Automating multi-label crisis detection in psychological support hotlines with pre-trained models

Psychological support hotlines provide immediate help to individuals in crisis, with operators assessing emotional states and suicide risk. However, increasing demand has led to a shortage of trained professionals, emphasizing the need for AI-driven crisis detection models. This study included 1,057 calls from the Hangzhou Hotline (2022-2023) to evaluate the effectiveness of deep learning and pre-trained models in detecting psychological crises using audio (Wave2Vec, Whisper) and transcribed text (RoBERTa, GPT). We adopted two strategies: deep learning classification with pre-trained models and Large Language Models (LLMs)-based prediction via prompt engineering (GPT-4 and DeepSeek series). The deep learning framework, employing GPT embeddings excelled in multi-label predictions compared to auditory model, achieving 80.48% [80.18%, 80.78%] F1 scores for identifying high-risk calls in prospective tests. Fusion experiments revealed that acoustic features offered negligible predictive value compared with text semantics. Notably, GPT-4o and DeepSeek-R1, utilizing few-shot learning, demonstrated performance comparable to the GPT-embedding deep learning model across multiple tasks. This suggests that their advanced Chain-of-Thought reasoning effectively mitigates data dependency gap, enabling LLMs to align with clinical domains using a few examples. Expert evaluation confirmed the clinical applicability of GPT-generated explanations. Taken together, these findings highlight the potential of LLMs in mental health crisis detection and lay the foundation for future research.