Development of End-To-End Clinical Decision Support Tools To Prevent Cardiotoxic Drug Response

摘要 药物引起的心脏毒性，以 QT 间期延长和尖端扭转型室性心动过速的形式，是一种不常见但 目前上市的一百多种药物具有毁灭性的副作用。药物引起的 QT 普遍存在 跨医学专业和病症的延长（diLQTS）给寻求以下服务的提供者带来了挑战： 开出已知的延长 QT 间期的药物，尤其是针对非心脏病的药物。我们小组的工作是 开发自动化临床决策支持 (CDS) 工具来提醒提供者患者风险已显示出希望 旨在减少高危人群的处方数量。然而，这些工具依赖于历史 心电图 (ECG) 和 QT 间期延长可识别高危患者，从而排除大量患者 在我们的系统中没有进行心电图检查的潜在高危人群。通过独特的机构 与 Google 合作，我们的整个电子健康记录 (EHR) 的副本存储在 Google 上 云平台（GCP），我们开发了初步的深度学习模型来预测diLQTS的风险。我们 还验证了几个现实世界中 QT 间期和 diLQTS 的遗传关联 使用聚合多基因风险评分的人群。通过创建机构生物库 结果临床应用的认证，以及 EHR 数据与遗传数据的基于云的集成， 我们有能力利用现有的基础设施来研究深度学习和遗传学的作用 降低 diLQTS 的风险。这项研究将结合我们独特的研究和临床 与我们的调查团队一起了解科罗拉多大学安舒茨医学校区的基础设施 由药物基因组学和医学信息学研究专家组成开发和研究 一种端到端 CDS 工具，结合遗传学和深度学习来预测 diLQTS 风险。的 该应用程序的具体目标包括以下内容：(1) 开发和测试基于云的深度学习模型 使用住院和门诊患者的 EHR 数据来预测 diLQTS 风险； (2) 使用以下方法验证 diLQTS 的遗传预测因子 机构生物库样本和多种族外部人群； (3) 使用以下方法开发和测试 CDS 工具 这些先进的方法可以降低 diLQTS 的风险。我们将使用通用数据模型（观察 医疗结果合作伙伴关系）从 EHR 数据映射，以及定制 DNA 阵列（多种族 基因分型阵列）设计用于对各种非欧洲血统进行插补，以确保我们的 本研究的预测模型和结果可以在其他机构和人群中复制 未来。通过这种方式，这项调查不仅可以深入了解机器学习的使用， 遗传学用于 diLQTS 风险预测，但它也将为未来高级 CDS 开发创建蓝图 对于其他条件。

SUMMARY Drug-induced cardiac toxicity, in the form of QT prolongation and torsade de pointes, is an uncommon but devastating side effect of over one hundred currently marketed drugs. The ubiquity of drug-induced QT prolongation (diLQTS) across medical specialties and conditions creates a challenge for providers seeking to prescribe known QT-prolonging medications, particularly for non-cardiac conditions. Work by our group to develop automated clinical decision support (CDS) tools that alert providers of patient risk has shown promise towards reducing the number of prescriptions to at-risk individuals. However, these tools rely on a history of an electrocardiogram (ECG) with QT prolongation to identify at-risk patients, and thus exclude a large number of potentially at-risk individuals who have not had an ECG within our system. Through a unique institutional partnership with Google, in which a copy of our entire electronic health record (EHR) is stored on the Google Cloud Platform (GCP), we have developed preliminary deep-learning models to predict risk of diLQTS. We have also validated the genetic association with the QT interval and diLQTS across several real-world populations using an aggregate polygenic risk score. Through creation of an institutional biobank with certification for clinical application of results, as well as cloud-based integration of EHR data with genetic data, we have the capability to leverage our existing infrastructure to study the role of deep learning and genetics to reduce the risk of diLQTS. This investigation will combine our unique research and clinical infrastructure on the University of Colorado Anschutz Medical Campus with our investigative team composed of experts in the study of pharmacogenomics and medical informatics to develop and study an end-to-end CDS tool incorporating genetics and deep learning to predict risk of diLQTS. The specific aims of this application include the following: (1) develop and test a cloud-based, deep-learning model using EHR data on in- and outpatients to predict risk of diLQTS; (2) validate genetic predictors of diLQTS using institutional biobank samples, and a multi-ethnic external population; and (3) develop and test CDS tools using these advanced methods to reduce the risk of diLQTS. We will use a common data model (Observational Medical Outcomes Partnership) mapped from EHR data, as well as a custom DNA array (Multi-Ethnic Genotyping Array) designed for imputation across a variety of non-European ancestries, to ensure that the our prediction model and findings from this study can be replicated in other institutions and populations in the future. In such a way, this investigation will not only provide insight into the use of machine learning and genetics for risk prediction of diLQTS, but it will also create a blueprint for future advanced CDS development for other conditions.