MindfulDiary: Harnessing Large Language Model to Support Psychiatric Patients' Journaling

Title of the Paper

MindfulDiary: Harnessing Large Language Model to Support Psychiatric Patients’ Journaling

Paper Information

• Field: Mental Health and Human-Computer Interaction
• Keywords: Mental health, automated journaling, large language models, clinical psychology, chatbots, patient records

Research Background and Problem

• Problems and Challenges:

  1. Psychiatric patients (especially those with depression) often struggle with emotional expression and daily journaling due to emotional apathy and cognitive load.
  2. Traditional paper-based journals often suffer from low engagement, lack of detail, difficulty in expression, and insufficient structure.
  3. Existing chatbots are limited by rule-based or retrieval-based approaches, which fail to adequately support diverse journaling needs and free-form conversations.
  4. While large language models (LLMs) excel in generating open-domain conversations, concerns about their safety and controllability in clinical settings remain.

• Significance: Journaling can help patients organize their thoughts, release emotions, and provide valuable clinical data to support mental health professionals (MHPs) in better understanding patients. However, improving patient and professional engagement in the journaling and analysis process remains a long-standing challenge.

• Motivation and Related Work:

  1. Leveraging LLMs can enhance the natural interaction capabilities of chatbots while providing patients with a safe and effective journaling tool for mental health issues.
  2. By incorporating guided question-and-answer processes and analyzing emotions and themes in conversations, clinical professionals can be supported with actionable data.

Solution

• Proposed Method or Solution: The authors propose MindfulDiary, a dual-module support system based on large language models, which includes:

  1. Patient Interface: Enables natural language conversations to help patients document daily experiences and emotions, automatically generating chapter-style diary summaries.
  2. Expert Dashboard: Visualizes patients’ daily records, analyzes emotional changes and key events, and supports clinical interviews.

• Innovations:

  1. Employing state machine design to control LLM-generated conversations, using a "phased" dialogue design (including "relationship building, exploration, and closure" phases) to ensure interactions are compliant and safe.
  2. Utilizing LLMs’ natural generative capabilities for mental health data collection, significantly increasing the richness of information compared to traditional open-ended journaling.
  3. Introducing automated analysis of emotions and key events in patient records to provide clinicians with additional contextual information.

• Implementation Steps and Key Techniques:

  1. Patient User Interface:
     • Conduct a mental health questionnaire (e.g., PHQ-9) assessment before formal conversations.
     • Facilitate phased free-form conversations to promote detailed emotional journaling.
     • Automatically generate summary texts using GPT-4, allowing users to review their entries.
  2. Dialogue Pipeline:
     • Designed a dialogue parser (handling state transitions and summary generation) and a response generator (producing dialogue content for specific phases).
     • Used a state machine to ensure the controllability and safety of generated content.
  3. Clinical Dashboard:
     • Provides interaction summaries (e.g., word clouds, event summaries, emotional trends) and raw conversation records to assist clinicians in decision-making and understanding.
     • Flags potential data biases and guides clinicians to verify the accuracy of LLM-generated content.

Research Outcomes

• Specific Results:

  1. MindfulDiary demonstrated its effectiveness in a four-week field study involving 28 patients with depression and 5 mental health professionals, helping patients improve journaling habits and detailed expression while enhancing clinicians’ understanding and empathy toward patients.
  2. The study showed that patients improved in detailed journaling, diverse perspectives, and structured summaries, facilitated by the chatbot’s guided question design.
  3. The expert dashboard enabled clinicians to gain insights into patients’ daily dynamics, emotional states, and significant events, providing data support for clinical decision-making.

• Advantages Compared to Existing Methods:

  1. Unlike standalone journaling tools, MindfulDiary significantly reduces users’ cognitive load through interactive conversations while providing multi-dimensional record content.
  2. Compared to rule-based chatbots, MindfulDiary is more flexible and better at capturing spontaneous or random topics.
  3. Automated summarization and analysis features save clinicians time in identifying key information.

• Experiments or Evaluations:

  • Patients generated an average of 0.62 diary entries per day (501 entries in total), with participation rates remaining consistently high, indicating the system’s effectiveness in motivating users to continue journaling.
  • Guided questioning strategies (e.g., emotional exploration, activity analysis, in-depth follow-ups) helped users produce rich content.
  • Mental health professionals reported that MindfulDiary provided valuable insights into patients’ daily lives and trends beyond clinical interviews.

• Limitations and Future Directions:

  1. Limitations:
     • The study was limited to a single hospital system, with a sample population primarily consisting of adolescents, which may not represent a broader patient demographic.
     • Potential biases and inaccuracies in GPT-generated content could impact clinical outcomes.
  2. Future Directions:
     • Expand research to include users from diverse backgrounds and address more complex mental health issues.
     • Refine mechanisms for verifying LLM-generated data, such as collaborating more closely with mental health experts to develop testing standards.
     • Explore ways to integrate patient-clinician interactions more closely with system functionalities.