Explainable Deep Learning Approach for Automatic Arousal and Sleep Stages Scoring, and Knowledge Discovery

用于自动唤醒和睡眠阶段评分以及知识发现的可解释深度学习方法

• 批准号: 10291585
• 负责人: Ao Li
• 金额: $ 11.51万
• 依托单位: UNIVERSITY OF ARIZONA
• 依托单位国家: 美国
• 财政年份: 2021
• 资助国家: 美国
• 起止时间: 2021-09-20 至 2023-08-31
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/10291585
• 关键词: Address; Algorithms; Arousal; Artificial Intelligence; Attention; Blood Pressure; Cardiac health; Clinic; Clinical; Coronary Arteriosclerosis; Data; Databases; Dependence; Diabetes Mellitus; Diagnosis; Diagnostic; Engineering; Evaluation; Event; Frequencies; Goals; Heart failure; Home; Home environment; Human; Hypertension; Knowledge; Knowledge Discovery; Lead; Length; Machine Learning; Manuals; Mediating; Medical Care Team; Medicine; Modeling; Monitor; Multi-Ethnic Study of Atherosclerosis; Outcome; Pattern; Performance; Physicians; Physiological; Polysomnography; Prevalence; Psychological Transfer; REM Sleep; Research; Risk Factors; Scientist; Severities; Signal Transduction; Sleep; Sleep Apnea Syndromes; Sleep Disorders; Sleep Fragmentations; Sleep Stages; Source; Stroke; System; Techniques; Technology; Testing; Training; Trust; Woman; automated algorithm; base; convolutional neural network; deep learning; deep learning algorithm; design; heart rate variability; improved; interdisciplinary approach; learning network; men; multi-task learning; multitask; neural patterning; non rapid eye movement; portability; recurrent neural network; relating to nervous system; respiratory; sleep quality; usability

Project Summary Sleep-disordered breathing (SDB) is potential remedial risk factor for hypertension, diabetes, stroke, coronary artery disease, and heart failure. The prevalence of SDB is estimated to be between 6.5% and 9% in women and between 17% and 31% in men. During polysomnography, which is often required for diagnosis, sleep stages and the frequency of cortical arousals are important metrics. A high frequency of arousals is indicative of sleep fragmentation. Additionally, cortical arousal events are also used to identify hypopneic events in sleep scoring. Currently, type III portable sleep monitors are commonly used for diagnosing SDB severity instead of more expensive polysomnography. However, most portable home sleep test (HST) monitors do not record electroencephalographic (EEG) data which are required for arousal identification, resulting in an underestimation of SDB severity in manual scoring of SDB events. Thus, there is a critical need to improve portable HST sleep monitors with advanced automatic scoring algorithms that can identify arousals associated with SDB events. Studies have found that cortical arousal is associated with sympathetic neural surges observed on electrocardiographic (ECG) and blood pressure signals. Additionally, changes in respiratory patterns, which can be observed from the ECG signal, have been found to be associated with specific EEG patterns. Furthermore, different autonomic neural patterns dominate in non-rapid eye movement (NREM) and rapid eye movement (REM) sleep. The RR interval and respiratory- mediated HF components of heart rate variability (HRV) increase from stages N1 to N3. Our hypothesis is that ECG signals can be used to automatically scoring sleep stages and arousals in HST. In this study, we plan to develop a deep learning-based multi-task learning algorithm for automatic arousal and sleep stage scoring. Instead of HRV based algorithms, we propose to employ an end-to-end deep learning network to acquire features from the raw ECG data. The proposed model consists of convolutional neural networks, recurrent neural networks, and an attention mechanism. It can: (1) accept varying length ECG data; (2) capture long-range dependencies in the ECG data; and (3) share knowledge among scoring tasks for arousal and sleep stages. We use HRVs to further analyze the ECG regions selected by the deep learning model. This is a critical step to understand the underpinnings of associations between sleep events and the ECG signal discovered by the proposed model. Our specific aims include: (1) developing an end-to-end multitask deep learning model for automatic arousal and sleep stages scoring by analyzing a modified lead II ECG signal which is commonly used in sleep studies; (2) advanced interpretation of deep learning model outcomes. Our current effort will evaluate the usability of deep learning approach in sleep medicine and will have a substantive and sustained impact on diagnosis outcomes for sleep disorders.

项目摘要 睡眠呼吸障碍（SDB）是高血压、糖尿病、脑卒中， 冠状动脉疾病和心力衰竭。SDB的患病率估计在6.5%至9%之间 在女性和男性中的比例分别为17%和31%。在多导睡眠图中， 诊断、睡眠阶段和皮层觉醒的频率是重要的度量。高频率的 觉醒是睡眠片段的指示。此外，皮层唤醒事件也用于识别 睡眠呼吸不足事件评分。目前，III型便携式睡眠监测器通常用于 诊断SDB的严重程度，而不是更昂贵的多导睡眠图。然而，大多数便携式家庭睡眠 测试（HST）监视器不记录唤醒所需的脑电图（EEG）数据 这导致在SDB事件的手动评分中低估了SDB的严重程度。因此，在本发明中， 迫切需要改进具有高级自动评分算法的便携式HST睡眠监测器 可以识别与SDB事件相关的觉醒。研究发现大脑皮层的唤醒 在心电图（ECG）和血压信号上观察到交感神经波动。 另外，已经发现了可以从ECG信号观察到的呼吸模式的变化 与特定的脑电图模式有关此外，不同的自主神经模式占主导地位， 非快速眼动（NREM）睡眠和快速眼动（REM）睡眠。RR间期和呼吸- 心率变异性（HRV）的介导的HF成分从N1阶段增加到N3阶段。我们的假设是 ECG信号可用于HST中的睡眠阶段和觉醒的自动评分。本研究 计划为自动唤醒和睡眠阶段开发基于深度学习的多任务学习算法 得分。我们建议采用端到端深度学习网络，而不是基于HRV的算法， 从原始ECG数据中获取特征。所提出的模型由卷积神经网络组成， 递归神经网络和注意力机制。它可以：（1）接受不同长度的心电数据;（2） 捕获ECG数据中的长程依赖性;以及（3）在唤醒的评分任务之间共享知识 睡眠阶段。我们使用HRV来进一步分析由深度学习模型选择的ECG区域。 这是了解睡眠事件和心电图之间联系的基础的关键一步 该模型发现的信号。 我们的具体目标包括：（1）开发端到端的多任务深度学习模型， 通过分析修改的导联II ECG信号自动唤醒和睡眠阶段评分， 用于睡眠研究;（2）深度学习模型结果的高级解释。我们目前的努力将 评估深度学习方法在睡眠医学中的可用性，并将有实质性和持续的 对睡眠障碍诊断结果的影响。