Statistical Methods for Integrative Analysis of Large Scale Neuroimaging Data

摘要 随着多模态多队列神经成像数据的迅速涌现， 神经科学在阿尔茨海默病（AD）研究中，许多研究依赖于多模态神经成像数据， 识别早期诊断AD的关键图像生物标志物。尽管在数据收集方面做了很大的努力， 缺乏严格的统计方法和有效的计算工具来正确整合大量的神经成像数据， 一个统计模型，并进行推理，以解决实际问题。失踪等重要问题 受试者间异质性的数据和调整仍然没有解决。在本提案中，我们建议 建立两个综合模型，一个处理多模态数据，另一个处理纵向多队列数据。 它们将建立在一个通用的M估计框架下，该框架涵盖了许多广泛使用的统计模型， 特殊情况。我们将为这些模型提供各种推理工具，并开发有效的算法来解决 存在块缺失值的M-估计问题。在目标1中，我们提出了一个因子调整的综合 为多模态数据建模，并提供一套完整的推理工具。这些工具可以测试 一个完整的数据模式以及来自一个或多个预测因子的多个线性组合的重要性 方式。在目标2中，我们提供了一个强大的计算工具来处理多模态数据的块缺失值。 这样的工具不需要对缺失值进行特别的插补，而是依赖于一个创新的迷你模型， 批量梯度下降算法，以产生一个很好的估计。在目标3中，我们将开发一个交互因素模型 联合模拟来自多个队列的纵向数据。我们表明，这样的模型包括标准 随机效应模型作为特例，更灵活地对纵向数据进行建模， 受试者间异质性。这项研究可能会改变我们分析神经成像数据的方式。 并加深我们对阿尔茨海默病及其与公共卫生关系的认识。

Abstract Integrative analysis methods are in great needs as multimodal multi-cohort neuroimaging data rapidly emerge in neuro science. In Alzheimer's Disease (AD) studies, many research relies on multimodal neuroimaging data to identify key image biomarkers for the early diagnosis of AD. Despite great endeavors in data collection, there still lacks rigorous statistical methods and efficient computational tools to properly integrate big neuroimaging data in a statistical model and carry out inference to address practical problems. Important problems such as missing data and adjustment for between-subject heterogeneity still remain unsolved. In this proposal, we propose to build two integrative models, one handles multimodal data and the other handles longitudinal multi-cohort data. They will be built under a generic M-estimation framework that covers many widely used statistical models as its special cases. We will provide various inference tools for these models and develop efficient algorithms to solve the M-estimation problem in presence of block missing values. In Aim 1, we propose a factor-adjusted integrative model for multimodal data and provide a complete set of inference tools. These tools can test the significance of one whole data modality as well as the significance of multiple linear combinations of predictors from one or more modalities. In Aim 2, we provide a powerful computational tool to handle block missing values of multimodal data. Such a tool does not need to perform ad-hoc imputation on missing values, but rather relies on an innovative mini- batch gradient descent algorithm to yield a good estimator. In Aim 3, we will develop an interactive factor model to jointly model longitudinal data coming from multiple cohorts. We show that such a model includes the standard random effects model as a special case and is more flexible modeling the longitudinal data and accounting for the between-subject heterogeneity. The proposed research will likely transform how we analyze neuroimaging data and enhance our understanding of Alzheimer's Disease and its relation to public health.