An Exploratory Deep Learning Approach for Predicting Subsequent Suicidal Acts in Chinese Psychological Support Hotlines

TL;DR

The paper tackles the problem of accurately predicting subsequent suicidal acts using long-sequence speech from a Chinese psychological support hotline, addressing limitations of scale-based assessments and small non-clinical datasets. It proposes a novel multi-task deep learning architecture that uses Whisper to extract long-sequence speech features, a Transformer encoder to model dependencies, and an LSTM decoder with two heads to predict both a suicide-risk scale and overall risk. On a large clinical dataset with 12-month follow-up, the model achieves a 71.15% F1-score, outperforming manual scales and eight competing models, demonstrating the potential for AI-assisted risk identification in hotline settings. The work lays groundwork for prospective clinical validation and human–machine collaboration to improve prevention workflows in mental health services.

Abstract

Psychological support hotlines are an effective suicide prevention measure that typically relies on professionals using suicide risk assessment scales to predict individual risk scores. However, the accuracy of scale-based predictive methods for suicide risk assessment can vary widely depending on the expertise of the operator. This limitation underscores the need for more reliable methods, prompting this research's innovative exploration of the use of artificial intelligence to improve the accuracy and efficiency of suicide risk prediction within the context of psychological support hotlines. The study included data from 1,549 subjects from 2015-2017 in China who contacted a psychological support hotline. Each participant was followed for 12 months to identify instances of suicidal behavior. We proposed a novel multi-task learning method that uses the large-scale pre-trained model Whisper for feature extraction and fits psychological scales while predicting the risk of suicide. The proposed method yields a 2.4\% points improvement in F1-score compared to the traditional manual approach based on the psychological scales. Our model demonstrated superior performance compared to the other eight popular models. To our knowledge, this study is the first to apply deep learning to long-term speech data to predict suicide risk in China, indicating grate potential for clinical applications. The source code is publicly available at: \url{https://github.com/songchangwei/Suicide-Risk-Prediction}.

Figures (3)

• Figure 1: The flowchart of the psychological support hotlines. The service process of a psychological support hotlines is divided into three main stages: suicide risk assessment, crisis intervention, and follow-up support. During the suicide assessment stage, interventionists usually spend 15 to 30 minutes using a suicide risk assessment scale to determine whether the caller has suicidal tendencies. In the crisis intervention stage, for those callers assessed as having suicidal tendencies, necessary support and assistance will be provided. Finally, in the follow-up support stage, these individuals will undergo continuous risk assessment and intervention over the next 12 months to ensure that they receive long-term and sustained care and support.
• Figure 2: The model architecture for this research. First, segmented speech is extracted as a series of embeddings using the pre-trained Whisper model. These feature embeddings are then processed through a transformer-based encoder to establish long-distance dependencies. Finally, the LSTM decoder performs multi-task learning by connecting to two classification heads—one for scale prediction and the other for overall suicide risk assessment.
• Figure 3: Distribution of speech duration (in minutes) for the psychological support hotlines.