Novel Statistical Integration Methods for Multi-View Data

在大数据时代，从常规临床就诊和医学研究中收集的患者数据的数量和种类都在以前所未有的速度增长，例如从医疗记录、基因组和其他神经成像数据源。联合分析从各种来源获得的此类数据(也称为多视角数据)，整合丰富的信息，无疑为精准健康带来了新的承诺，但尽管如此，它也带来了重要的新兴挑战。首先，每个来源的数据通常都是高维的，具有比主题数量更多的特征。第二，来自不同来源的数据可以是不同的类型(例如，连续的、分类的、计数和事件发生的时间)和不同格式的数据(例如，多元变量、矩阵变量和张量变量)。最后，在大规模的数据采集过程中，数据丢失是不可避免的。虽然已经开发了许多技术来分别对这些单独的数据类型或格式中的每一种进行建模，但目前很少有方法可以在联合分析高维混合多视图数据的同时正确处理通常不完整的多视图。拟议的研究将解决这些问题，并将重点放在以下具体目标上：目标1将制定联合建模策略，在多视图数据中提取低维特征，同时区分不同数据源特定的共享和独特特征，并同时评估它们预测感兴趣结果的预测能力。目标2将开发联合建模策略来聚类多视图数据，同时分离相关和不相关的特征，并同时将相关联的子组成员资格与感兴趣的结果相关联。目标3将提出在联合建模框架中嵌入贝叶斯加性回归树的方法，以将多视图数据与结果数据相关联。这项建议侧重于开发新的贝叶斯分层建模方法，该方法非常适合于合并具有潜在变量的复杂数据结构，并重要地实现精确推理。受医学研究中混合多视图数据分析的激励，建议的研究将提供新颖且计算高效的分析工具和策略(即用户友好的R包)，使医学研究中的科学家能够从整体上探索他们的潜在数据，并使我们更接近于实现精确健康的最终目标。同时，所提出的方法论创新也广泛适用于涉及异质多视角数据的其他科学领域。

In the era of big data, both the volume and variety of patient data collected from routine clinical visits and medical research, are growing at an unprecedented pace, e.g., from medical records, genomic and neuroimaging data sources. Jointly analyzing such data obtained from various sources (also known as multi-view data) to integrate rich information has undoubtedly brought new promises for precision health, but nonetheless it also poses important and emerging challenges. First, the data from each source are typically high dimensional, with more features than the number of subjects. Second, the data from different sources could be of different types (e.g., continuous, categorical, counts and time-to-event) and of different formats, (e.g., multivariate, matrix-variate and tensor-variate). Lastly, missing data are inevitable during the large-scale data collection process. While many techniques have been developed to model each of these individual data types or formats separately, there are currently few methods that can jointly analyze high-dimensional mixed multi-view data while also properly handling often incomplete multi-view. The proposed research will address these issues and focus on the following specific aims:  Aim 1 will develop joint modeling strategies to extract low-dimensional features in the multi-view data, while distinguishing shared and unique features specific to different data sources, and simultaneous evaluate their prognostic abilities to predict the outcome of interest. Aim 2 will develop joint modeling strategies to cluster multi-view data, while separating relevant and irrelevant features, and simultaneously relate the associated subgroup memberships to the outcome of interest. Aim 3 will propose methods to embed the Bayesian Additive Regression Trees in the joint modeling framework to relate multi-view data to the outcome data.  This proposal focuses on developing novel Bayesian hierarchical modelling approaches, which are well suited to incorporate complex data structures with latent variables and importantly enable exact inference. Motivated by analyzing mixed multi-view data in medical studies, the proposed research will provide novel and computationally efficient analytic tools and strategies (i.e., user friendly R packages) to allow scientists in medical research to explore their data holistically and to bring us closer to meeting the ultimate goals in precision health. In the meanwhile, the methodological innovations proposed are widely applicable in other scientific domains that involve heterogenous multi-view data.