基于自编码器和隐马尔可夫模型的睡眠呼吸暂停检测方法

阻塞性睡眠呼吸暂停容易引发心血管并发症。作为睡眠呼吸暂停诊断的金标准,多导睡眠仪的检测费用昂贵且影响患者睡眠质量。鉴于心肺高度耦合,心电信号已被广泛应用于睡眠呼吸暂停检测中。然而,大多数基于心电信号的研究专注于人工特征的设计,依赖于专家先验知识。基于深度学习的方法能够减少特征提取过程中的人为因素。提出一种基于自编码器和隐马尔可夫模型的睡眠呼吸暂停检测方法。首先,利用栈式稀疏自编码器,直接从RR间期序列中进行半监督特征学习,先在预训练阶段进行无监督学习,随后在微调阶段引入标签进行有监督学习。然后,构建支持向量机和人工神经网络,分别结合隐马尔可夫模型之后,组成决策融合分类器,隐马尔可夫模型引入片段之间的时间依赖性,决策融合可整合不同分类器之间的优势。基于Physio Net的apnea-ECG数据库70例整夜睡眠数据,实验结果显示:阻塞性睡眠呼吸暂停片段识别准确率、敏感性和特异性分别为84.7%、88.9%和82.1%,个体识别准确率达到100%。基于自编码器的特征提取方法相较于特征工程,能够降低先验知识限制,使特征提取过程更加自动化、智能化。此外,决策融合分类器相较于单一分类器,不仅可提升片段识别准确率,而且能缓解识别结果中敏感性和特异性之间的不平衡性。

Sleep Apnea Detection Based on Auto-encoder and Hidden Markov Model

Obstructive sleep apnea (OSA) is prone to cardiovascular complications. As a gold standard for the diagnosis of sleep apnea,polysomnography is expensive and affects the sleep quality of patients. Because of the high coupling between heart and lung,electrocardiogram (ECG) signals are widely used in sleep apnea detection. However,most of the studies based on ECG signals focus on the design of artificial features,relying on the prior knowledge of experts. Methods based on deep learning can reduce human factors during feature extraction. In this study,we proposed a sleep apnea detection method based on auto-encoder and hidden Markov model (HMM). Firstly,a stacked sparse auto-encoder was used to perform semi-supervised feature learning directly from the RR interval sequence. Unsupervised learning was performed during the pre-training phase,and labels were then introduced for supervised learning during the fine-tuning phase. Then,a decision fusion classifier based on support vector machine (SVM) and artificial neural network (ANN) combined with HMM was constructed. HMM introduced the temporal dependence between segments. Decision fusion integrated the advantages between different classifiers. Experimental results based on the sleep data of 70 cases of all-night in PhysioNet′s apnea-ECG database showed that the accuracy,sensitivity and specificity of per-segment OSA detection was 84.7%,88.9% and 82.1% respectively,and per-subject detection accuracy was 100%. Compared with feature engineering,the feature extraction method based on auto-encoder could reduce the limitation of prior knowledge and make the feature extraction process more automatic and intelligent. In addition,compared with the single classifier,the decision fusion classifier not only improved the accuracy of per-segment OSA detection,but also alleviated the imbalance between sensitivity and specificity in detection results.