Bayesian machine learning for complex missing data and causal inference with a focus on cardiovascular and obesity studies

项目摘要 该提案将开发贝叶斯机器学习方法，通过贝叶斯非参数(BNP)来处理 不可忽视的缺失(在结果和协变量中)，并对电子健康记录进行因果推断 (EHR)，以解决多变量纵向数据中的缺失问题，以及因果调解问题。丢失 数据在临床研究中仍然是一个问题，特别是对于使用EHR的研究。在临床研究中，更多 E↵排序被用来试图将错失量降到最低，但它仍然是一个问题，而错失率是 在基于EHR的研究中，这是一个持续存在的问题(而且更不可控)。此外，在这方面的工作有限。 在电子病历中使用辅助信息，以增强处理丢失数据的能力。解决问题的方法 多变量纵向数据中的缺失性不发达，并且与许多临床试验设置相关 从成本e↵经济分析到不完全时变辅助协变量(或混杂因素)再到因果中介 多个令人感兴趣的结果。治疗↵有效性的机制对行为学特别感兴趣。 审判。具体地说，不同的过程是如何调节干预的过程的？fi↵？这可以方便地构建 未来的干预措施。然而，确定这些“调解人”对结果的因果关系是不正确的。我们将 开发新的方法来在复杂的群组随机试验环境中确定这些↵影响 几乎没有做过什么工作。对于所有这些设置，贝叶斯方法是理想的，因为它允许人们适当地 描述有关不可验证的fi假设(存在于所有这些问题中)的不确定性，并允许 贝叶斯非参数模型的fl灵活性。求解BNP问题的MCMC算法有时收敛速度较慢， 对于大的n是站不住脚的。我们将扩展现有的方法来解决这两个复杂情况 考虑到数据的大小和复杂性不断增加，对所有考虑的应用程序以及一般应用程序都很重要。 这些方法是由NHLBI资助的几项研究推动的，这些研究的PI是这项提案的联合调查者， 将被开发来帮助回答许多重要的临床问题，包括行为机制 体重管理的变化以及与护理挂钩(和参与度)对治疗e-↵效率的影响 用于测量血压结果。这些方法也将帮助我们评估潜在的协同效应，当药物↵ 与潜在的糖尿病致病因子同时使用，对癌症预后的影响是否因不同而不同 Di↵曾经做过减肥手术。 整个研究团队的合作历史将有助于产生最好的科学和 促进我们的方法学和临床fi标准的传播。我们将发布这些方法的代码 (通过PI的GitHub页面和软件论文)，以确保调查人员随时可以使用这些方法 参与心血管、肥胖、糖尿病和癌症研究。

Project Summary This proposal will develop Bayesian machine learning approaches via Bayesian nonparametrics (BNP) to handle nonignorable missingness (in outcomes and covariates) and conduct causal inference for electronic health records (EHRs), to address missingness in multivariate longitudinal data, and for causal mediation problems. Missing data remains a problem in clinical studies and in particular, for studies using EHRs. In clinical studies, more e↵ort is spent to try to minimize the amount of missingness, but it still remains a problem and missingness is a constant issue (and less controllable) in studies based on EHRs. In addition, there has been limited work on the use of auxiliary information in EHRs that can enhance the ability to deal with missing data. Approaches for missingness in multivariate longitudinal data is underdeveloped and relevant across many clinical trials settings from cost e↵ectiveness analysis to incomplete time-varying auxiliary covariates (or confounders) to causal mediation to multiple outcomes of interest. The mechanisms of treatment e↵ectiveness are of particular interest in behavioral trials. Specifically, how do di↵erent processes mediate the e↵ect of an intervention? This can facilitate constructing future interventions. However, determining the causal e↵ect of such 'mediators' on outcomes is di"cult. We will develop new approaches to identify these e↵ects in the complex setting of cluster randomized trials for which little work has been done. For all these settings, a Bayesian approach is ideal as it allows one to appropriately characterize uncertainty about unverifiable assumptions (which are present in all these problems) and allows the flexibility of Bayesian nonparametric models. MCMC algorithms for BNP can sometimes converge slowly and can be untenable for large n. We will extend existing approaches to address both these complications which will be important for all the applications considered and in general, given the increasing size and complexity of data. The methods are motivated by several NHLBI funded studies, whose PI's are co-investigators on this proposal, and will be developed to help answer numerous important clinical questions including the mechanisms of behavior change in weight management and the impact of linkage (and engagement) to care on treatment e↵ectiveness for blood pressure outcomes. The methods will also help us evaluate potentially synergistic e↵ects when drugs with potential diabetogenic e↵ects are used concomitantly and whether the impact on cancer outcomes varies by di↵erent bariatric surgeries. The history of the the collaborations among the entire study team will help produce the best science and facilitate dissemination of our methodological and clinical findings. We will disseminate code for these methods (via the PI's github page and software papers) to ensure the methods will be readily usable by investigators involved in cardiovascular, obesity, diabetes, and cancer studies.