Day 5. Suicidal Ideation Detection: A Review of Machine Learning Methods and Applications综述

Title:
Suicidal Ideation Detection: A Review of Machine Learning Methods and Applications
自杀意念检测：机器学习方法及应用综述

Keywords:
Deep learning 深度学习
feature engineering 特征工程
social contents 社交内容
suicidal ideation detection (SID) 自杀意念检测

Abstract：
Suicide is a critical issue in modern society. Early detection and prevention of suicide attempts should be addressed to save people’s life. Current suicidal ideation detection (SID) methods include clinical methods based on the interaction between social workers or experts and the targeted individuals and machine learning techniques with feature engineering or deep learning for automatic detection based on online social contents. This article is the first survey that comprehensively introduces and discusses the methods from these categories. Domain-specific applications of SID are reviewed according to their data sources, i.e., questionnaires, electronic health records, suicide notes, and online user content. Several specific tasks and data sets are introduced and summarized to facilitate further research. Finally, we summarize the limitations of current work and provide an outlook of further research directions.
自杀是一个严重的现代问题。早期发现和预防自杀未遂应该致力于拯救人们的生命。目前的自杀意念检测方法主要包括基于社会工作者或专家与目标个体互动的临床方法和基于特征工程或深度学习的机器学习技术，用于基于在线社交内容的自动检测。本文是第一次全面介绍和讨论这些分类方法的调查。根据SID的数据来源，即调查问卷、电子健康记录、自杀记录和在线用户内容，对SID的特定领域应用进行了审查。介绍和总结了几个具体的任务和数据集，以便于进一步的研究。最后，总结了目前工作的局限性，并对下一步的研究方向进行了展望。

1.INTRODUCTION
MENTAL health issues, such as anxiety and depression, are becoming increasingly concerned in modern society, as they turn out to be especially severe in developed countries and emerging markets. Severe mental disorders without effective treatment can turn to suicidal ideation or even suicide attempts. Some online posts contain much negative information and generate problematic phenomena, such as cyberstalking and cyberbullying. Consequences can be severe and risky since such lousy information is often engaged in some form of social cruelty, leading to rumors or even mental damage. Research shows that there is a link between cyberbullying and suicide [1]. Victims overexposed to too many negative messages or events may become depressed and desperate; even worse, some may commit suicide.
焦虑和抑郁等心理健康问题在现代社会越来越受到关注，在发达国家和新兴市场，焦虑和抑郁尤为严重。严重的精神障碍如果得不到有效的治疗，就会变成自杀意念，甚至是自杀未遂。一些网络帖子含有大量负面信息，并产生一些问题现象，如网络跟踪和网络欺凌。后果可能是严重的和危险的，因为这些糟糕的信息往往参与某种形式的社会残酷，导致谣言甚至精神损害。研究表明，网络欺凌和自杀之间有联系[1]。过度暴露在负面信息或事件中的受害者可能会变得沮丧和绝望；更糟糕的是，有些人可能会自杀。

The reasons that people commit suicide are complicated. People with depression are highly likely to commit suicide, but many without depression can also have suicidal thoughts [2]. According to the American Foundation for Suicide Prevention (AFSP), suicide factors fall under three categories: health factors, environmental factors, and historical factors [3]. Ferrari et al. [4] found that mental health issues and substance use disorders are attributed to the factors of suicide. O’Connor and Nock [5] conducted a thorough review of the psychology of suicide and summarized psychological risks as personality and individual differences, cognitive factors, social factors, and negative life events.
人们自杀的原因很复杂。抑郁症患者极有可能自杀，但许多没有抑郁症的人也可能有自杀念头[2]。根据美国自杀预防基金会（AFSP），自杀因素分为三类：健康因素、环境因素和历史因素[3]。Ferrari等人[4]发现精神健康问题和物质使用障碍是导致自杀的因素。O’Connor和Nock [5] 对自杀心理进行了全面的回顾，并将心理风险归纳为个性和个体差异、认知因素、社会因素和消极生活事件。

Suicidal ideation detection (SID) determines whether the person has suicidal ideation or thoughts by given tabular data of a person or textual content written by a person. Due to the advances in social media and online anonymity, an increasing number of individuals turn to interact with others on the Internet. Online communication channels are becoming a new way for people to express their feelings, suffering, and suicidal tendencies. Hence, online channels have naturally started to act as a surveillance tool for suicidal ideation, and mining social content can improve suicide prevention [6]. Strange social phenomena are emerging, e.g., online communities reaching an agreement on self-mutilation and copycat suicide. For example, a social network phenomenon called the “Blue Whale Game” in 2016 uses many tasks (such as self-harming) and leads game members to commit suicide in the end. Suicide is a critical social issue and takes thousands of lives every year. Thus, it is necessary to detect suicidality and prevent suicide before victims end their life. Early detection and treatment are regarded as the most effective ways to prevent potential suicide attempts.
自杀意念检测（SID）是通过给定的个人表格数据或个人书写的文本内容来判断一个人是否有自杀意念。由于社交媒体和在线匿名技术的进步，越来越多的人开始在互联网上与他人互动。在线交流渠道正在成为人们表达情感、痛苦和自杀倾向的新方式。因此，网络频道自然开始充当自杀意念的监控工具，挖掘社交内容可以提高自杀预防[6]。奇怪的社会现象正在出现，例如，网上社区就自残和模仿自杀达成协议。例如，2016年一个名为“蓝鲸游戏”的社交网络现象使用了许多任务（如自残），并导致游戏成员最终自杀。自杀是一个重要的社会问题，每年夺去成千上万人的生命。因此，有必要在受害者结束生命之前发现自杀行为并防止自杀。早期发现和治疗被认为是防止潜在自杀企图的最有效方法。

Potential victims with suicidal ideation may express their thoughts of committing suicide in fleeting thoughts, suicide plans, and role-playing. SID is to find out these risks of ntentions or behaviors before tragedy strikes. A meta-analysis conducted by McHugh et al. [7] shown statistical limitations of ideation as a screening tool but also pointed out that people’s expression of suicidal ideation represents their psychological distress. Effective detection of early signals of suicidal ideation can identify people with suicidal thoughts and open a communication portal to let social workers mitigate their mental issues. The reasons for suicide are complicated and attributed to a complex interaction of manyfactors [5], [8]. To detect suicidal ideation, many researchers conducted psychological and clinical studies [9] and classified responses of questionnaires [10]. Based on their social media data, artificial intelligence (AI) and machine learning techniques can predict people’s likelihood of suicide [11], which can better understand people’s intentions and pave the way for early intervention. Detection on social content focuses on feature engineering [12], [13], sentiment analysis [14], [15], and deep learning [16]–[18]. Those methods generally require heuristics to select features or design artificial neural network (ANN) architectures for learning rich representation. The research trend focuses on selecting more useful features from people’s health records and developing neural architectures to understand the language with suicidal ideation better.
有自杀意念的潜在受害者可能会在自杀想法、自杀计划和角色扮演中表达他们的自杀想法。SID是在悲剧发生之前找出这些意图或行为的风险。McHugh等人进行的荟萃分析[7]显示了意念作为一种筛选工具的统计局限性，同时也指出人们对自杀意念的表达代表了他们的心理困扰。对自杀意念的早期信号进行有效检测，可以识别出有自杀想法的人，并打开一个沟通门户，让社会工作者缓解他们的心理问题。自杀的原因是复杂的，并归因于许多复杂因素的相互作用[5], [8]。为了检测自杀意念，许多研究人员进行了临床研究 [9] 和问卷调查分类[10]。基于他们的社交媒体数据，人工智能（AI）和机器学习技术可以预测人们自杀的可能性[11]，这可以更好地了解人们的意图，为早期干预铺平道路。对社交内容的检测集中在特征工程[12], [13]、情感分析[14], [15]和深度学习[16]–[18]。这些方法通常需要启发式来选择特征或设计人工神经网络（ANN）结构来学习丰富的表示。研究趋势集中在从人们的健康记录中选择更有用的特征，开发神经结构，以便更好地理解带有自杀意念的语言。

Mobile technologies have been studied and applied to suicide prevention, for example, the mobile suicide intervention application iBobbly [19] developed by the Black Dog 2 Institute. Many other suicide prevention tools integrated with social networking services have also been developed, including Samaritans Radar 3 and Woebot 4. The former was a Twitter plugin that was later discontinued because of privacy issues. For monitoring alarming posts, the latter is a Facebook chatbot based on cognitive behavioral therapy and natural language processing (NLP) techniques for relieving people’s depression
and anxiety.
移动技术已经被研究并应用于自杀预防，例如黑狗研究所2开发的移动自杀干预应用iBobbly [19]。还开发了许多其他与社交网络服务相结合的自杀预防工具，包括Samaritans Radar 3和Woebot 4。前者是一个Twitter插件，后来因为隐私问题而停止使用。后者是一个基于认知行为疗法和自然语言处理（NLP）技术的Facebook聊天机器人，用于监控人们的抑郁和焦虑。
2https://blackdoginstitute.org.au/research/digital-dog/programs/ibobbly-app
3https://samaritans.org/about-samaritans/research-policy/internet-suicide/samaritans-radar
4https://woebot.io

Applying cutting-edge AI technologies for SID inevitably comes with privacy issues [20] and ethical concerns [21]. Linthicum et al. [22] put forward three ethical issues, including the influence of bias on machine learning algorithms, the prediction on time of suicide act, and ethical and legal questions raised by false positive and false negative prediction. It is not easy to answer ethical questions for AI as these require algorithms to reach a balance between competing values,issues, and interests [20].
将尖端人工智能技术应用于SID不可避免地会带来隐私问题[20]和伦理问题[21]。Linthicum等人[22] 提出了三个伦理问题，包括偏差对机器学习算法的影响、对自杀行为时间的预测以及假阳性和假阴性预测所引发的伦理和法律问题。要回答人工智能的伦理问题并不容易，因为这些问题需要算法在相互竞争的价值观、问题和利益之间达到平衡[20]。

AI has been applied to solve many challenging social problems. Detection of suicidal ideation with AI techniquesis one of the potential applications for social good and should be addressed to improve people’s wellbeing meaningfully. The research problems include feature selection on tabular and text data and representation learning on natural language. Many AI-based methods have been applied to classify suicide risks. However, there remain some challenges. There are a limited number of benchmarks for training and evaluating SID. AI-powered models, sometimes, learn statistical clues but fail to understand people’s intentions. Moreover, many neural models are lack of interpretability. This survey reviews SID methods from the perspective of AI and machine learning and specific domain applications with social impact. The categorization from these two perspectives is shown in Fig. 1. This article provides a comprehensive review of the increasingly important field of SID with machine learning methods. It proposes a summary of current research progress and an outlook of future work. The contributions of our survey are summarized as follows.
人工智能已被应用于解决许多具有挑战性的社会问题。用人工智能技术检测自杀意念是切实改善人民生活的潜在应用之一。研究问题包括表格和文本数据的特征选择和自然语言的表示学习。许多基于人工智能的方法被应用于自杀风险的分类。然而，仍然存在一些挑战。训练和评估SID的基准数量有限。人工智能驱动的模型，有时会学习统计线索，但无法理解人们的意图。此外，许多神经网络模型缺乏可解释性。这项调查从人工智能和机器学习以及具有社会影响的特定领域应用的角度回顾了SID方法。图1显示了从这两个角度进行的分类。本文全面回顾了机器学习方法在SID中日益重要的领域。对目前的研究进展进行了总结，并对今后的工作进行了展望。我们调查的贡献总结如下。

1)To the best of our knowledge, this is the first survey that conducts a comprehensive review of SID, its methods, and its applications from a machine learning perspective.
2)We introduce and discuss the classical content analysis and modern machine learning techniques, plus their application to questionnaires, EHR data, suicide notes, and online social content.
3)We enumerate existing and less explored tasks and discuss their limitations. We also summarize existing data sets and provide an outlook of future research directions in this field.
1）据我们所知，这是第一次从机器学习的角度对SID、其方法和应用进行全面回顾的调查。
2）我们介绍和讨论了经典内容分析和现代机器学习技术，以及它们在问卷调查、EHR数据、自杀笔记和在线社交内容中的应用。
3）我们列举现有的和较少探索的任务，并讨论它们的局限性。我们还总结了现有的数据集，并对该领域未来的研究方向进行了展望。

The remainder of this article is organized as follows. Methods and applications are introduced and summarized in Sections II and III, respectively. Section IV enumerates specific tasks and some data sets. Finally, we have a discussion and propose some future directions in Section V.
本文的其余部分组织如下。方法和应用分别在第二节和第三节中介绍和总结。第四节列举了具体任务和一些数据集。最后，我们在第五部分进行了讨论，并提出了未来的发展方向。

图1。自杀意念检测的分类：方法和领域。左侧部分表示方法分类，而右侧部分显示域的类别。箭头和实心点表示子类别。

2.METHODS AND CATEGORIZATION
Suicide detection has drawn the attention of many researchers due to an increasing suicide rate in recent years and has been studied extensively from many perspectives.The research techniques used to examine suicide also span many fields and methods, for example, clinical methods with patient–clinic interaction [9] and automatic detection from user-generated content (mainly text) [12], [17]. Machine learning techniques are widely applied for automatic detection.
近年来，随着自杀率的不断上升，自杀检测引起了许多研究者的关注，并从多个角度进行了广泛的研究。用于检测自杀的研究技术也跨越了许多领域和方法，例如，具有患者-诊所交互的临床方法 [9] 和从用户生成的内容（主要是文本）的自动检测方法 [12], [17]。机器学习技术广泛应用于自动检测。

Traditional suicide detection relies on clinical methods, including self-reports and face-to-face interviews. Venek et al. [9] designed a five-item ubiquitous questionnaire for the assessment of suicidal risks and applied a hierarchical classifier on the patients’ response to determine their suicidal intentions. Through face-to-face interaction, verbal and acoustic information can be utilized. Scherer [23] investigated the prosodic speech characteristics and voice quality in a dyadic interview to identify suicidal and nons