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

Advances in artificial intelligence, particularly in natural language processing, offer promising tools for addressing mental health challenges in online contexts, potentially identifying at-risk individuals and informing timely interventions. This study investigates the potential of Large Language Models (LLMs) for automatically triaging social media posts expressing psychological distress. Using a dataset of 425 Italian-language Reddit posts, we compared the triage performance of three state-of-the-art LLMs - ChatGPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro - with trained clinician assessment using an adapted version of the Mental Health Triage Scale (MHTS), a validated instrument used in psychiatric screening services. A zero-shot prompting approach, with and without role assignment (simulating a clinician's perspective), evaluated the models' capability to assess intervention urgency. Results revealed that LLMs consistently overestimated urgency compared to human raters, although correlations with human judgments were moderate to strong, with GPT-4o and Claude 3.5 Sonnet demonstrating higher agreement. GPT-4o achieved the best classification performance, highlighting its potential for this task. Claude 3.5 Sonnet showed high sensitivity but lower precision, indicating a tendency toward false positives, while Gemini 1.5 Pro exhibited more balanced but generally lower performance. These findings suggest that while LLMs show promise for mental health triage in social media, their tendency to overestimate urgency and model-specific variations in performance underscore the need for careful interpretation, and human oversight when applying LLMs in mental health contexts.