CALLM: Enhancing Clinical Interview Analysis Through Data Augmentation With Large Language Models.

Area of Science:

• Artificial Intelligence in Healthcare
• Computational Psychiatry
• Machine Learning for Mental Health

Background:

• Global rise in mental health disorders incurs significant economic burden.
• Scarcity and imbalance of clinical data challenge automated mental health diagnosis.
• Existing machine learning models struggle with limited, uneven datasets.

Purpose of the Study:

• Introduce a novel clinical transcript data augmentation framework using large language models (CALLM).
• Generate realistic synthetic clinical data to address data scarcity and imbalance.
• Enhance the effectiveness of machine learning algorithms in automated mental health diagnosis.

Main Methods:

• Developed a CALLM framework employing "patient-doctor role-playing" for data generation.
• Introduced a "Textbook-Assignment-Application" (T-A-A) partitioning for systematic dataset creation.
• Utilized a "Response-Reason" prompt engineering paradigm for authentic transcript generation.

Main Results:

• Achieved 0.77 balanced accuracy, 0.70 F1-score, and 0.78 AUC on the E-DAIC PTSD dataset using a fine-tuned DistilBERT model.
• Demonstrated robust adaptability in Zero-Shot Learning (ZSL) and Few-Shot Learning (FSL) scenarios.
• CALLM framework showed consistent improvements over other augmentation methods and PTSD diagnostic works, at less than 1% of traditional data collection costs.

Conclusions:

• The CALLM framework effectively generates high-quality synthetic clinical data for mental health diagnosis.
• This approach significantly enhances machine learning model performance, particularly in low-data regimes.
• CALLM offers a cost-effective and scalable solution for augmenting clinical datasets, advancing automated mental health diagnosis.