Leveraging ambulatory assessment data and machine learning to develop personalized prediction models of suicidal ideation

A.7.项目总结/摘要 拟议的研究将利用主动和被动的动态评估（AA）方法和机器 学习在青年临床样本中开发个性化的自杀意念（SI）预测模型， 有自杀倾向这项工作是及时和重要的，因为自杀是目前第二大 美国青少年的死亡原因SI和自杀行为的发生率在过去稳步上升 20年SI是预测模型的关键目标，因为它是可识别的，可靠的和可修改的前提 自杀行为开发能够有效预测SI的预测模型可以为just-in- 在高峰风险的精确时间进行干预，从而在发生之前预防自杀行为。 然而，尽管经过几十年的研究，我们准确预测SI的能力仍然很差，可能是因为 自杀行为是由情境化动态过程之间复杂的相互作用引起的， 具体到每个人。然而，大多数研究都试图预测SI，一种高度个人特异性的 现象，从群体层面的数据-这可以充分确定谁是在风险，但不是当一个人是 这对预防至关重要。因此，为了提高我们对SI的理解和预测， 为了采取适当地考虑个体可变性的方法（即，个性化或精确医学）， 由此模型适合人而不是相反。 动态评估、移动的计算和机器学习的进步允许收集， 管理和分析开发个性化风险预测所需的动态高分辨率数据 模型我建议联合收割机结合这些方法的进步，开发个性化的SI模型 预测.具体来说，在自杀高危青年人群中，本研究将使用生态学方法， 瞬时评估（EMA），每天两次评估SI严重程度，并收集连续被动传感器数据 连续100天都在使用智能手机。该项目将被动收集的传感器数据映射到 从经验和理论上与自杀风险相关的变量，如身体活动和流动性， 沟通和社会互动。SI严重性的每日评级和连续被动 传感器数据将为开发个性化风险计算器提供必要的数据， SI严重性的可变性作为被动传感器数据的函数。这项研究将进一步推动当前的概念化 自杀风险和预测使用先进的方法和计算方法，并提供 培训创新方法，有可能实时预测SI风险，这是响应 目标2.2和4.1的NIMH的战略计划，并持有巨大的希望，改善自杀 预防工作。

A.7. PROJECT SUMMARY/ABSTRACT The proposed research will leverage active and passive ambulatory assessment (AA) methods and machine learning to develop personalized suicidal ideation (SI) prediction models among a clinical sample of youth at high-risk for suicidality. This work is timely and important, given that suicide is currently the second-leading cause of death among youth in the U.S., and rates of SI and suicidal behavior have risen steadily over the past 20 years. SI is a critical target for predictive models, as it is an identifiable, reliable, and modifiable antecedent of suicidal behavior. Developing predictive models that can effectively predict SI may pave the way for just-in- time interventions delivered at the precise time of peak risk, thus preventing suicidal behavior before it occurs. However, despite decades of research, our ability to accurately predict SI remains poor, likely because suicidality results from a complex interaction between contextualized dynamic processes that are largely specific to each individual. Yet, most research has attempted to predict SI, a highly person-specific phenomenon, from group-level data—which can adequately identify who is at risk but not when an individual is at risk, which is critical for prevention. Thus, to improve our understanding and prediction of SI it is imperative to take an approach that properly accounts for individual variability (i.e., personalized or precision medicine), whereby the model is fit to the person rather than vice versa. Advancements in ambulatory assessment, mobile computing, and machine learning allow for the collection, management, and analysis of dynamic, high-resolution data required to develop personalized risk prediction models. I propose to combine these methodological advancements to develop personalized models of SI prediction. Specifically, among a population of youth at high-risk for suicidality, this study will use ecological momentary assessment (EMA) to assess SI severity twice daily and collect continuous passive sensor data from smartphones for 100 consecutive days. This project will map passively collected sensor data onto variables that are empirically and theoretically linked to suicide risk, such as physical activity and mobility, communication, and social interaction. This combination of daily ratings of SI severity and continuous passive sensor data will provide the necessary data to develop personalized risk calculators that model each person’s variability in SI severity as a function of passive sensor data. This study will further current conceptualizations of suicide risk and prediction using advanced methodological and computational approaches, and provide training in innovative methods that have the potential to predict SI risk in real-time, which is responsive to Objectives 2.2 and 4.1 of NIMH’s Strategic Plan and holds tremendous promise for improving suicide prevention efforts.