Related Experiment Video
Updated: May 15, 2026
07:31
Implementation of a Real-Time Psychosis Risk Detection and Alerting System Based on Electronic Health Records using CogStack
Published on: May 15, 2020
Automating multi-label crisis detection in psychological support hotlines with pre-trained models.
Shuying Rao1,2,3, Guifeng Deng1,2,3, Haidong Song1
1Affiliated Mental Health Center and Hangzhou Seventh People's Hospital and School of Brain Science and Brain Medicine, Zhejiang University School of Medicine, Hangzhou, China.
PLOS Digital Health
|May 13, 2026
Summary
AI models show promise in detecting psychological crises from hotline calls. Large Language Models (LLMs) like GPT-4o and DeepSeek-R1, using few-shot learning, achieved high accuracy, reducing reliance on extensive training data for mental health support.
Area of Science:
- Artificial Intelligence
- Mental Health Technology
- Computational Linguistics
Background:
- Psychological support hotlines are crucial for crisis intervention.
- Increasing demand strains human resources, necessitating automated solutions.
- AI-driven models are needed for efficient and accurate crisis detection.
Purpose of the Study:
- To evaluate deep learning and pre-trained models for detecting psychological crises in audio and text data.
- To compare the effectiveness of different AI strategies, including LLMs, for crisis detection.
- To assess the clinical applicability of AI-generated explanations in mental health contexts.
Main Methods:
- Utilized 1,057 calls from the Hangzhou Hotline (2022-2023).
- Employed deep learning with pre-trained models (Wave2Vec, Whisper, RoBERTa, GPT) and LLMs (GPT-4, DeepSeek series) via prompt engineering.
- Adopted strategies including deep learning classification and LLM-based prediction with few-shot learning and Chain-of-Thought reasoning.
Main Results:
- Deep learning with GPT embeddings achieved 80.48% F1 score in identifying high-risk calls, outperforming auditory models.
- Text semantics were more valuable than acoustic features for crisis prediction.
- LLMs (GPT-4o, DeepSeek-R1) with few-shot learning matched deep learning performance, demonstrating reduced data dependency.
Conclusions:
- LLMs show significant potential for mental health crisis detection, comparable to traditional deep learning models.
- Few-shot learning and advanced reasoning in LLMs enable effective application in clinical domains with minimal data.
- AI-generated explanations are clinically applicable, paving the way for future AI integration in mental health services.
