Applying Innovative Artificial Intelligence Approaches to a Large Sleep Physiologic Biorepository to Integrate Sleep Disruption in Cardiovascular Risk Calculation

项目概要： 心血管疾病（CVD）每年造成> 800，000人死亡，即，美国32%的死亡病例， 预计到2035年，总成本将达到2.5万亿美元。实验和流行病学数据确定睡眠障碍- - 最近在美国心脏协会生命的基本8-作为独立的预防目标， 减轻下游主要不良心血管事件（MACE）。阻塞性睡眠呼吸暂停综合征（OSA）是一种 通过间歇性缺氧和交感神经通路最常涉及CV风险的疾病 神经系统激活然而，我们小组和其他人的新兴科学已经确定， 睡眠中断的各个方面，如睡眠减少和睡眠结构中断，也会增加CV风险。 增强的表型不仅是OSA--超越了标准使用的呼吸暂停低通气指数的限制 (AHI)- 而且其他睡眠障碍也可以改善表征睡眠相关病理生理学的能力， MACE预测。然而，导致CV风险的重叠睡眠表型难以表征， 考虑到对大型数据集的需求。此外，睡眠障碍的“困倦”表型与以下因素有关： CV风险增加;然而，对于如何将其整合到CV风险预测中的理解有限。 因此，我们建议利用现有的多模式心肺和神经系统的临床登记， 生理睡眠数据，即> 186，000份存档的睡眠研究。工作范围包括进行分析 来自多导睡眠图（PSG）数据集的CVD生物学上合理的聚合生物标志物， 利用人工智能模型从结构化数据和原始PSG信号数据中识别模式， MACE（非致死性心肌梗死、致死性冠心病、非致死性或致死性卒中）的发生率，以及 研究嗜睡表型的影响。我们将进一步研究将自动 当前临床CV风险分层方案中的PSG分析。这项工作将为外部 在其他临床队列和NHLBI国家睡眠研究资源中进行的验证工作， 地理上不同的45,000多项睡眠研究的汇编。这项工作提供了一个创新的 有机会评估睡眠研究的能力，即，PSG生物标志物，以预测个体的风险增加， CVD使用我们小组建立的方法。创新还在于使用最先进的深度学习 策略，包括用于PSG直接生理学的低维表示的Transformers模型 信号.鉴于我们的专业知识和经验， 我们在睡眠医学、心血管和计算机科学的研究中。

PROJECT SUMMARY: Cardiovascular disease (CVD) accounts for >800,000 deaths annually, i.e., 32% of all deaths in the US, with total costs projected to reach $2.5 trillion by 2035. Experimental and epidemiologic data identify sleep disorders- -recently recognized in American Heart Association Life’s Essential 8--as independent preventative targets to mitigate downstream major adverse cardiovascular events (MACE). Obstructive sleep apnea (OSA) is the sleep disorder most consistently implicated in CV risk operating via pathways of intermittent hypoxia and sympathetic nervous system activation. Emerging science, however, from our group and others, has identified that other facets of sleep disruption, such as curtailed sleep and sleep architectural disruption, also increase CV risk. Enhanced phenotyping of not only OSA--beyond the limitations of the standardly used apnea-hypopnea index (AHI) --- but also other sleep disorders could refine the ability to characterize sleep-related pathophysiology and MACE prediction. However, overlapping sleep phenotypes contributing to CV risk are difficult to characterize, given the need for large datasets. Moreover, the “sleepy” phenotype of sleep disorders is associated with increased CV risk; however, there is limited understanding of how to integrate this into CV risk prediction. Therefore, we propose leveraging an existing clinical registry of multimodal cardiorespiratory and neurologic physiologic sleep data, i.e.,>186,000 archived sleep studies. The scope of work involves conducting an analysis of biologically plausible aggregate biomarkers of CVD from datasets of polysomnograms (PSG) that combine with artificial intelligence models to identify patterns from structured data and raw PSG signal data to forecast the incidence of MACE (nonfatal myocardial infarction, fatal coronary heart disease, nonfatal, or fatal stroke) and examine the influence of the sleepy phenotype. We will further examine the utility of incorporating automatic PSG analysis in the current clinical CV risk stratification schema. This work will set the stage for external validation work in other clinical cohorts and the NHLBI National Sleep Research Resource, a pooled geographically diverse compilation of >45,000 sleep studies. The proposed work provides an innovative opportunity to assess the ability of sleep study, i.e., PSG biomarkers, to predict individuals at increased risk for CVD using methods established by our group. Innovation also lies in the use of state-of-the-art deep learning strategies, including Transformers models for low-dimensional representation of PSG direct physiological signals. Our group is well-positioned to undertake the following study aims, given the expertise and experience we have in sleep medicine, cardiovascular, and computer science research.