Scalable Bayesian methods for big imaging data analysis

摘要！ 该提案将解决大成像数据统计分析中最及时和最重要的问题。我们 该项目的动机是“青少年大脑认知发展（ABCD）研究”，这是最大的 这是一项关于美国大脑发育和儿童健康的长期研究， 卫生研究院（NIH）。该方案的创新之处在于：1）我们开发了一种新的贝叶斯图像- 向量回归模型与新的稀疏和光滑高斯过程先验。它使执行 大脑活动的高分辨率图像与社会活动的高维向量之间的关联分析 环境因素和临床变量。据我们所知，现有的方法不能有效地 并联合分析高分辨率图像和协变量的高维向量， 系统的建模框架; 2）我们开发了一个新的贝叶斯标量图像神经网络模型， 稀疏、平滑和空间变化系数。这种新模式有很大的潜力，使更好的 与所有现有方法相比，预测青少年开始使用药物的风险;以及 重要的是，它将确定与物质使用模式相关的重要成像生物标志物。这 将提供一个更好的理解的病理物质使用启动; 3）我们提出了贝叶斯 通过组合图像-向量对多模态成像数据进行高维中介分析的模型 回归和具有修改的标量图像回归。在潜在成果框架下， 我们还将定义环境因素/电子健康记录对精神病理学的直接影响， 作为它们通过脑功能和/或结构的变化介导的间接影响。4)我们 为所有提出的模型开发可扩展的后验计算算法。这些有效的计算 工具将使在临床和转化研究中应用统计方法成为可能， 应用.我们的方法可以解决关于青少年大脑认知发展的两个关键问题：1） 他们将确定重要的童年经历和社会环境因素，如体育，视频， 游戏，社交媒体，不健康的睡眠模式和吸烟，影响大脑发育; 2）了解 推断大脑发育对物质使用开始和模式的风险，包括详细的数量， 频率、给药途径和共同使用模式。! !

ABSTRACT! This proposal will address the most timely and important issues in statistical analysis of big imaging data. Our project is motivated by "The Adolescent Brain Cognitive Development (ABCD) Study”, which is the largest long-term study of brain development and child health in the United States and is funded by the National Institutes of Health (NIH). Innovative aspects of this proposal are: 1) We develop a new Bayesian image-on- vector regression model with novel sparse and smooth Gaussian process priors. It enables to perform association analysis between high-resolution images of brain activity and high-dimensional vectors of social environmental factors and clinical variables. To the best of our knowledge no existing methods can efficiently and jointly analyze high-resolution images and high-dimensional vectors of covariates simultaneously under a systematic modeling framework; 2) We develop a new Bayesian scalar-on-image neural network model with sparse, smooth, and spatially-varying coefficients. This new model has great potential to make better predictions about the risk of an adolescent initiating substance use compared to all existing methods; and more importantly, it will identify important imaging biomarkers that are associated with substance use patterns. This will provide a better understanding of the pathology of substance use initiation; 3) We propose a Bayesian model for high-dimensional mediation analysis of multimodality imaging data by combining image-on-vector regression and scalar-on-image regression with modifications. Under the potential outcome framework, ! we will define the direct effects of environmental factors/electronic health records on psychopathology, as well as their indirect effects that are mediated through the changes in brain functions and/or structures. 4) We develop scalable posterior computation algorithms for all of the proposed models. These efficient computation tools will enable the possibility to apply the statistical methods in the clinical and translational research and applications. Our methods can address two key questions about adolescent brain cognitive development: 1) they will identify important childhood experiences and social environmental factors, such as sports, video games, social media, unhealthy sleep patterns, and smoking, that affect brain development; 2) understand the inferences of brain development on the risk of substance use initiation and patterns, including detailed quantity, frequency, route of administration, and co-use patterns. ! !