Efficient and Effective Fine-tuning Method for Depression Detection from Conversation.

Area of Science:

• Artificial Intelligence
• Computational Psychiatry
• Natural Language Processing

Background:

• Depression significantly impacts society, necessitating early detection for effective treatment.
• Large Language Models (LLMs) show promise in depression detection, with prompt-engineering outperforming fine-tuning in recent studies.
• Fine-tuning LLMs for depression detection traditionally requires substantial memory, limiting accessibility.

Purpose of the Study:

• To develop a fine-tuning method for LLMs that achieves higher accuracy than prompt-engineered models for depression detection.
• To minimize GPU memory requirements during the LLM training process.
• To validate the model's performance on both public and private datasets.

Main Methods:

• Quantized the parameters of the LLM to reduce memory footprint during training.
• Retained the original LLM structure, training it as a generative model without adding a classification layer.
• Evaluated the model on the DAIC-WOZ and a private PROMPT dataset for depression detection.

Main Results:

• Achieved an F1 score of 84% for depression detection on the DAIC-WOZ dataset using low GPU memory.
• Demonstrated superior performance compared to prompt-engineered models.
• Attained an F1 score of 82% on the private PROMPT dataset, indicating robustness beyond pre-training data.

Conclusions:

• The proposed fine-tuning method effectively detects depression with high accuracy and low memory requirements.
• This approach offers a more memory-efficient alternative to prompt-engineering for LLM-based depression detection.
• The model's strong performance on a private dataset suggests generalizability and reduced reliance on publicly available pre-training data.