打鼾的无监督聚类分析识别阻塞性睡眠呼吸暂停严重程度的声学生物标志物。

文献信息

中文摘要

背景： 阻塞性睡眠呼吸暂停（OSA）是一种常见的公共卫生问题，与心血管、代谢和神经认知后果相关。尽管多导睡眠监测（PSG）是诊断金标准，但其可及性可能有限且等待时间长。打鼾是OSA的常见症状，但并非所有打鼾者都患有OSA。在本研究中，我们假设打鼾的特征可用于检测OSA。我们应用无监督聚类方法对打鼾模式进行分类，并检查其与OSA参数的相关性。
方法： 我们招募了287名有打鼾症状的成年人，他们在台湾2家教学医院接受了整夜PSG。记录打鼾声音，并分析主频和响度。首先使用训练队列（n = 50）开发K-means聚类模型，并将其应用于将频率模式分为5个聚类。根据响度阈值（70 dB），每个聚类进一步分为2个子聚类。使用独立验证队列（n = 237）开发呼吸暂停低通气指数（AHI）预测模型，并使用受试者工作特征分析评估判别性能。
结果： 聚类1、2和5与较高的AHI、氧减饱和度指数和呼吸暂停指数显著相关。相反，聚类3和4显示很少或没有关联。使用年龄、BMI和显著声学聚类开发了AHI预测模型。对于PSG定义的OSA阈值AHI ≥ 5、≥15和≥30次/小时，AUC值分别为0.90、0.83和0.82，敏感性范围为70.5%至77.5%，特异性范围为73.7%至100%。
结论： 我们的研究表明，基于AI的打鼾模式聚类分析可用于评估OSA严重程度。打鼾模式分析也可作为OSA的额外筛查工具。

英文摘要

BACKGROUND: Obstructive sleep apnea (OSA) is a prevalent public health concern linked to cardiovascular, metabolic, and neurocognitive consequences. Although polysomnography (PSG) is the diagnostic standard, the availability can be limited with long wait time. Snoring is a common symptom of OSA, but all individuals with snoring do not have OSA. In this study, we hypothesized the characteristics of snoring may be used to detect OSA. We applied an unsupervised clustering method to classify snoring patterns and examine their correlations with OSA parameters.
METHODS: We enrolled 287 adults with snoring symptoms who underwent overnight PSG at 2 teaching hospitals in Taiwan. Snoring sounds were recorded, and the dominant frequency and loudness were analyzed. A K-means clustering model was first developed using a training cohort (n = 50) and applied to classify frequency patterns into 5 clusters. Each cluster was further divided into 2 subclusters based on a loudness threshold (70 dB). An independent validation cohort (n = 237) was used to develop an apnea-hypopnea index (AHI) prediction model and evaluate discriminative performance using receiver operating characteristic analysis.
RESULTS: Clusters 1, 2, and 5 were significantly associated with higher AHI, oxygen desaturation index, and apnea index. In contrast, clusters 3 and 4 showed little or no association. An AHI prediction model was developed using age, BMI, and significant acoustic clusters. For PSG-defined OSA thresholds of AHI ≥ 5, ≥15, and ≥30 events/h, the AUC values were 0.90, 0.83, and 0.82, respectively, with sensitivity ranging from 70.5% to 77.5% and specificity ranging from 73.7% to 100%.
CONCLUSIONS: Our study demonstrates an AI-based clustering analysis of snoring patterns can be used to evaluate OSA severity. The snoring pattern analysis can also be an additional screening tool for OSA.