Ethics and Machine Learning: Adolescent Suicide Prediction on Social Media

自杀是英国10-24岁青少年的主要死亡原因(公共卫生英格兰，2018年)。对于临床医生来说，识别自杀风险增加的青少年是一项重大挑战，因为传统的评估工具(例如自杀意念量表和Beck的自杀启发量表)的积极预测价值较低(PPV；Runeson等人，2017)。社交媒体的快速增长及其与青少年日常生活的日益融合，为研究人员提供了通过新的行为数据来源预测自杀风险的新机会(例如，McClellan，Ali，Mutter，Kroutil，&Landwehr，2017；Torous等人，2018)。这一过程由机器学习算法(MLA)实现，该算法使用归纳方法生成有关社交媒体数据集中发现的模式的新理论(McClellan等人，2017年)。这些方法被大学、医院和社交媒体公司使用。在美国，Facebook使用MLA扫描账户，寻找有自杀风险的人(Kaste，2018)。虽然他们的算法没有披露，但如果Facebook认为一个人有自杀的“迫在眉睫的风险”，它会通知当地的紧急响应人员，后者可能会反过来拜访这个人进行健康检查。据报道，2017年，Facebook向当地应急人员发出了超过3500次的警报(Kaste，2018年)。伦理学家已经研究了成年人被算法认为有自杀风险时可能出现的担忧，但他们的探索仍然是理论上的(戈麦斯·德·安德拉德、波森、穆列洛、多纳霍和瓜达尼奥，2018；麦科南、克莱顿和沃尔什，2018；塔克、塔克特、格利克曼和雷格，2018)。此外，还没有类似的讨论或实证研究侧重于这种做法对青少年的影响。将MLA与社交媒体数据一起使用可能会给青少年带来独特的伦理问题，因为与成年青少年相比，他们在社交媒体上花费的时间更多，使用的平台也更多(Smith&Anderson，2018)。此外，青少年和成年人之间的自杀风险和自杀企图在形式和死亡率上也不同(Parellada等人，2008年)。关于同意行医的立法可能是与年龄有关的差异的另一个原因，因为16岁以下的年轻人只能提供与其能力相称的独立同意，因此不能自动有权同意行医(Gillick诉西诺福克案和Wisbech AHA案，1985年)。因此，法定监护人可能会代表他们的孩子做出决定。在这一背景下考虑一系列道德问题，对于最大限度地减少对青少年的伤害的可能性和最大限度地提高干预措施的任何益处是至关重要的。因此，我在精神病学系的DPhil的目的是研究当青少年被基于算法的自杀预测和干预作为目标时可能出现的伦理问题。大多数关于成人自杀预测的伦理学文献使用原则驱动的方法和三个最重要的生物伦理学原则：自主、慈善和隐私(戈麦斯·德·安德拉德等人，2018年；McKernan等人，2018年；塔克等人，2018年)。在我的DPhil的第一阶段，我将创建一个基于这些原则的生命伦理学框架，用于在社交媒体上使用MLA预测青少年自杀。

Suicide is a leading cause of death for British adolescents aged 10-24 (Public Health England, 2018). Identifying adolescents who are at increased risk of suicide is a major challenge for clinicians because traditional assessment instruments (e.g. the Scale for Suicide Ideation and Beck's Suicide IntentScale) have low Positive Prediction Value (PPV; Runeson et al., 2017). The rapid growth of social media combined with its increasing integration into the daily lives of adolescents presents researchers with new opportunities for predicting suicide risk through new sources of behavioral data (e.g. McClellan, Ali,Mutter, Kroutil, & Landwehr, 2017; Torous et al., 2018). This process is enabled by Machine Learning Algorithms (MLAs), which use inductive methods to generate new theories about patterns found in social media datasets (McClellan et al., 2017). These methods are used by universities, hospitals, and socialmedia companies. In the United States, Facebook uses MLAs to scan accounts for people at-risk of suicide (Kaste, 2018). While their algorithm is undisclosed, if it deems a person to be at "imminent risk" of suicide Facebook will alert local emergency responders, who may in turn visit the person for a wellnesscheck. In 2017 Facebook reportedly alerted local emergency responders over 3,500 times (Kaste, 2018). Ethicists have studied the concerns that may arise when adults are deemed "at-risk" of suicide by algorithms, however their exploration remains theoretical (Gomes de Andrade, Pawson, Muriello,Donahue, & Guadagno, 2018; McKernan, Clayton, & Walsh, 2018; Tucker, Tackett, Glickman, & Reger, 2018). In addition, there have been no comparable discussions or empirical studies focusing on the impact of this practice on adolescents. Using MLAs with social media data may cause unique ethical concerns foradolescents who spend more time on social media and use a greater variety of platforms compared to their adult counterparts (Smith & Anderson, 2018). In addition, suicide risk and attempts differ in form and fatality between adolescents and adults (Parellada et al., 2008). Legislation on consent to medical practicesmay be another cause of age-related differences, as young people under the age of 16 can only provideindependent consent proportionate to their competence and therefore are not automatically entitled to give consent to medical practices (Gillick v. West Norfolk and Wisbech AHA, 1985). As such, a legal guardian may be making decisions on their child's behalf. Considering the range of ethical issues within this context is fundamental to minimize the potential for harm and maximize any benefits of interventions for adolescents. Therefore, the aim of my DPhil within the Department of Psychiatry is to study the ethical questions that may arise when adolescents are targeted by algorithm-based suicide predictions and interventions. Most of the ethics literature on suicide prediction with adults uses a principle-driven approach and three overarching bioethical principles: autonomy, beneficence, and privacy (Gomes de Andrade et al., 2018; McKernan et al., 2018; Tucker et al., 2018). Within the first phase of my DPhil, I will create a bioethics framework founded on these principles for the use of MLAs in adolescent suicide prediction on social media.