Computational Notebooks as Co-Design Tools: Engaging Young Adults Living with Diabetes, Family Carers, and Clinicians with Machine Learning Models

Title of the Paper

"Computational Notebooks as Co-Design Tools: Engaging Young Adults Living with Diabetes, Family Carers, and Clinicians with Machine Learning Models"

Bibliographic Information

• Research Domain: Co-design of interactive machine learning tools and their application in diabetes management
• Keywords: Co-design, machine learning, human-computer interaction, diabetes, computational notebooks, multidisciplinary collaboration, user experience, health technology

Research Background and Problem Statement

• Issues and Challenges:

  • Artificial Intelligence (AI) and Machine Learning (ML) are rapidly permeating fields such as healthcare, but algorithm design often fails to adequately incorporate end-user needs and expectations, potentially leading to inequitable health outcomes.
  • Current AI design processes are heavily focused on data science, with workflows differing significantly from user experience design, making human-machine collaboration and user trust difficult to achieve.
  • Computational notebooks (e.g., Jupyter Notebook) are widely used in data science education and team collaboration, but their potential as co-design tools for non-technical users remains underexplored.

• Significance:

  • Cross-disciplinary collaboration and user-centered participatory design are critical for ensuring the fairness and usability of AI systems.
  • By actively involving end-users (e.g., young adults with diabetes, family carers, clinicians), the development of human-machine collaborative systems and health prediction models can better address specific needs.

• Research Motivation and Related Work:

  • Combining interactive data science tools with user experience design can bridge the gap between data science workflows and user needs.
  • In current health systems, collaboration based on traditional AI development processes remains limited to specialized groups, failing to effectively engage diverse user populations (e.g., non-technical individuals).

Solution

• Methodology or Solution:

  • The authors conducted a co-design study combining computational notebooks (Jupyter Notebook) with online workshops to guide user groups—including diabetes patients, family carers, and clinicians—in collaboratively designing machine learning models.
  • The study involved multi-stage design processes: from static images to interactive data visualization tools, ultimately enabling users to create virtual feature importance models tailored to their needs.

• Innovative Contributions:

  • Proposed a novel participatory design framework utilizing computational notebooks as "boundary objects" to facilitate knowledge sharing and co-design within multidisciplinary teams.
  • Empowered non-technical users to create virtual ML models that reflect their personal health needs and values, enhancing creativity and interactive learning experiences.

• Implementation Steps and Key Techniques:

  1. Computational Notebook Design:
     • Developed co-design notebooks integrating static explanations (e.g., dataset descriptions) and interactive ML models (e.g., risk visualizations).
     • Included five core components: user health data experiences, dataset sample demonstrations, basic machine learning concepts, health risk prediction models, and user-defined virtual model activities.
  2. Experimental Design:
     • Recruited young diabetes patients (6 participants), family carers (4 participants), and clinicians (3 participants) to participate in five online workshops.
     • Participants engaged in interactive activities (e.g., adjusting risk sliders) to learn about ML concepts, discuss potential model impacts, and define specific health functionality requirements.
  3. Coding Practice:
     • Used simplified Python scripts to guide users in inputting key feature variables and generating virtual feature importance graphs based on their needs.

Research Outcomes

• Specific Results:

  • Methodological Success: Most participants successfully created virtual feature importance models, demonstrating learning curves and breakthroughs despite initial unfamiliarity with technical tools.
  • Benefits of ML Co-Design:
    • Enabled participants to predict health risks using interactive computational notebooks and discuss potential benefits and harms (e.g., resource allocation).
    • Enhanced understanding and acceptance of complex ML systems among patients, families, and clinicians.
  • User Feedback:
    • Participants reported that interactive elements in the notebooks made complex ML topics accessible and enriched exploration.
    • Users proposed numerous practical and creative feature suggestions, including detailed estimation models based on multifactor inputs (e.g., diet, sleep, BG levels).

• Comparative Advantages Over Existing Solutions:

  • Compared to traditional AI design methods, computational notebooks as collaborative tools improved participants' comprehension and creativity.
  • Diverse groups effectively engaged in ML explanation, critique, and innovation processes.

• Limitations and Future Directions:

  • Dataset Constraints: Using existing datasets may limit users' ability to fully express and address their needs.
  • User Experience Challenges: Some non-technical users felt confused by programming interfaces and task operations.
  • Future Directions:
    • Develop more customizable, user-friendly notebook interfaces to reduce technical complexity.
    • Explore collaborative approaches starting from dataset design to ensure inclusivity and specificity.