Elwali Ahmed, Moussavi Zahra
Biomedical Engineering Program, University of Manitoba, Winnipeg, MB, Canada.
Ann Biomed Eng. 2017 Mar;45(3):839-850. doi: 10.1007/s10439-016-1720-5. Epub 2016 Sep 6.
Screening for obstructive sleep apnea (OSA) disorder during wakefulness is challenging. In this paper, we present a set of tracheal breathing sounds characteristics with classification power for separating individuals with apnea/hypopnea index (AHI) ≥ 10 (OSA group) from those with AHI ≤ 5 (non-OSA group) during wakefulness. Tracheal breathing sound signals were recorded during wakefulness in supine position; subjects were instructed to have a few deep breaths through their nose, then through their mouth. Study participants were 147 individuals (80 males) referred to overnight polysomnography (PSG) assessment; their AHI scores were collected after their overnight-PSG study was completed. The signals were normalized; then, their power spectra were estimated. After conducting a multi-stage process for feature extraction and selection on a subset of training data, two spectral features showing significant differences between the two groups were selected for classification. These features showed a correlation of 0.42 with AHI. A 2-class support vector machine classifier with a linear kernel was used. Following this an exhaustive leave-two-out cross-validation was performed. The overall accuracies were 83.83 and 83.92% for training and testing datasets, respectively, while the overall sensitivity and specificity of the test datasets were 82.61 and 85.22%, respectively. We also applied the same method for anthropometric information (i.e., age, weight, etc.) as features, and they resulted in an overall accuracy of 77.6 and 76.2% for training and testing datasets, respectively. The results of this study show a superior classification power of respiratory sound features compared to anthropometric features for a quick screening of OSA during wakefulness. The relationship of the sound features and known morphological upper airway structure of OSA subjects are also discussed.
在清醒状态下筛查阻塞性睡眠呼吸暂停（OSA）障碍具有挑战性。在本文中，我们提出了一组具有分类能力的气管呼吸音特征，用于在清醒状态下将呼吸暂停/低通气指数（AHI）≥10的个体（OSA组）与AHI≤5的个体（非OSA组）区分开来。在清醒状态下，受试者仰卧位时记录气管呼吸音信号；指示受试者先通过鼻子进行几次深呼吸，然后通过嘴巴进行深呼吸。研究参与者为147名个体（80名男性），他们被转诊进行夜间多导睡眠图（PSG）评估；在他们的夜间PSG研究完成后收集他们的AHI分数。对信号进行归一化处理；然后，估计其功率谱。在对训练数据的一个子集进行特征提取和选择的多阶段过程后，选择了两组之间显示出显著差异的两个频谱特征进行分类。这些特征与AHI的相关性为0.42。使用了具有线性核的二类支持向量机分类器。在此之后，进行了详尽的留二法交叉验证。训练数据集和测试数据集的总体准确率分别为83.83%和83.92%，而测试数据集的总体敏感性和特异性分别为82.61%和85.22%。我们还将相同的方法应用于人体测量信息（即年龄、体重等）作为特征，训练数据集和测试数据集的总体准确率分别为77.6%和76.2%。这项研究的结果表明，与人体测量特征相比，呼吸音特征在清醒状态下快速筛查OSA方面具有更高的分类能力。还讨论了声音特征与OSA受试者已知的形态学上气道结构之间的关系。
