Mental-LLM: Leveraging Large Language Models for Mental Health Prediction via Online Text Data.

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Mental Health Informatics

Background:

• Large language models (LLMs) show potential in various applications.
• Limited research exists on LLM capabilities and enhancements for mental health tasks.

Purpose of the Study:

• To comprehensively evaluate multiple LLMs on mental health prediction tasks.
• To investigate the impact of prompting techniques and fine-tuning on LLM performance in mental health.

Main Methods:

• Evaluated Alpaca, Alpaca-LoRA, FLAN-T5, GPT-3.5, and GPT-4 using zero-shot, few-shot, and instruction fine-tuning.
• Conducted experiments on online text data for mental health prediction.
• Performed an exploratory case study on LLM reasoning capabilities for mental health.

Main Results:

• Zero-shot and few-shot prompting showed promising but limited performance.
• Instruction fine-tuning significantly improved LLM performance across all tasks.
• Fine-tuned models (Mental-Alpaca, Mental-FLAN-T5) outperformed GPT-3.5 and GPT-4 prompt designs and matched state-of-the-art models.
• GPT-4 demonstrated promising reasoning capabilities.

Conclusions:

• Instruction fine-tuning is a key method for enhancing LLMs in mental health.
• While promising, LLMs require further development to address limitations like bias and ethical risks before real-world deployment.
• Guidelines are provided for enhancing LLM capabilities in mental health applications.