Applying statistical learning tools to personalize cardiovascular treatment

摘要 心血管疾病(CVD)的治疗通常由风险分层工具指导(以决定治疗谁)，以及 随机对照试验(以决定选择哪些治疗方法)。先前的心血管疾病研究揭示了两个主要的 改进我们的治疗方法的障碍：(1)无法获得重新校准风险的纵向队列数据 用于当地估计的分层工具(按邮政编码)，或用于主要促进心血管疾病的人 合并症(如慢性肾脏疾病)；和(2)随机试验的平均治疗效果可为 如果投射到不同于试验中“普通”参与者的个体身上，那就大错特错了。心血管疾病风险- 如果我们克服了关键障碍，分层和治疗效果评估可以得到改进和个性化 要取得进展：正确评估来自新的大型参与者数据存储库的风险和治疗效果，这 有更大的人口规模，并包括比普通队列研究更多合并疾病的患者， 并允许根据个人级别的数据开发个性化的风险/收益预测工具。我们的前辈 研究表明，我们可以通过对这些数据应用新的统计学习方法来关键地推动该领域的发展， 解决：(I)多次检测的假阳性；(Ii)依赖无法解释的标准回归 对于因素之间的非线性、复杂的相互作用；以及(Iii)在众多因素中确定最佳方法 可选的统计学习方法。我们建议将我们在这些领域的工作应用于(目标1)开发CVD 常见队列研究中样本量不足患者的风险分层工具。我们将加强 心血管疾病风险分层，包括当地调整(按邮政编码)和影响 心血管疾病风险(例如，慢性肾脏疾病)。我们还将(目标2)开发个性化治疗效果 预测工具，用于指导具有高潜在收益和高风险的心血管疾病治疗的决策，以及 试验中的个人参与者数据是可用的。我们已经从大量的参与者中获得了个人数据 显示心血管疾病风险降低和严重不良事件风险增加差异很大的随机试验 来自三类药物：非维生素K拮抗剂口服抗凝剂，强化血压治疗，以及 治疗糖尿病的钠-葡萄糖协同转运蛋白2抑制剂。我们的初步研究表明，传统的 回归方法无法区分哪些患者最有可能受益或受到此类疗法的伤害，但 我们的统计学习方法可以。最后，我们将(目标3)开发开源工具，以提高 研究人员为他们的数据集和问题选择一种最佳的统计学习方法。虽然数量众多 文献中提出了统计学习方法，这是生物医学科学家面临的一个关键问题 无法获得随机对照试验数据的问题是：我应该使用哪种方法来从观察数据中估计治疗效果？ 在我们为观测数据确定最佳推断方法的创新方法的基础上，我们将 构建一个开源工具来比较各种方法，确定哪种方法最常产生最佳结果 在标准化指标上最大限度地减少错误和最大化性能的治疗决策。

Abstract Cardiovascular disease (CVD) treatment is often guided by risk stratification tools (to decide who to treat), and randomized controlled trials (to decide which treatments to select). Prior CVD research reveals two major obstacles to improving our treatment approach: (i) longitudinal cohort data are unavailable for recalibrating risk stratification tools for local-area estimation (by zip code), or for people with major CVD-promoting comorbidities (e.g., chronic kidney disease); and (ii) the average treatment effect in randomized trials can be highly erroneous when projected onto individuals that vary from the ‘average’ participant in a trial. CVD risk- stratification and treatment effect estimation can be improved and personalized if we overcome a critical barrier to progress: correctly estimating risk and treatment effect from new, large participant data repositories, which have greater population size and include patients with more co-morbid conditions than common cohort studies, and which permit personalized risk/benefit prediction tool development from individual-level data. Our prior studies show that we can critically advance the field by applying novel statistical learning methods to this data, to address: (i) false-positives from multiple testing; (ii) the reliance on standard regressions that cannot account for non-linear, complex interactions between factors; and (iii) identifying the optimal approach among many alternative statistical learning methods. We propose to apply our work in these areas to (Aim 1) Develop CVD risk stratification tools for patients with inadequate sample sizes in common cohort studies. We will enhance CVD risk stratification to include local-area adjustment (by zip code) and major co-morbid conditions affecting CVD risk (e.g., chronic kidney disease). We will additionally (Aim 2) develop personalized treatment effect prediction tools to guide decisions for CVD therapies with high potential benefit and risk, for therapies where individual participant data from trials are available. We have obtained the individual participant data from the large randomized trials that reveal wide variations in CVD risk reduction and serious adverse event risk increase from three drug classes: non-vitamin K antagonist oral anticoagulants, intensive blood pressure treatment, and sodium-glucose co-transporter 2 inhibitors for diabetes. Our preliminary research shows that traditional regression methods cannot distinguish which patients are most likely to benefit or be harmed by such therapies, but our statistical learning methods can. Finally, we will (Aim 3) develop open-source tools to improve the ability of researchers to choose an optimal statistical learning approach for their dataset and problem. While numerous statistical learning methods have been proposed in the literature, a key problem for biomedical scientists without access to RCT data is: which method should I use to estimate treatment effects from observational data? Building on our innovative approach to identify the optimal inference method for observational data, we will construct an open-source tool to compare methods, identifying which method most often results in optimal treatment decisions that minimize error and maximize performance on standardized metrics.