ECG-AI Based Prediction and Phenotyping of Heart Failure with Preserved Ejection Fraction

项目摘要/摘要 在美国，有600多万成年人患有心力衰竭。心力衰竭与 高死亡率的同时也降低了生活质量。及早认识到心力衰竭并及时 干预措施有助于减轻个人和整个医疗系统的疾病负担。然而， 超过一半的心力衰竭患者左心室射血分数(HFpEF)保持不变，而大多数患者 现有的心衰治疗方法主要针对左心室射血分数(HFrEF)降低的心衰患者。这是因为 HFpEF是一种异质性综合征，其病因尚不清楚。一项由美国国立卫生研究院资助的新倡议， HeartShare研究旨在填补这一知识空白，以确定潜在的不同的HFpEF亚型 使用深度表型、多组学和机器学习方法的治疗选择。然而，仍然有 需要低成本和可获得的工具1)对大量患者人群进行HFpEF风险筛查以支持 预防性风险调整策略和2)确定HFpEF亚型以辅助靶向治疗。 这项辅助研究的目标是利用低成本和可访问的心电数据通过人工 用于预测发生HFpEF风险和流行HFpEF亚型的智能(AI)。 我们和其他人已经证明，将人工智能应用于心电数据可以区分减少和保留的患者 EF具有高精度[1-5]。我们最近开发并验证了一项基于心电的10年心力衰竭风险预测 使用人工智能(AI)的模型[6，7]。这些发现使我们假设人工智能应用于心电数据 可以预测HFpEF风险并识别特定的HFpEF亚型。这项辅助研究的目的是测试我们的 利用回顾的心电图和临床数据进行假说：a)由美国国立卫生研究院资助的黄金标准研究 确定HFpEF和b)来自三个大型医疗系统的真实世界的心电和临床数据(WFU- 维克森林大学，北卡罗来纳州温斯顿-塞勒姆；德克萨斯州-田纳西大学健康科学中心，孟菲斯， TN；和芝加哥Luc-Loyola大学)和c)心脏共享研究的数据。基于我们的专业知识，我们 建议开展基于心电图的HFpEF亚型风险预测和分类，完成三项工作 目标： 目标1.使用NIH资助的研究数据开发事件HFpEF预测模型：我们将利用 来自美国国立卫生研究院资助的研究的高质量和准确的数据，以开发预测HFpEF事件风险的人工智能模型。 目标2.使用真实世界的电子健康记录(EHR)开发事件HFpEF预测模型- 派生数据：我们将首先利用非常大和多样化的基于EHR的真实世界数据来开发事件 HFpEF风险预测模型。然后，我们将通过迁移学习将其与NIH基于数据的模型进行协调。 目的3.开发、测试和实现基于心电信号的HFpEF表型。这一目标将利用来自 对流行的HFpEF患者进行HFpEF亚型划分。

Project Summary/Abstract More than 6 million adults are suffering from heart failure in the United States. Heart failure is associated with high mortality rate while also reducing the quality of life. Early recognition of heart failure and timely interventions can help reducing the disease burden to individuals and to overall healthcare system. However, more than half of HF patients are HF with preserved left ventricular ejection fraction (HFpEF) while the majority of existing HF treatments are for HF with reduced left ventricular ejection fraction (HFrEF). This is because HFpEF is a heterogenous syndrome, and its etiology is not well understood. A new NIH-funded initiative, HeartShare Study, aims to fill this knowledge gap to identify subtypes of HFpEF potentially with different treatment options using deep phenotyping, multi-omics, and machine learning approach. However, there is still a need for low cost and accessible tools 1) for screening large patient populations for HFpEF risk to support preventive risk modification strategies and 2) for identifying HFpEF subtypes to assist targeted therapeutics. The goal of this ancillary study is to utilize low cost and accessible electrocardiogram (ECG) data via artificial intelligence (AI) for prediction of incident HFpEF risk and subtyping of prevalent HFpEF. We and others have shown that AI applied to ECG data can discriminate patients with reduced and preserved EF with high accuracy [1-5]. We recently developed and validated an ECG-based 10-year HF risk prediction model using artificial intelligence (AI) [6, 7]. These findings led us to hypothesize that AI applied to ECG data can predict HFpEF risk and identify specific HFpEF subtypes. The goal of this ancillary study is to test our hypothesis by leveraging retrospective ECG and clinical data from: a) NIH-funded studies with gold standard ascertainment of HFpEFand b) real-world ECG and clinical data from three large healthcare systems (WFU- Wake Forest University, Winston-Salem, NC; UT-University of Tennessee Health Science Center, Memphis, TN; and LUC-Loyola University Chicago) and c) data from the HeartShare Study. Building on our expertise, we propose developing ECG-based risk prediction and classification of HFpEF subtypes by completing three Aims: Aim 1. Develop an incident HFpEF prediction model using data from NIH-funded studies: We will utilize high quality and accurate data from NIH-funded studies to develop AI model predicting risk for incident HFpEF. Aim 2. Develop an incident HFpEF prediction model using real-world Electronic Health Records (EHR)- derived data: We will first utilize very larger and diverse EHR-based real world data to develop incident HFpEF risk prediction model. We will then harmonize it with the NIH-data based model via transfer learning. Aim 3. Develop, test and implement ECG-based HFpEF phenotyping. This aim will utilize data from prevalent HFpEF patients to classify HFpEF subtypes.