Accelerating research to advance care for adults with congenital heart disease through development of validated scalable computational phenotypes

项目摘要 20世纪后半叶治疗先天性心脏病（CHD）的手术的出现改变了 从婴儿期致命疾病的缓解到终身慢性病的管理， 成年目前，美国有超过150万成年CHD患者。这些患者 心血管和其他医学合并症的负担很大， 不良后果的风险，如心律失常、心力衰竭、脑血管意外和过早死亡。 这一人群的出现需要新的临床护理模式以及新的治疗方法的开发。 研究工具和基础设施，以满足这些患者的独特特征和医疗保健需求。 成人CHD的特点是治疗策略的高度复杂性、时代依赖性异质性， 和随时间变化的终身疾病的影响。这个迅速增长的人口是不足的研究， 对组成疾病的病理生理学仍不完全了解。计费和其他行政 电子病历中可用的代码对CHD诊断既不敏感也不特异， 充分描述了许多其他突出的临床特征。因此，大型管理系统中的结构化数据 数据库不太适合研究成人CHD患者，即使目标只是识别一组CHD患者， 有明确诊断的患者。这是研究工作的主要障碍， 这是迄今为止进行的有限的基于人群的研究的根本障碍。成人CHD研究将 从建立协调、大规模、多中心数据集的方法中受益匪浅。 虽然计费代码是不够的，准确分类成人CHD所需的信息已经 在电子病历中以临床笔记的形式提供，主要由非结构化（“免费”） 短信了手动数据提取是费力的、资源密集型的，并且因此是不可扩展的。我们建议申请 对电子病历中的非结构化文本采用尖端的自然语言处理方法， 为成人CHD研究的基本变量开发可计算分类器。我们将使用两个 波士顿儿童医院和布里格姆妇女医院的独特机构数据资源， 已经填充了专家裁定的标签，以训练分类器用于定义不明确的关键表型 行政代码。这些分类器将在范德比尔特的独立患者队列中进行验证 大学医学中心，并在新的疾病特异性风险预测模型中进行测试。这项工作承诺， 通过大规模增加可以研究的患者队列的规模， 建立一个基础，为这一得不到充分服务的人口提供更好的循证决策支持。

PROJECT SUMMARY The advent of surgery to treat congenital heart disease (CHD) in the second half of the 20th century shifted the care paradigm from palliation of disease fatal in infancy to management of lifelong chronic disease through adulthood. There are now more than 1.5 million adults with CHD living in the United States. These patients have a substantial burden of cardiovascular and other medical comorbidities, as well as markedly increased risk for adverse outcomes such as arrhythmia, heart failure, cerebrovascular accident, and premature death. The emergence of this population requires new clinical care models as well as the development of novel research tools and infrastructures to address these patients' unique characteristics and healthcare needs. Adult CHD is characterized by substantial complexity, era-dependent heterogeneity in treatment strategies, and time-varying implications of lifelong disease. This burgeoning population is understudied, and the pathophysiology of the component diseases remains incompletely understood. Billing and other administrative codes available in the electronic medical record are neither sensitive nor specific for CHD diagnosis and do not adequately describe many other salient clinical features. As a result, structured data in large administrative databases are not well suited to studying adults with CHD, even when the goal is simply to identify a cohort of patients with a given diagnosis. This constitutes a major impediment to research efforts and is the primary barrier underlying the limited population-based research performed to date. Adult CHD investigation would benefit immensely from methods to establish harmonized, large-scale, multi-center datasets. While billing codes are inadequate, the information needed to accurately classify adults with CHD is already available in the electronic medical record in the form of clinical notes, comprised mainly of unstructured (“free”) text. Manual data extraction is laborious, resource intensive, and, therefore, not scalable. We propose to apply cutting-edge natural language processing approaches to unstructured text in the electronic medical record to develop computable classifiers for variables fundamental to the study of adults with CHD. We will use two unique institutional data resources at Boston Children's Hospital and Brigham and Women's Hospital that are already populated with expert-adjudicated labels to train classifiers for key phenotypes that are poorly defined by administrative codes. These classifiers will be validated in an independent patient cohort at Vanderbilt University Medical Center and tested in new disease-specific risk prediction models. This work promises to accelerate CHD research by massively increasing the scale of the patient cohorts that can be studied and by establishing a foundation for improved evidence-based decision support for this underserved population.