睡眠微觉醒（Sleep Arousal，Arousal）是睡眠过程中脑电和肌电频率发生突变的一种现象，大量研究表明微觉醒与阻塞性睡眠呼吸暂停低通气综合征（Obstructive Sleep Apnea-Hypopnea Syndrome, OSAHS）有紧密关系。微觉醒检测主要通过多导睡眠监测（Polysomnography，PSG）对受试者采集数据后，经由专业医师判读，应用于OSAHS疾病的辅助筛查。但人工判读微觉醒较复杂，相关的诊疗资源极为缺乏。然而睡眠障碍的普遍存在且患者数量呈上升的趋势，因此监测工作家庭化、筛查设备轻量化、检测算法更加准确高效的工作则显得愈发重要。本课题依托国家科技部的重大研发专项《睡眠呼吸疾病数字化集成设备分级诊疗体系研究》，提出了一种基于PSG信号和深度学习方法的微觉醒检测模型，旨在开发更准确、更高效的微觉醒检测算法，探索更多能用于微觉醒检测的信号通道组合，为未来轻量化设备的微觉醒检测提供创新思路。本文所提出的高精度的微觉醒自动检测模型是一种以残差网络作为基础框架的改进型卷积神经网络。模型主要由增加了自注意力机制的残差块组成的基本特征提取单元、在卷积网络后串接的类循环神经网络结构双向长短期记忆网络（Bi-LSTM，Bi-Directional Long Short-Term Memory）组成。增加自注意力的特征提取单元改变了原有的PSG信号多通道的并行计算规则，使得模型具有更多非线性，更好地拟合通道之间的复杂相关性；而Bi-LSTM可以同时捕捉前后双向编码信息，合理地考虑到时序信号的迟滞信息和前后相关性。此外，本研究充分利用了PSG采集到的各项生理信号，通过通道组合分析来探究除了以呼吸和脉搏相关通道为代表的其它可用通道。基于上述模型结构，本文在合作医院提供的临床数据集以及公开数据集中分别进行验证，性能表现出色：1）在临床数据集中取得90.47%的精确率、91.86%的召回率、99.46%的特异度和99.11%的准确率；2）在公开数据集中取得84.00%的精确率、87.80%的召回率、99.49%的特异度和98.89%的准确率；3）使用胸腔和腹腔呼吸努力、脉搏波和脉搏血氧饱和度4通道在临床数据集中取得83.34%的精确率、86.45%的召回率、99.16%的特异度和98.57%的准确率。本文的算法研究在检测性能上较现有研究处于领先水平，且为使用轻量化可穿戴设备实现微觉醒检测提供创新思路，具有高效、准确、稳定、应用和推广性。
Sleep Arousal (Arousal) is a phenomenon of sudden changes in EEG and EMG frequencies during sleep, which has been shown to be closely related to Obstructive Sleep Apnea-Hypopnea Syndrome (OSAHS) in numerous studies. Arousal detection is mainly performed by polysomnography (PSG) and interpreted by a medical professional as an aid to screening for OSAHS disease. However, manual interpretation of arousal is complex, and there is a lack of resources for diagnosis and treatment. The prevalence of sleep disorders and the increasing number of patients make it more and more important to make the monitoring work at home, lighten the screening equipment, and improve the accuracy of the detection algorithm.Based on the major R&D project of the Ministry of Science and Technology of China, “Research on the Hierarchical Diagnosis and Treatment System of Sleep Respiratory Diseases with Digital Integrated Equipment”, this project proposes a arousal detection model ArousalNet based on PSG signals and deep learning methods, aiming to develop more accurate and efficient arousal detection algorithms, explore more combinations of signal channels for arousal detection, and provide innovative ideas for future use of lightweight devices.The proposed ArousalNet is a modified convolutional neural network with residual network as the base framework. The model mainly consists of a basic feature extraction unit composed of residual blocks based on self-attention mechanism, and Bi-Directional Long Short-Term Memory (Bi-LSTM), a recurrent neural network-like structure cascaded after the convolutional network. Adding the self-attention residual convolution layer changes the original parallel computation rules of multi-channel signals, which makes the model more nonlinear and can better fit the complex correlations between channels; Bi-LSTM can capture the front and back bidirectional coding information at the same time and reasonably take into account the hysteresis information and front and back correlations of temporal signals. In addition, this study makes full use of the physiological signals acquired by PSG, and explores other available signal channels without using EEG and EMG channels, but using respiratory and pulse-related channels as inputs through channel combination analysis. Based on the above model structure and experimental ideas, this thesis was validated on the clinical dataset of Tongren Hospital and the public dataset of SHHS, and the performance was excellent: (1)90.47% accuracy, 91.86% recall, 99.46% specificity and 99.11% accuracy were achieved in the clinical dataset; (2)84.00% accuracy, 87.80% recall, 99.11% accuracy were achieved in the public dataset, 87.80% recall, 99.49% specificity and 98.89% accuracy; (3)83.34% precision, 86.45% recall, 99.16% specificity and 98.57% accuracy in the clinical dataset using 4 channels of thoracic and abdominal respiratory effort, pulse wave and pulse oximetry.The algorithm study in this thesis is at the leading level in terms of detection performance compared to existing studies, and provides innovative ideas for implementing arousal detection using lightweight wearable devices with accuracy, stability, application and scalability.
