Assessing the accuracy and consistency of large language models in triaging social media posts for psychological distress

Area of Science:

• Artificial Intelligence
• Mental Health Technology
• Computational Linguistics

Background:

• Online platforms facilitate communication of psychological distress.
• Artificial intelligence (AI), especially natural language processing (NLP), offers tools for mental health support.
• Large Language Models (LLMs) present opportunities for automated analysis of online mental health content.

Purpose of the Study:

• To evaluate the efficacy of state-of-the-art LLMs in triaging social media posts indicating psychological distress.
• To compare the performance of ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro against trained clinician assessments.
• To assess LLMs' ability to determine intervention urgency using a zero-shot prompting approach.

Main Methods:

• Utilized a dataset of 425 Italian-language Reddit posts expressing psychological distress.
• Compared LLM triage performance (ChatGPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro) with clinician assessments using the Mental Health Triage Scale (MHTS).
• Employed zero-shot prompting with and without role assignment to simulate clinical judgment.

Main Results:

• LLMs consistently overestimated intervention urgency compared to human clinicians.
• Moderate to strong correlations were observed between LLM and human judgments.
• GPT-4o demonstrated the highest classification performance; Claude 3.5 Sonnet showed high sensitivity but lower precision; Gemini 1.5 Pro had balanced but lower performance.

Conclusions:

• LLMs show promise for automated mental health triage on social media.
• The tendency of LLMs to overestimate urgency necessitates careful interpretation and human oversight.
• Model-specific performance variations highlight the importance of selecting appropriate AI tools for mental health applications.