Statistical Models of Suicidal Behavior and Brain Biology Using Large Data Sets

摘要--项目6 孔特中心的主要目标是更好地了解自杀行为的原因， 最终，将这些知识转化为有可能检测出有较高风险的患者的工具 更多致命的自杀未遂。为了实现这一目标，正在收集大量有价值的数据。这个项目是 主要涉及非常高维的数据，包括脑成像数据、DNA遗传和表观遗传学 数据和基因表达数据。特别是，我们对将这些数据用作统计中的预测因子感兴趣 模特们。我们的最终目标是建立和实施可以用来区分抑郁症自杀的模型 来自其他抑郁症患者的尝试。 最大的障碍是处理非常高维的数据，这种情况下经典的统计建模 技术不具备处理的能力。我们使用这种高维数据建模的方法寻求 实现三个主要目标：(1)提供良好的预测精度；(2)建立可解释的模型， 对各种预测因素和结果变量之间的关系给予有意义的洞察；(3) 通过验证研究确保模型的稳定性。为了实现这些目标，我们的方法将 在模型选择和拟合过程中结合正则化成分 模型复杂性。在申请中，处罚金额将通过交叉验证或其他方式确定 相关技术。在模型开发和应用中，我们将考虑广泛的结果 变量包括连续或几乎连续的(例如，抑郁的严重程度、攻击性特征)、序数 (例如自杀企图的致命性)、分类(例如诊断)和二元(是/否，例如自杀企图)，如 以及事件发生时间数据(例如，自杀未遂的时间)，这将利用生存分析技术。 这个项目特别关注可能的预测者的数量在几十个的情况下 数以千计甚至更多。现有数据和要作为Conte中心的一部分收集的数据示例包括 脑成像数据(5-羟色胺系统的正电子发射断层扫描(PET)图像、扩散张量 成像(DTI)扫描、功能磁共振成像(FMRI)扫描、磁共振波谱 (MRS)扫描)、全基因组DNA基因分型和甲基化数据以及基因表达数据。

SUMMARY-PROJECT 6 The primary objective of the Conte Center is to better understand the causes of suicidal behavior and, ultimately, to translate that knowledge into tools that have the potential for detecting patients at higher risk for more lethal suicide attempts. Towards that aim, a great deal of valuable data is being collected. This project is primarily concerned with very high-dimensional data, including brain imaging data, DNA genetic and epigenetic data, and gene expression data. In particular, we are interested in using such data as predictors in statistical models. Our ultimate goal is to build and implement models that can be used to distinguish depressed suicide attempters from other depressed patients. The greatest hurdle is dealing with very high dimensional data, a situation that classical statistical modeling techniques are not equipped to handle. Our approach to modeling with such high-dimensional data seeks to achieve three primary goals: (1) to provide good predictive accuracy; (2) to build interpretable models that can give meaningful insight into the relationship between the various predictors and the outcome variable; (3) to ensure stability of the models through validation studies. To accomplish these goals our approach will incorporate a regularization component in the model selection and fitting process by applying a penalty to model complexity. In applications, the amount of penalization will be determined by cross-validation or another related technique. In both model development and application, we will consider a wide range of outcome variables including continuous or nearly continuous (e.g., severity of depression, aggressive traits), ordinal (e.g., lethality of suicide attempts), categorical (e.g., diagnosis), and binary (yes/no, e.g., suicide attempt), as well as time-to-event data (e.g., time to a suicide attempt), which will draw on survival analysis techniques. This project focuses specifically on situations in which the number of possible predictors is in the tens of thousands or more. Examples of existing data and data to be gathered as part of the Conte Center include brain imaging data (positron emission tomography (PET) images of the serotonin system, diffusion tensor imaging (DTI) scans, functional magnetic resonance imaging (fMRI) scans, magnetic resonance spectroscopy (MRS) scans), genome-wide DNA genotyping and methylation data, and gene expression data.