Detecting suicide risk in adolescents and young adults: A machine learning-based analysis of nonverbal behaviors exhibited during suicide assessments

项目总结 自杀是美国15-24岁青少年的第二大死因。一项具有挑战性的 自杀预防的组成部分是发现高危青年。先前的研究表明， 绝大多数自杀死者在与精神健康人士的最后一次谈话中否认有自杀念头 提供商。因此，只有15%的心理健康专业人士表示非常自信也就不足为奇了。 评估青少年自杀风险。行为标记物为更客观地确定风险提供了一条途径。 尽管在这一领域取得了进展，但行为标记物主要以反应的形式运作 时间和任务的表现，只是触及了可能的表面，具有丰富、动态的性质 行为数据。计算科学的最新进展为建立行为模型提供了机会 人类不容易量化甚至察觉到的信息。这项研究旨在利用 基于机器学习的自杀过程中非语言行为特征研究 评估，并测试这些行为是否可以用来识别有自杀倾向的青少年和年轻人。 具体地说，我们将自动提取副语言特征、自发的面部动作和头部 青少年和青壮年以及他们的临床面试者所展示的运动。我们将使用传统的 假设检验检验一组非语言行为是否从先前的研究中获得信息 区分有自杀倾向(即过去一年的积极自杀意念)和非自杀倾向(即没有自杀的终生病史 思想/行为)青少年和青壮年(目标1)。然后我们将使用机器学习来测试 任何额外的、由经验决定的非语言行为可能有助于我们识别自杀的能力 参与者(目标2)。数据将来自自我伤害思想和自我伤害的录音管理 行为访谈--对有自杀倾向和非自杀倾向的青少年和年轻人(n=232；12-19岁)进行修订， 和哥伦比亚-自杀严重程度评定量表的视频管理，有自杀倾向和 非自杀青年(n=70；18-24岁)。着眼于自杀行为的前瞻性预测 未来的研究，这一行的长期目标是利用计算方法来量化非 可以用来客观和大规模地检测自杀风险的言语行为。

PROJECT SUMMARY Suicide is the second leading cause of death among 15-24-year-olds in the United States. A challenging component of suicide prevention is the detection of high-risk young people. Prior research suggests that the vast majority of suicide decedents deny suicidal ideation in their last conversation with a mental health provider. It is thus unsurprising that only 15% of mental health professionals report feeling very confident assessing youth suicide risk. Behavioral markers offer one avenue for more objective risk determination. Despite progress in this area, behavioral markers have been operationalized primarily in the form of reaction times and task performance, only scratching the surface of what is possible with the rich, dynamic nature of behavioral data. Recent advances in computational science offer an opportunity to model behavioral information that is not easily quantifiable or even perceivable to human beings. This study aims to employ machine learning-based approaches to characterize non-verbal behaviors exhibited during suicide assessments, and test whether these behaviors can be used to identify suicidal adolescents and young adults. Specifically, we will automatically extract paralinguistic characteristics, spontaneous facial action, and head motion exhibited by adolescents and young adults, and their clinical interviewers. We will use traditional hypothesis testing to examine whether a set of non-verbal behaviors informed by previous research differentiate suicidal (i.e., past year active suicidal ideation) and nonsuicidal (i.e., no lifetime history of suicidal thoughts/behaviors) adolescents and young adults (Aim 1). We will then use machine learning to test whether any additional, empirically-determined non-verbal behaviors may contribute to our ability to identify suicidal participants (Aim 2). Data will be drawn from audio-recorded administrations of the Self-Injurious Thoughts and Behaviors Interview-Revised with suicidal and nonsuicidal adolescents and young adults (n=232; 12-19 yrs), and video-recorded administrations of the Columbia-Suicide Severity Rating Scale with suicidal and nonsuicidal young adults (n=70; 18-24 yrs). With an eye toward prospective prediction of suicidal behavior in future research, the long-term goal of this line of work is to harness computational methods to quantify non- verbal behaviors that can be used to detect suicide risk objectively and at scale.