"I spent 14 hours debugging just one assignment": Toward Computer-Mediated Personal Informatics for Computer Science Student Mental Health

Research Background and Issues

• Problems and Challenges
  The authors identified significant mental health issues among computer science (CS) students, with rates of anxiety and depression being twice as high as those of other undergraduates and 5-10 times higher than the general population. The specific academic factors contributing to these problems remain unclear. The authors found that the complexity of debugging code, lack of self-awareness regarding stress, and impostor syndrome are major challenges.

• Significance
  These issues not only affect students' learning efficiency but may also lead to further deterioration of physical and mental health, academic disengagement, and even a reduced willingness to pursue careers in computing-related fields. Additionally, the impact on women and minority groups could result in a regression in the representation of these groups in the future tech industry.

• Research Motivation and Related Work
  Current research on student mental health primarily focuses on populations with pre-existing mental health conditions, lacking in-depth exploration of academic stress among the broader CS student population. While digital mental health tools are widely developed, there is a design gap between these tools and students' actual needs and experiences. Furthermore, the emotional impact of debugging and real-time stress monitoring are critical research issues at the intersection of computer education and mental health.

Solution

• Proposed Solution
  The authors designed and developed EmotionStream, an algorithm-assisted personal informatics (PI) tool for continuous monitoring and providing context- and emotion-based insights. The system combines automated emotion recognition (AER) with user self-reports to facilitate self-reflection.

• Innovations

  1. Integrates real-time facial emotion recognition with user behavior data (keystrokes, mouse interactions) and task context for real-time stress prediction.
  2. Employs a hybrid approach (combining active data collection with passive monitoring) to effectively address the excessive burden of traditional tracking tools.
  3. Features a visualization dashboard that helps students examine the relationship between their emotional states and task contexts, enabling improvement.

• Implementation Steps and Technical Details

  1. Data Collection: Tracks students' facial expressions using AER models such as DeepFace and Residual Masking Network, while recording keyboard and mouse activity and application usage. Ensures privacy protection by using local storage to prevent the leakage of personally identifiable information.
  2. User Feedback: Based on the Experience Sampling Method (ESM), the system prompts users every 20 minutes to self-report their emotions and stress levels, then summarizes behavioral and emotional trends on a dashboard.
  3. Model Evaluation: Validates the accuracy of the AER model and uses a random forest model combined with user behavior context for stress prediction, achieving a high F1 score (0.88).

Research Outcomes

• Specific Findings

  1. Qualitative and quantitative analyses reveal that debugging tasks and late-night activity times are significantly associated with high stress.
  2. EmotionStream was well-received among students, with over 70% of participants rating the tool as "good" or "very good."
  3. Although there were significant differences between AER model predictions and self-reported emotions, the performance of the stress prediction model improved significantly when contextual information was incorporated.

• Comparative Advantages Over Existing Solutions

  1. Addresses the mismatch between traditional digital mental health tools and students' actual needs.
  2. Innovatively integrates contextual data to dynamically provide refined insights, promoting self-reflection and behavioral adjustments among students.
  3. High engagement and adaptability of the tool indicate strong potential for daily application.

• Experimental or Evaluation Results

  1. The RMN and DeepFace models achieved emotion classification accuracies of 42% and 47%, respectively, highlighting their relative strengths in recognizing neutral emotions.
  2. When additional contextual data was used for stress prediction, the F1 score reached 0.88.
  3. Participant feedback indicated that the visualization output of the dashboard enabled them to identify stressors and adjust behaviors, effectively promoting mental health management.

• Limitations and Future Directions
  The authors identified the following limitations of the tool:

  1. Limited sample diversity, with only a small representation of minority groups, making it difficult to comprehensively assess racial, cultural, or gender biases.
  2. Low accuracy of facial emotion recognition models for non-neutral emotions, necessitating improvements in the diversity of training data for future algorithms.
  3. The tool currently supports only the Windows operating system, restricting its potential for widespread adoption.
  4. Long-term data on stress and academic performance tracking has not yet been sufficiently collected, requiring further validation of the tool's long-term effectiveness.

Through the development and evaluation of EmotionStream, the authors not only explored the academic stressors faced by CS students but also laid a solid foundation for designing more practical digital mental health tools. Future research should focus more on sample diversity, long-term application performance, and cross-platform compatibility to further enhance the tool's universality and accessibility.