阻塞性睡眠呼吸暂停（OSA）是一种常见的睡眠呼吸障碍疾病，睡眠分期是诊断OSA及其他睡眠相关疾病的关键方法。现阶段，睡眠分期主要由专业的睡眠分期医生进行，费时费力。虽然目前已经有很多研究者开发了基于深度学习的自动睡眠分期算法，但是目前主流的自动睡眠分期算法一般是基于有监督学习的，模型的泛化性能和效果依赖于有标签数据的量以及标签的质量，但是睡眠分期任务使用的PSG信号的标签需要专业的医生给出，标签获取代价高，且受限于医生的水平，不同的医生给出标签的一致性也不能得到保障。此外，PSG信号在采集过程中容易受到干扰，导致数据质量参差不齐。这些问题制约着自动睡眠分期算法的推广与应用。针对上述问题，首先，对于PSG信号的数据质量参差不齐，导致其在训练过程中对模型产生不良影响的问题，本文提出了一种适用于PSG信号的数据质量分级算法，能够通过对数据质量进行评估和分级，将不同质量等级的数据以不同的方式引入到模型训练中，提升模型的泛化性能和效果。此外，对于睡眠分期任务使用的标注数据获取代价高的问题，本文提出了一种适用于PSG信号的无标注数据信息引入算法，通过非对称网络结构和通道dropout来构造正则化样本对，在标注数据有限的前提下通过有效地引入无标注数据的信息来提升模型的性能。然后，结合上述两种算法以及Teacher-Student架构，本文提出了基于半监督学习的自动睡眠分期算法，在标注数据量一致的前提下，在公开数据集上相较于目前在睡眠分期任务上性能SOTA的有监督算法在准确率上有4-7个百分点的提升，相较于目前主流的一些半监督学习的方案也有1-5个百分点的提升，能够在标注数据仅为有监督学习算法十分之一以下的情况下达到相近的性能，具有广泛的应用价值。最后，本文还提出了以自动睡眠分期算法为重要组成部分的OSA辅助诊断流程的相关算法，涵盖OSA的预警、是否患病、疾病类型确定三个流程，对于提升OSA的诊疗效率具有重要意义。
Obstructive Sleep Apnea (OSA) is a common sleep breathing disorder, and sleep staging is a key method for diagnosing OSA and other sleep-related disorders. Currently, sleep staging is mainly performed by professional sleep staging physicians, which is time-consuming and labor-intensive. Although many researchers have developed automatic sleep staging algorithms based on deep learning, the mainstream automatic sleep staging algorithms are generally based on supervised learning, and the model‘s generalization performance and effectiveness depend on the amount and quality of labeled data. However, the labels of PSG signals used in sleep staging tasks need to be given by professional doctors, which is costly and limited by the level of doctors, and the consistency of labels given by different doctors cannot be guaranteed. In addition, PSG signals are susceptible to interference during the collection process, leading to uneven data quality. These problems hinder the promotion and application of automatic sleep staging algorithms.Regarding the above issue, firstly, this thesis proposes a data quality grading algorithm suitable for PSG signals to address the problem that the uneven data quality of PSG signals has an adverse effect on the model during training. By evaluating and grading the data quality, this algorithm can introduce data of different quality levels into the model training in different ways, thereby improving the generalization performance and effectiveness of the model.In addition, for the problem of high cost in obtaining labeled data used in sleep staging tasks, this thesis proposes an unlabeled data information introduction algorithm applicable to PSG signals. By constructing regularization samples through asymmetric network structures and channel dropout, the performance of the model can be improved by effectively introducing information from unlabeled data under the premise of limited labeled data.Then, combining the above two algorithms and the Teacher-Student architecture, this thesis proposes an automatic sleep staging algorithm based on semi-supervised learning. Under the premise of consistent amounts of labeled data, compared with the current state-of-the-art supervised algorithms in sleep staging tasks on public datasets, the proposed algorithm achieves a 4%-7% improvement in accuracy, and a 1%-5% improvement compared to some mainstream semi-supervised learning schemes. It can achieve similar performance with only one-tenth of the labeled data of supervised learning algorithms and has broad application value.At last, this thesis proposes an OSA-assisted diagnostic process and its related algorithms, with an automated sleep staging algorithm as a crucial component. This process includes three steps: OSA warning, disease status determination, and disease type identification, which are of significant importance in improving the efficiency of OSA diagnosis and treatment.
