Synthetic patient and interview transcript creator: an essential tool for LLMs in mental health

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Mental Health Technology

Background:

• High-quality training data is crucial for developing specialized large language models (LLMs), particularly in sensitive fields like mental health.
• Real patient data faces privacy and legal restrictions, hindering LLM development.
• Synthetic data generation offers a potential solution to overcome these data access challenges.

Purpose of the Study:

• To design and implement a synthetic patient and interview generation framework for creating training data for mental health LLMs.
• To ensure the synthetic data is contextually rich, demographically accurate, and lexically diverse.
• To address privacy and legal constraints associated with using real patient data.

Main Methods:

• Utilized two locally run instances of Llama 3.3:70B models: one as an interviewer and one as a patient.
• Developed a customizable question bank to structure interview transcripts.
• Employed a hybrid approach for patient profile generation, combining predefined variables with LLM-generated content.

Main Results:

• Generated interview transcripts exhibited lexical diversity comparable to human conversation (median Distinct-1 scores of 0.44 for patient, 0.33 for interviewer).
• Synthetic patient profiles showed demographic distributions not significantly different from real-world data and high word usage diversity (average Distinct-1 score of 0.8).
• The framework successfully produced contextually rich and realistic synthetic patient interactions.

Conclusions:

• The developed framework effectively generates high-quality synthetic training data for mental health LLMs.
• This approach overcomes privacy and legal barriers, facilitating the adoption of LLMs in mental health.
• The system's ability to tailor demographics and ensure data diversity supports the development of robust and ethical AI tools for mental healthcare.