data.

• Future Directions:

  • Expand the user base, particularly targeting older adults with significant cognitive or physical impairments.
  • Address potential risks of LLMs in medical contexts, such as misinformation and logical errors.
  • Conduct longitudinal studies to assess the system’s long-term impact on health awareness and medication adherence habits.

Conclusion

This study developed a personalized voice assistant powered by large language models, effectively addressing key challenges in health awareness and adherence to medical instructions among older adults. The system’s usability and user experience were validated, demonstrating its potential. Despite some limitations, the research provides valuable insights for future applications of voice assistant technology in health management.