智能制造技术与系统研究中心知识库

随着生活节奏的加快，越来越多的人都承受着不同程度的睡眠问题，改善睡眠质量的前提是及时发现睡眠疾病，睡眠呼吸事件监测是发现睡眠疾病的重要手段。目前临床上使用的睡眠呼吸事件监测的“金标准”为多导睡眠监测，但是多导睡眠监测是一种接触式监测方法，其佩戴需要专业的知识，价格昂贵且不易推广。随着互联网与智能医疗的发展，移动医疗是未来趋势，基于鼾声的睡眠呼吸事件监测方法具有数据易获取、非接触且易推广等优点，具有重要的研究意义和实用价值。

本文的研究目的是设计一个从鼾声检测到鼾声分析的睡眠呼吸事件监测框架，并对其中的每个任务开展了深入的研究，设计了相应的模型，本文的主要内容与创新点如下：

1. 提出了两种基于深度学习的鼾声检测方法。
基于传统机器学习的多阈值鼾声检测方法在实际应用中往往由于特征表达能力不强而检测性能不佳。本文舍弃了传统的检测方法，使用基于深度学习的方法研究鼾声检测，探讨了在不使用明确的鼾声定义前提下进行鼾声检测的方法。相比于先给出一个基于经验的鼾声定义，然后进行传统的鼾声检测，基于深度学习的鼾声检测算法，可以巧妙地避开人工干预的过程通过自主学习鼾声特征能取得更好的检测性能。具体地，本文首先提出了一种基于一维卷积网络的深度学习模型，该模型以音频数据作为输入，从多个特征层进行鼾声检测，最终输出检测结果。此外，在该模型的基础上，本文提出了一个模仿人耳听觉特性的特征增强模块，该模块通过使用多尺度卷积来模仿人耳对于音频频率的捕捉特性。通过在鼾声检测数据集上的实验，验证了本文所提方法的先进性。

2. 提出了一种基于信号自适应分解的鼾声分类方法。
睡眠呼吸事件检测面临着呼吸事件定义复杂、判别规则众多、模式难以识别等诸多挑战，本文摒弃了使用口鼻气流、胸腹带等信号来直接检测呼吸事件的方法，深入研究了鼾声与呼吸事件之间的联系。我们发现在发生呼吸事件时一般都伴随着特定的鼾声出现，不同呼吸事件相关的鼾声具有一定的事件辨识性。基于此，本文提出了一种基于差分算子的信号分解方法来提取能够反应上气道结构信息的鼾声频谱趋势，然后对该趋势进行倒谱分析得到一种新的鼾声声学特征。基于此声学特征，本文提出了一种鼾声分类方法。实验结果表明了该方法的有效性和优越性。
3. 提出了两种基于多尺度统计特征的鼾声分类方法。
阻塞性睡眠呼吸暂停低通气综合征的手术治疗方案成功与否的一个重要因素是上气道中阻塞部位的定位。准确地定位是制定手术方案的前提条件。考虑到鼾声产生的物理过程，如果把上气道看作一个声音产生系统，不同的阻塞部位等价于该系统处在不同的状态，则不同状态下产生的鼾声应该有着不同的声学特性和系统复杂度。基于以上发现，本文分别在时域与频域上研究了基于阻塞部位的鼾声分类方法。首先，在频域上，本文利用基于鼾声频谱趋势的特征提出了一种阻塞部位相关的鼾声分类方法。其次，在时域上，本文创新性地将基于多尺度熵的生理信号复杂度分析方法引入到鼾声分析领域来分析鼾声的复杂度特征，并结合声学特征梅尔频率倒谱系数，提出了一种基于鼾声声学特征和复杂度特征的鼾声分类方法。本文提出的方法利用了鼾声产生过程中的物理及生理学特点，在阻塞部位相关的鼾声分类公开数据集中取得了比现有方法更好的分类效果。

Multifarious pressures in modern life give rise to varying degrees of sleep problems. To improve sleep quality, the premise is timely detection of a sleep disease, by the means of monitoring sleep respiratory events. In current clinical practice, the gold standard for monitoring sleep respiratory events is polysomnography (PSG). However, PSG is contact, expensive and difficult to be popularized since it requires professional instruction to wear the PSG. With the rapid development of the Internet and intelligent medical care, mobile medical care is the future trend. The snore-based monitoring of sleep respiratory event has the advantages of easy to collect data, non-contact and easy to be promoted, which endow it with great research significance and practical value. The goal of this dissertation is to design a framework for monitoring sleep respiratory events, from snore detection to snore analysis tasks. In this framework, we have conducted detailed studies and designed a specific model for each task.

The main contributions of this dissertation are summarized as follows:
1.This dissertation proposes two deep learning based methods of snore detection. Traditional threshold-based methods of snore detection tend to perform barely satisfactory due to their weak ability to express features in practical applications. This dissertation distinguishes itself from the existing framework of traditional detection methods by presenting a deep learning-based snore detection model and studying snore detection without using a definition of snore event. Compared with those snore detection methods which require a definition of the snore at first, the deep learning-based methods can effectively avoid manual intervention during the process and achieve higher detection performance. Specifically, this dissertation first proposes a deep learning model based on a one-dimensional convolutional network, which takes audio recording as input and detects snores from multiple feature layers, and finally outputs snore detection results. In addition to this model, a feature enhancement module is developed, to simulate the
auditory characteristics of the human ear when capturing audio frequencies, by using a multi-scale convolution operation. Experimental results on the snore detection dataset demonstrate the superiority of the proposed methods.

2. This dissertation proposes a snore classification method based on adaptive signal decomposition. The detection of sleep respiratory events is faced with many challenges such as complex definitions, numerous discriminant rules, and difficult recognition of patterns. This dissertation abandons the method of directly detecting respiratory events on the basis of airflow and chest-abdominal respiration signals, and explores the connection between snores and respiratory events in a different way. It is found that a respiratory event is usually accompanied by some specific snores. The event-related snores enable us to identify these two kinds of events. Based on the above, this dissertation proposes a signal decomposition method relying on a difference operator to extract
the snore spectrum trend which can reflect the upper airway structure information. And then we perform cepstrum analysis on the trend to obtain a new snore acoustic feature. According to the feature, this dissertation proposes a snore classification method. Experimental results show the effectiveness and superiority of the method.

3. This dissertation proposes two snore classification methods based on the multiscale statistical feature. One of the cardinal factors that determine the success of surgical treatment for obstructive sleep apnea hypopnea syndrome is the identification of obstruction sites in the upper airway. In other words, the accurate localization of obstruction sites is the precondition for designing an operation scheme. Considering the
physical process of snore production, if we regard the upper airway as a sound production system, different obstructive sites are equivalent to different states of the system. Thus, the snores generated under different states should have different systematic complexity. Based on the above findings, comprehensive and thorough researches are conducted on snore classification approaches relating to obstructive sites in the time
domain and frequency domain respectively. Firstly, in the frequency domain, this dissertation proposes an obstructive-site-related snore classification method based on the snore spectrum trend features. Secondly, in the time domain, this dissertation studies the complexity characteristics of snoring by innovatively introducing the complexity analysis method of physiological signals based on multi-scale entropy into the field of snore analysis. Combined with the Mel-frequency cepstrum coefficient, we propose a snore classification method based on acoustic characteristics and complexity characteristics. The proposed method takes advantage of the physical and physiological characteristics of the snore production process and obtains better results on the public snore classification dataset related to obstructive site than the existing methods.