Evaluating Large Language Models in Crisis Detection: A Real-World Benchmark from Psychological Support Hotlines.

Area of Science:

• Artificial Intelligence
• Clinical Psychology
• Natural Language Processing

Background:

• Psychological support hotlines face increasing demand and resource limitations.
• Large language models (LLMs) show potential for crisis assessment, but their real-world clinical efficacy is understudied.

Purpose of the Study:

• Introduce PsyCrisisBench, a benchmark for evaluating LLMs in crisis intervention.
• Assess LLM performance across mood status recognition, suicidal ideation detection, suicide plan identification, and risk assessment.

Main Methods:

• Evaluated 64 LLMs (closed- and open-source) on 540 annotated crisis transcripts.
• Utilized zero-shot, few-shot, and fine-tuning approaches.
• Compared LLM performance against trained human operators.

Main Results:

• LLMs achieved high F1 scores for suicidal ideation detection (0.880), suicide plan identification (0.779), and risk assessment (0.907).
• Few-shot prompting and fine-tuning significantly improved LLM performance.
• LLMs matched or exceeded human performance in suicide plan identification and risk assessment.
• Mood status recognition remained a challenge (max F1 = 0.709).
• A smaller fine-tuned model (Qwen2.5-1.5B) outperformed larger models on specific tasks.

Conclusions:

• LLMs demonstrate performance comparable to human operators in text-based crisis assessment.
• LLMs possess complementary strengths to human operators across different assessment tasks.
• PsyCrisisBench offers a framework for developing and deploying AI in clinical mental health.