Deep learning of awake and sleep electrocardiography to identify atrial fibrillation risk in sleep apnea

项目摘要 心房颤动（AF）是最常见的心律失常，导致显著的发病率和死亡率 负担阻塞性睡眠呼吸暂停（OSA）是一种常见的睡眠障碍，但在 阻塞性睡眠呼吸暂停被认为是房颤的一个危险因素。 由于许多共同的风险因素，如肥胖，OSA和AF一直具有挑战性。没有研究 已经证明了关于OSA的信息是否可以改善对未来AF风险的预测。特别是， 确定“OSA患者”中哪些人有房颤风险尚不清楚。更好地识别群体 OSA患者中最容易发生AF的患者将告知临床医生和患者 治疗决策所需的信息。OSA评估的一个主要挑战是， 在评估中使用的常规度量，例如呼吸暂停低通气指数（AHI）不足以 捕获下游心血管（CV）反应。我们和其他人已经确定了有希望的生理- 驱动的多导睡眠图（PSG）标记物，可更好地捕获OSA的严重程度并改善CV风险 分层我们的初步研究表明，阻塞性睡眠呼吸暂停（OSA）的心率反应（HRR） 事件，而不是AHI，与社区居住的老年男性AF事件相关。心电图 (ECG)是一种容易获得的诊断工具，可以捕获心脏的电活动。深度学习（DL） 在各种临床结局（包括“清醒”时的AF）的检测和风险预测方面显示出巨大的前景 只有心电图。“睡眠”ECG受睡眠状态、呼吸，特别是病理性呼吸的影响 例如OSA事件。基于此，我们提出了目标1：评估新的基于HRR的OSA度量是否 改善了AF的风险预测，超出了当前的AF风险预测模型。我们将使用一个组合 社区动脉粥样硬化风险研究（ARIC）-睡眠心脏健康研究（SHHS）的前瞻性队列， 心血管健康研究（CHS）-SHHS和动脉粥样硬化多种族研究（梅萨）（N~5000，AF 事件~800）。目的2：使用清醒ECG（10秒12导联）和睡眠ECG开发和测试DL模型 （单导联）预测“OSA”一般人群新发AF。我们将开发一个卷积 使用ARIC + CHS队列的神经网络（CNN）模型（合并N与OSA约1500，AF事件约400）和 在梅萨队列中进行外部验证（OSA~1000，AF事件~100）。性能将与 CHARGE-AF风险预测模型。目标3：与目标2相同，不同之处在于它将是预测新的 在临床实践中，基于目标2的CNN模型，我们将开发一个 使用来自单个学术医疗中心（N= 2000，AF~200）的临床ECG数据的单独CNN模型， 在真实的世界临床实践中可能更相关。50%的数据集将用于训练，50%用于 验证。这项研究的结果将提供关键的信息，未来的应用DL在 改善OSA患者的CV风险分层。

Project Summary Atrial fibrillation (AF) is the most common cardiac arrhythmia responsible for significant morbidity and mortality burden. Obstructive sleep apnea (OSA) is a common sleep disorder but disproportionately more common in patients with AF. OSA has been proposed as a risk for AF. However, clarifying the association between the OSA and AF has been challenging due to many commonly shared risk factors such as obesity. No studies have demonstrated whether information about OSA improves prediction of future risk of AF. In particular, identifying who “among those with OSA” would be at risk for AF is unclear. Better identification of the group most vulnerable to developing AF among those with OSA will inform clinicians and patients of critical information needed for therapeutic decision making. One major challenge in OSA evaluation is that conventional metrics used in the evaluation, such as the apnea hypopnea index (AHI) do not adequately capture downstream cardiovascular (CV) responses. We and others have identified promising physiologically- driven polysomnography (PSG) markers that better capture the severity of OSA and improve CV risk stratification. Specifically related to AF, our preliminary study shows that heart rate response (HRR) to OSA events, but not AHI, is associated with incident AF in community dwelling elderly men. Electrocardiography (ECG) is a readily available diagnostic tool that captures electrical activity of the heart. Deep learning (DL) has shown great promise in detection and risk prediction of various clinical outcomes including AF from `awake' ECGs alone. `Sleep' ECG is affected by sleep state, respiration and particularly by pathological respiration such as OSA events. Based on this, we propose Aim 1: To evaluate whether novel HRR-based OSA metrics improves risk prediction of AF beyond the current AF risk prediction model. We will use a combined prospective cohort of Atherosclerosis Risk in Communities Study (ARIC)-Sleep Heart Health Study (SHHS), Cardiovascular Health Study (CHS)-SHHS and Multi-Ethnic Study of Atherosclerosis (MESA) (N~5000, AF events~800). Aim 2: To develop and test the DL model using an awake ECG (10 sec 12 lead) and sleep ECG (single lead) to predict a new onset AF in general population “with OSA”. We will develop a convolutional neural network (CNN) model utilizing ARIC + CHS cohorts (combined N with OSA~1500, AF events ~400) and externally validate in MESA cohort (OSA~1000, AF events ~100). The performance will be compared with the CHARGE-AF risk prediction model. Aim 3: Same as Aim 2 except it will be the DL model in prediction of new onset AF patients with OSA in clinical practice. Building upon the CNN model from Aim 2, we will develop a separate CNN model using clinical ECG data from a single academic medical center (N= 2000, AF~200) that may be more relevant in real world clinical practice. 50% of the dataset will be used for training and 50% for validation. The findings of this study will provide critical information about the future application of DL in improving CV risk stratification of people with OSA.