Deep Learning-based Emulation Analysis: Methodological Developments and Case Studies

项目摘要 客观地量化药物、设备和治疗程序对生存的相对有效性 心血管疾病(CVD)的结局，严格设计和执行随机临床试验 (RCT)仍然是黄金标准。然而，对于许多问题，RCT要么失败了，要么不可行。 幸运的是，电子病历(EMR)和保险索赔数据库的快速发展使得 有可能挖掘大量观测数据，并有效补充随机对照试验。在可用的 旨在得出RCT类型结论的观测数据分析技术，仿真已经出现为 考虑到它类似试验的体系结构、可解释性和可伸缩性，它特别有吸引力。它已被应用于心血管疾病 二十多年，并导致了许多重要的发现。 这项研究有两个目的。第一个目标是开发基于深度学习的仿真分析 流水线、方法和软件。现有的大多数仿真分析都是基于“经典”回归 技巧。最近，我们团队率先开发了基于动态链接库的仿真分析，并应用于 心血管病。与回归相比，DL具有更好的模型拟合度和灵活的适应性 未指明的非线性效应。在我们最近成功的基础上，这个项目将在方法上取得重大进展 通过开发具有更有效估计的基于动态链接库的尖端仿真分析(这具有更多的 期望的健壮性和显著改进的稳定性和可解释性)，全面和有效 推论(对于对治疗效果作出明确结论是必不可少的，但在大多数DL中是缺失的 学习)和友好的软件(便于广泛使用)。这种方法论的努力可以大大扩展 仿真分析、深度学习、因果推理、观测数据分析和医学的范围 记录/保险索赔数据分析。第二个目标是进行两个具有高度临床意义的案例研究。 第一个案例研究是关于评价ICD(植入式心脏复律除颤器)对全因心脏复律的影响 退伍军人事务部老年人口死亡率。旨在解决以下问题的临床试验 由于注册人数较少，此问题失败。作为退伍军人事务部因果倡议的一部分，模拟被提议为 一个可行的解决方案来“取代”试验。第二个案例研究是关于评价两种药物的比较疗效。 利伐沙班与达比卡特兰在医疗保险人群中对房颤患者死亡率的比较 在FDA批准和已经广泛使用的两种药物的情况下，不太可能进行随机对照试验。不仅仅是直接告知 临床实践，这一目标下的研究也可以补充和促进VA因果倡议 作为拟议方法未来应用的原型。

Project Summary To objectively quantify the relative effectiveness of drugs, devices, and treatment procedures on survival outcomes of cardiovascular diseases (CVDs), rigorously designed and executed randomized clinical trials (RCTs) remain as the gold standard. However, for many problems, RCTs either have failed or are not feasible. Luckily, the fast development of electronic medical record (EMR) and insurance claims databases makes it possible to mine a large amount of observational data and efficiently complement RCTs. Among the available observational data analysis techniques that aim to draw RCT-type conclusions, emulation has emerged as especially attractive, given its trial-like architecture, interpretability, and scalability. It has been applied to CVDs for over twenty years and led to many important findings. This study has two aims. The first aim is to develop a deep learning (DL)-based emulation analysis pipeline, methods, and software. Most of the existing emulation analyses are based on “classic” regression techniques. Very recently, our group was the first to develop DL-based emulation analysis with application to CVDs. Compared to regression, DL excels by having superior model fitting and flexibly accommodating unspecified nonlinear effects. Built on our recent success, this project will methodologically significantly advance by developing cutting-edge DL-based emulation analysis with more effective estimation (that has the much- desired robustness property and significantly improved stability and interpretability), comprehensive and valid inference (which is essential for making definitive conclusions on treatment effects but missing in most DL studies), and friendly software (to facilitate broad utilization). This methodological effort can substantially expand the scope of emulation analysis, deep learning, causal inference, observational data analysis, and medical record/insurance claims data analysis. The second aim is to conduct two clinically highly significant case studies. The first case study is on evaluating the effect of ICD (Implantable Cardioverter Defibrillator) on all-cause mortality in the VA (Department of Veterans Affairs) elderly population. The clinical trial targeting at addressing this problem failed because of low enrollment. As part of the VA CAUSAL Initiative, emulation was proposed as a viable solution to “replace” the trial. The second case study is on evaluating the comparative efficacy of Rivaroxaban versus Dabigatran on the mortality of AF (atrial fibrillation) patients in the Medicare population, for which an RCT is unlikely with both drugs FDA-approved and already popularly used. Beyond directly informing clinical practice, research under this aim can also complement and advance the VA CAUSAL Initiative as well as serve as a prototype for future applications of the proposed approach.