Enhanced Identification of Ocular Phenotypes and Outcomes in Electronic Health Record Data

摘要 从纸质图表到电子健康记录（EHR）的过渡，计算的进步 电力和存储容量，以及复杂机器的可用性和可访问性 学习算法已经彻底改变了研究人员利用大数据的能力， 用它来回答各种重要的临床问题。然而，最大化所有人的效用 这些来自EHR和临床注册的丰富临床数据的预测是基于研究人员的能力， 准确识别哪些患者患有特定疾病;准确分类条件 根据他们的疾病严重程度;确定哪些患者正在改善、稳定或恶化; 并适当识别和量化临床相关结果。目前，几乎所有 在眼科中使用大数据的研究人员完全依赖于管理计费 识别常见眼部疾病和关注结果的代码。然而，研究表明， 仅仅依靠账单代码充满了局限性，并且没有充分利用 结构化和自由文本EHR数据中常规捕获的大量有用信息。在这 我们开发、严格测试和验证一种创新方法的建议， 为了更准确地识别和分类患有常见视力威胁性眼病的患者， 并捕获从较轻到较严重疾病状态的转变以及感兴趣的关键结果。 根据我们进行的初步研究，我们认为我们的方法可以增强眼部 表型鉴定实质上比完全依赖计费代码更准确。在 目的1，我们使用这种方法对EHR数据进行识别和分类， 不可逆视力丧失的常见原因-青光眼、糖尿病视网膜病变和黄斑 退化在目标2中，我们通过尝试识别新的 这3种情况的形式;我们将使用聚类分析来识别临床特征组 与这三种共同分离的眼部疾病相关。我们还将测试 其中一些集群与更好或更差的临床结果相关。在目标3中，我们应用 我们的方法是在EHR数据中识别关键的眼部结果，如疾病稳定性， 对于3种感兴趣的眼部疾病中的每一种，从较低阶段进展到较晚期阶段。通过 为了实现这一建议的目标，我们将允许全国和世界各地的研究人员 在EHR和临床登记中更准确地识别重要的眼部疾病和结局 数据集。这将作为一个关键的构建块，允许研究人员将大数据纳入 机器学习和人工智能应用，基因型-表型关联研究， 临床试验的患者招募以及许多其他临床和转化研究项目。

ABSTRACT The transition from paper charts to electronic health records (EHRs), advances in computing power and storage capacity, and the availability and accessibility of sophisticated machine learning algorithms have revolutionized the ability for researchers to tap into Big Data and make use of it to answer all sorts of important clinical questions. However, maximizing the utility of all of this rich clinical data from EHRs and clinical registries is predicated on the ability for researchers to accurately identify which patients have specific diseases; to accurately classify conditions based on their disease severity; to ascertain which patients are improving, stable, or deteriorating; and to appropriately identify and quantify clinically relevant outcomes. Currently, nearly all researchers who work with Big Data in ophthalmology rely exclusively on administrative billing codes to identify common ocular diseases and outcomes of interest. Yet, research has shown that sole reliance on billing codes is fraught with limitations and does not take full advantage of the plethora of useful information routinely captured in structured and free-text EHR data. In this proposal we develop, rigorously test, and validate an innovative approach to permit researchers to more accurately identify and classify patients with common sight-threatening ocular diseases and capture transitions from less to more severe disease states and key outcomes of interest. Based on preliminary studies we performed, we believe our approach to enhanced ocular phenotype identification is substantially more accurate than exclusive reliance on billing codes. In Aim 1, we use this approach to EHR data to identify and categorize patients with 3 of the most common causes of irreversible vision loss—glaucoma, diabetic retinopathy, and macular degeneration. In Aim 2, we extend enhanced phenotype identification by trying to identify novel forms of these 3 conditions; we will use cluster analysis to identify groups of clinical features associated with these 3 ocular diseases that co-segregate together. We will also test whether some of these clusters are associated with better or worse clinical outcomes. In Aim 3, we apply our approach to identify key ocular outcomes in EHR data such as disease stability and progression from less to more advanced stages for each of the 3 ocular diseases of interest. By fulfilling the aims of this proposal, we will permit researchers throughout the country and the world to more accurately identify important ocular diseases and outcomes in EHR and clinical registry datasets. This will serve as a key building block to permit researchers to incorporate Big Data into machine learning and artificial intelligence applications, genotype-phenotype association studies, patient recruitment for clinical trials, and many other clinical and translational research projects.