Semiparametric Analysis of Big Censored Data

项目摘要 该项目的广泛的长期目标是开发半参数回归方法，用于分析 删失数据，这是在慢性疾病的生物医学研究中经常遇到的。此续订 应用程序的重点是解决涉及匈奴的大数据分析中的计算挑战， 数千到数千万的个体，数千到数千万的变量。规格 目标是开发：（1）基于半参数效率的通信效率，分布式提升算法， 用于将考克斯比例风险模型拟合到各种大删失数据的得分函数;（2）a 通信效率高的分布式增强算法，该算法将随机特征集选择方案嵌入到 高维环境中的变量选择;（3）通信效率，分布式提升算法， 将具有潜在因子的考克斯模型拟合到具有缺失值的多种类型的高维特征;以及（4） 一种分布式EM算法，该算法结合了用于矩阵求逆的预条件共轭梯度法， 数值积分的拉普拉斯近似的一种新的锡永和随机效应考克斯的拟合 有大量基因相关个体的模型。每一个目标都涉及到重要的新挑战， 今天的大型生物医学研究所产生的问题。所提出的方法和算法是基于似然 和其他合理的统计原则。所需的渐近性质的估计将建立钻机- 通过创新性地使用现代经验过程理论和其他先进的数学工具。的 我们将通过模拟真实的数据，对提出的方法和算法进行广泛的评估 并在云计算环境下进行了测试，提供了高数据安全保障和可扩展的COM， 建设基础设施。此外，这些方法和算法将应用于我们正在进行的生物医学研究， 包括NHLBI的Trans-Omics for Precision Medicine项目和英国生物库。最后，高效，可靠， 并将制作具有适当文档的方便用户的开放源码软件。的总体影响 拟议的工作将是为生存分析创造新的范例，推进美国的生物医学研究， 美国和其他国家，并加快寻找有效的战略，以预防和治疗心血管疾病 疾病、癌症、艾滋病和其他对全球公共卫生至关重要的疾病。

Project Summary The broad, long-term objectives of this project are to develop semiparametric regression methods for analyzing censored data, which are commonly encountered in biomedical research on chronic diseases. This renewal application is focused on addressing the computational challenges in the analysis of big data involving hun- dreds of thousands to tens of millions of individuals with thousands to tens of millions of variables. The specific aims are to develop: (1) a communication-efficient, distributed boosting algorithm based on semiparametric effi- cient score functions for fitting the Cox proportional hazards model to a wide variety of big censored data; (2) a communication-efficient, distributed boosting algorithm that embeds a random feature-set selection scheme into variable selection in high-dimensional settings; (3) a communication-efficient, distributed boosting algorithm for fitting a Cox model with latent factors to multiple types of high-dimensional features with missing values; and (4) a distributed EM algorithm that incorporates both the preconditioned conjugate-gradient method for matrix inver- sion and a novel modification of the Laplace approximation to numerical integration for fitting a random-effect Cox model with a large number of genetically related individuals. Each of these aims addresses important new chal- lenges arising from today's big biomedical studies. The proposed methods and algorithms are based on likelihood and other sound statistical principles. The desired asymptotic properties of the estimators will be established rig- orously through innovative use of modern empirical process theory and other advanced mathematical tools. The proposed methods and algorithms will be evaluated extensively through simulation studies mimicking real data and tested in the cloud computing environment, which provides high data security guarantees and scalable com- puting infrastructures. In addition, the methods and algorithms will be applied to our ongoing biomedical studies, including the NHLBI Trans-Omics for Precision Medicine program and the UK Biobank. Finally, efficient, reliable, and user-friendly open-source software with proper documentation will be produced. The overall impact of the proposed work will be to create new paradigms for survival analysis, advance biomedical research in the United States and other countries, and accelerate the search for effective strategies to prevent and treat cardiovascular diseases, cancers, AIDS, and other diseases of utmost importance to global public health.