Prediction for Retrospection: Integrating Algorithmic Stress Prediction into Personal Informatics Systems for College Students' Mental Health

Title of the Paper

Prediction for Retrospection: Integrating Algorithmic Stress Prediction into Personal Informatics Systems for College Students’ Mental Health

Paper Information

• Research Domain: Human-Computer Interaction, Personal Informatics, Mental Health, Stress Management
• Keywords: Personal Informatics, Mental Health, Stress Management, Algorithmic Experience, Explainability

Research Background and Problem Statement

• Identified Problems or Challenges:

  • College students are at high risk for the onset of mental health issues, with stress significantly impacting their physical and psychological well-being.
  • Traditional methods struggle to effectively track and understand stress responses, hindering users from taking appropriate measures to address stress sources.
  • Existing smart Personal Informatics (PI) systems primarily focus on state detection and reporting, rather than supporting users in deep reflection and behavioral change.

• Significance:

  • Stress management is crucial for improving college students’ productivity and overall health.
  • Data-driven reflection can enhance users’ understanding of their stress patterns and sources, encouraging positive behavioral changes.

• Research Motivation and Related Work:

  • Personal Informatics systems have been shown to help users reflect on health data and take action, but algorithm-assisted reflection mechanisms in stress management remain underdeveloped.
  • Explainable Artificial Intelligence (XAI) can improve user experience and algorithm transparency, but excessive explanations may lead to cognitive overload and reduced trust, requiring careful design.

Solution

• Method or Solution:

  • Develop a stress management system called “MindScope,” which infers users’ stress levels using smartphone data and machine learning algorithms.
  • Design three levels of stress prediction explanations (no explanation, category-level explanation, detailed explanation) to explore the impact of explainability on user reflection.
  • Prediction reports include data on social activities, locations, actions, sleep, and smartphone usage behavior.

• Innovations:

  • Combining algorithmic predictions with explanation mechanisms to support users in actively reflecting on stress patterns and sources.
  • Introducing a continuous model update mechanism based on user feedback.
  • Exploring how moderate explanations can promote user-led reflection and stress intervention planning.

• Implementation Steps and Techniques:

  • Modeling Phase: Collect 10 days of user stress data through Ecological Momentary Assessment (EMA) and device sensor data to build personalized stress prediction models.
  • Prediction Phase: Provide 15 days of stress prediction reports, allowing users to confirm or adjust system predictions while receiving intervention suggestions.
  • Use SHAP-based explanation generation techniques to present key stress factors as category-level or specific behavioral information.
  • Utilize push notifications to suggest micro-tasks as stress interventions.

Research Outcomes

• Specific Outcomes:

  • MindScope helps users identify stress patterns, recall stress-related events, and develop detailed stress relief plans.
  • During a 25-day field study, participants’ stress levels significantly decreased, demonstrating MindScope’s potential effectiveness in stress management.

• Advantages Over Existing Solutions:

  • MindScope integrates algorithmic predictions with user-driven interactions, reducing memory biases and encouraging vivid and detailed recall of stress events.
  • Provides different levels of explanations to meet user needs, facilitating both real-time reflection and detailed analysis of stress sources.

• Experimental or Evaluation Results:

  • Users rated Type 2 (category explanation) and Type 3 (detailed explanation) higher than Type 1 (no explanation).
  • Type 3 reports were more actionable but were influenced by user trust when model accuracy was low, while Type 2 reports encouraged users to actively recall and generate hypotheses.
  • Issues related to data privacy and model performance in long-term use remain prominent for some user groups.

• Limitations and Future Directions:

  • Expanding the study duration and participant pool is necessary to validate the impact of different explanation levels on a broader population.
  • Cold start issues and sensitivity of algorithms to individual differences require further optimization.
  • Future work could explore longer-term studies and incorporate more interactive design elements (e.g., zooming, filtering) for flexible explanation mechanisms.

Conclusion

This study developed and deployed the MindScope system to provide college students with algorithm-based stress predictions and corresponding explanations, supporting users in better reflecting on and managing stress. The experiment revealed the potential of predictive algorithms and explainability in fostering reflection, while offering numerous design recommendations to improve user experience and system reliability. In the field of personal informatics for mental health, this paper opens a new avenue, advocating for open algorithm design to assist users in deeply understanding their states and behavioral patterns, while promoting positive behavioral changes.