Statistical and Machine Learning Methods to Improve Dynamic Treatment Regimens Estimation Using Real World Data

项目总结/摘要 2型糖尿病（T2 D）是一种全球流行病，影响全球约4.62亿人。Cur- 租赁医疗指南主要依赖于随机对照试验（RCT）的数据， 效果，这远远不足以为真实的世界的患者做出个性化的决定。这种限制是 更糟糕的是，在异质性人群中发现动态治疗方案（DTR）， 在疾病过程的一个或多个阶段做出决定。这种局限性可以通过补充来部分解决。 用真实的世界数据（RWD）补充RCT数据，如疾病登记、前瞻性观察性研究， 调查和电子健康记录，以改善医疗决策。尽管承诺结合 RWD和RCT在方法和算法开发方面存在若干重大挑战。这些包括 当应用于真实的世界患者时，RCT结果缺乏普遍性或实用性;偏倚 未观察到的混杂因素;以及对长期副作用/风险的担忧。该提案旨在解决每个 这些挑战。具体来说，在目标1中，我们通过提出一个新的框架来解决可推广性问题 使用来自RWD的证据来改善试验中的DTR学习。该框架使用RWD来选择信息， 积极的定制功能，平衡人口分布，并通过双重稳健提高统计效率 估计。在目标2中，为了提高DTR的实用性，我们提出了一种鲁棒的方法来首先推断个体 RWD的治疗选择/偏好，然后使用 试验数据。由此产生的DTR不仅在统计学上有效，而且与患者/临床医生的偏好相一致 在真实的世界人口中。在目标3中，为了减少由于RWD中隐藏的混杂因素而导致的偏倚，我们建议联合 将试验数据与RWD结合起来的半参数模型;我们提出的模型允许不同的幅度 治疗效应量和控制RWD中隐藏混杂因素导致的可能偏倚。目标4：解决 考虑到对长期风险的关注，我们考虑了一种估计DTR的一般程序，以最大限度地提高效率 结果，同时确保与推荐DTR相关的长期副作用保持低于 一定的门槛。然后，我们提出了一种新的同时学习算法来估计最佳DTR跨 所有阶段。对于所有四个目标，我们将使用以下工具提供严格的假设和理论依据： 浓度不等式，统计学习理论，经验过程和半参数推断。我们将 进行广泛的模拟研究，以研究所提出的方法在各种情况下的表现， 环，并比较其性能与现成的方法。我们将应用所提出的方法来估计 使用临床试验数据和取自哥伦比亚大学电子健康记录的RWD的T2 D DTR， 俄亥俄州州立大学医疗中心以及Allof Us精准医学研究。我们的方法和发现 通过软件开发进行宣传;该软件将根据用户反馈经常更新。

Project Summary/Abstract Type 2 diabetes (T2D) is a global epidemic affecting approximately 462 million individuals world-wide. Cur- rent medical treatment guidelines rely largely on data from randomized controlled trials (RCTs) that study average effects, which is far from adequate for making individualized decisions for real world patients. This limitation is even worse for discovering dynamic treatment regimens (DTRs) in a heterogeneous population where treatment decisions are made over one or more stages of disease course. This limitation can be partially addressed by sup- plementing RCT data with real world data (RWD), such as disease registries, prospective observational studies, surveys and electronic health records, to improve medical decision making. Despite of the promise of combining RWD and RCT, there are several significant challenges in method and algorithm development. These include lack of generalizability or practical utility for the findings from RCTs when applied to real world patients; bias due to unobserved confounders; and concern about long-term side effects/risks. This proposal aims to address each of these challenges. Specifically, in Aim 1, we address the generalizability issue by proposing a novel framework that uses evidence from RWD to improve learning DTRs in the trials. The framework uses RWD to select infor- mative tailoring features, balance population distributions and improve statistical efficiency through doubly robust estimation. In Aim 2, to improve the practical utility of DTRs, we propose a robust method to first infer individual treatment choice/preference from RWD, then incorporate this estimated preference into learning DTRs using the trial data. The resulting DTRs are not only statistically valid but also compatible with patient/clinician preference in real world populations. In Aim 3, to lessen the bias due to hidden confounders in RWD, we propose joint semiparametric models to combine the trial data with RWD; the models we propose allow different magnitudes of treatment effect sizes and control for possible bias due to hidden confounders in RWD. In Aim 4, to address the concern about long-term risks, we consider a general procedure for estimating DTRs that maximizes efficacy outcomes while ensuring that long-term side effects associated with the recommended DTRs remain below a certain threshold. We then propose a novel simultaneous learning algorithm to estimate the optimal DTRs across all stages. For all four aims, we will provide rigorous assumptions and theoretical justifications using tools from concentration inequalities, statistical learning theory, empirical processes and semiparametric inference. We will conduct extensive simulation studies to study the performance of the proposed approaches in a variety of set- tings, and compare their performance with off-the-shelf methods. We will apply the proposed methods to estimate DTRs for T2D using clinical trial data and RWD taken from electronic health records in Columbia University and Ohio State University medical centers as well as Allof Us precision medicine study. Our methods and findings will be publicized through software development; the software will receive frequent updates based on user feedback.