Interpretable Deep Forecasting of Hazardous Substance Use during High School

摘要 药物滥用障碍 (SUD) 影响着超过 2000 万美国人，造成个人冲突和损失。一个专业 SUD 风险因素是在高中 (HUSH) 期间危险使用物质 (HU)，此时大脑继续 发展，使其对环境因素特别敏感。确定 HUSH 的影响和风险 通常侧重于固定（即创伤、人口统计）和可修改（即， 心理健康、情绪、环境、行为、睡眠）因素，偶尔还有大脑发育特征 区分物质使用和不使用的群体。结果对风险的改善有限 评估。因此，我们提出 HUSH 研究的范式转变，取代测量选择和 通过将个体映射到综合的多维度量来进行群体分裂。的目标 我们新颖的数据驱动流程是确定预测 HUSH 的固定和可修改因素的星座 个人。由于我们的分析基于包括脑成像在内的公共数据集，因此我们将记录 这些星座与神经回路的相互作用，以确定 HUSH 的神经机制基础。 影响有害物质使用的因素有很多，例如性别和家族史等固定因素 南苏丹;不健康的睡眠习惯、同伴压力和冒险倾向等可改变的因素；和大脑 发展的特点是情绪和控制网络之间的非典型失衡。我们将建模 这种异质性通过机器（深度）学习技术识别在线测量星座 我们关于预防的假设，即与异常相互作用的可修改行为 神经适应预测 HUSH 的启动。目标 1 将预测在高风险的最后几年开始 HUSH 学校基于全国无/低物质使用者 16 岁之前最接近的访问 青春期酒精与神经发育联盟（NCANDA，N = 350）并确认了以下研究结果 更大的青少年大脑认知发展队列（ABCD，N>11K）。 HUSH 将定义为 通过年度自我报告记录物质使用标准，并通过每周调查进行完善 手机。目标 2 将创建一个自我监督学习模型，明确跟踪跨时间的交互 可改变的行为、固定因素和对假设检验很重要的大脑回路。我们将交叉验证 该模型通过识别每个高中年级的 HUSH 并根据高中之前收集的数据进行预测 学校。我们将探索在获取数据时动态更新模型以预测弹性，即青年 尽管存在创伤等危险因素，但在高中期间仍避免使用有害物质的人 以及不愉快的新冠大流行经历。每个目标都与有关以下因素的假设检验相关： 可以进行更改以减轻 HUSH 的风险。 该项目将是第一个提供在个人层面准确预测 HUSH 风险的模式的项目。 准确计算这种风险是加强预防工作的基础。

ABSTRACT Substance use disorder (SUD) affects over 20 million Americans, causing personal strife and cost. A major SUD risk factor is hazardous use (HU) of substances during high school (HUSH), when the brain continues to develop, rendering it especially sensitive to environmental factors. Identifying the effects of and risk for HUSH generally focuses on selected interactions between fixed (i.e., trauma, demographics) and modifiable (i.e., mental health, emotion, environment, behavior, sleep) factors, and occasionally brain development features differentiating substance using and non-using cohorts. Results have yielded limited improvements to risk assessment. Thus, we propose a paradigm shift in the study of HUSH, replacing measurement selection and population splitting with mapping individuals to comprehensive multi-dimensional measures. The objective of our novel data-driven process is to identify constellations of fixed and modifiable factors forecasting HUSH in individuals. As our analysis is based on public data sets that include brain imaging, we will document interactions of those constellations with neural circuits to determine neuromechanistic underpinnings of HUSH. Myriad factors influence hazardous substance use, such as the fixed contributors of sex and family history of SUD; the modifiable factors of unhealthy sleep habits, peer pressure, and risk-taking propensity; and brain development characterized by an atypical imbalance between emotion and control network. We will model this heterogeneity via machine (deep) learning technology identifying constellations of measurements in line with our hypotheses regarding prevention, i.e., modifiable behaviors interacting with anomalies in neuroadaptation forecasting HUSH initiation. Aim 1 will forecast initiation of HUSH in the last years of high school based on the closest visit before turning age 16 years for no/low substance users in National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA, N =350) and confirm findings on the larger Adolescent Brain Cognitive Developmental cohort (ABCD, N>11K). HUSH will be defined by substance use criteria recorded through annual self-reports and refined by weekly surveys administered via cell phones. Aim 2 will create a self-supervised learning model explicitly tracking over time interactions across modifiable behaviors, fixed factors, and brain circuits important for hypothesis testing. We will cross-validate the model by identifying HUSH for each high school year and forecast based on data collected prior to high school. We will explore dynamically updating the model as data are acquired to predict resilience, i.e., youth who abstain from hazardous substance use during high school, despite having risk factors such as traumatic and untoward COVID pandemic experiences. Each aim is linked to hypothesis testing concerning factors that can be altered to mitigate the risk of HUSH. This project will be the first to provide patterns accurately forecasting the risk of HUSH on an individual level. Accurately computing this risk would be foundational for enhancing prevention efforts.