基于深度多任务学习的心理援助热线来电者自杀危险评估模型的构建与检验

自杀是我国居民的主要死因之一，心理援助热线是常用的有效自杀预防手段。在热线预防自杀实践中，用传统量表方法评估自杀危险程度需要逐一询问来电者较多问题，评估可行性和结果准确性受到来电者配合度的影响。本项目拟在心理援助热线通话期间，由接线员聚焦于自杀心理危机问题与来电者进行半结构式交谈，采用人工智能技术从热线通话中实时抽取文本情感、对话问答、音频及声谱图数据，构建深度多任务学习模型，自动计算来电者自杀危险程度。用已贴自杀结局标签的1万条既往来电录音训练和调试该模型。而后将该模型挂载于北京市心理援助热线工作系统，自动评估每个来电者的自杀危险程度，用传统量表评估的结果平行验证该模型的性能。将1500名接受了该模型评估的来电者纳入队列，开展前瞻性随访研究。以随访期间这些来电者是否发生了自杀行为作为金标准来检验该自动评估模型的预测能力。本项目的评估与随访过程不干扰现有的热线心理危机干预流程。

Suicide is one of the leading causes of death among residents in China. Psychological support hotline is one of commonly used and effective methods for suicide prevention. In the previous suicide prevention activities, routine scales were used to evaluate the suicidal risk of hotline callers. However, the accuracy and feasibility of the evaluation were affected by the cooperation level of the callers. In the present project, the artificial intelligence technique would be used to extract text emotional features, dialogue response features of key facts, audio frequency based emotional features, and spectrogram from the ongoing hotline dialogue, while the hotline operators have a semi-structured conversation focused on the suicidal crisis with hotline callers. A deep multi-task learning model would be developed to real-time calculate the severity of suicidal risk of the callers. The model would be trained and adjusted by ten thousand hotline tape-records which were labeled with suicidal outcomes and suicidal risk scores. The deep multi-task learning model will be loaded at the operating system of the Beijing Psychological Support Hotline. For each callers with themselves’ psychological problem, a suicidal risk score would be given by the model, based on the ongoing hotline dialogue. To estimate the parallel validity of the model, the routine scale suicidal risk score of the callers would be also given by the hotline operator. Approximate 1500 hotline callers who were given deep learning model suicidal risk score were recruited and followed up for up to 12 month. The suicidal outcomes were identified via follow-up, which would be used to test the predictive validity of the deep multi-task learning model. The suicidal risk assessment by the model and the follow-up would not interfere with the routine suicide intervention process.