基于多尺度熵的呼吸紊乱事件相关鼾声信号的识别研究

The low diagnostic rate of obstructive sleep apnea hypopnea syndrome (OSAHS) is a clinical challenge. By monitoring snoring and respiratory disturbance events in real time, snoring can be categorized into seven categories: former-hypopnea, during-hypopnea, post-hypopnea, former-apnea, during-apnea, post-apnea and common snoring. Its linear time and frequency domain features have been extensively studied; but snoring is essentially a complicated physiological phenomenon, its nonlinear features have not been elucidated yet. Entropy can be used to characterize and quantify nonlinear events. The decision tree model can recognize common snoring, during-hypopnea, during-apnea, and post-apnea snoring, and the accuracy is 81.13%. The accuracy of linear features and sample entropy for identifying snoring and non-snoring is 94.54%, which suggests that entropy can be used to snoring. We found that multi-scale entropy could recognize 83% of common snoring, but lower for recognizing snoring related to respiratory disturbance events. In this study, we plan to further optimize the methods to categorize snoring and develop appropriate algorithms to accurately and automatically distinguish different types of snoring via increasing sample size, combining linear features with multi-scale entropy and using machine learning. The expected results will contribute to accurate diagnosis of OSAHS.

阻塞性睡眠呼吸暂停低通气综合征(OSAHS)低诊断率是临床面临的一大挑战。我们将鼾声和呼吸紊乱事件实时结合，初步将鼾声分为发生在低通气和呼吸暂停前、中、后的鼾声及普通鼾声七类。鼾声线性时域、频域特征已被广泛研究。但鼾声本质上是一种复杂的生理信号，其非线性特征未被重视，熵可以量化复杂性，是非线性分析指标。我们前期利用鼾声线性特征通过决策树模型能区分出普通鼾声及低通气中、呼吸暂停中、呼吸暂停后鼾声，准确率为81.13%；通过线性特征与样本熵组合识别鼾声的准确度达94.54%，显示了熵在鼾声研究中的应用前景。初步研究发现多尺度熵对普通鼾声识别率达83%，但对呼吸事件相关鼾声识别率较低。本研究拟通过进一步加大样本量、分析不同类型鼾声信号的线性特征与多尺度熵、利用机器学习，对鼾声分类进行进一步优化，并探索合适的特征组合及算法用于准确地自动化区分不同类型鼾声，为通过鼾声分析准确诊断OSAHS奠定基础。

阻塞性睡眠呼吸暂停低通气综合征(OSAHS)低诊断率是临床面临的一大挑战。鼾声信号的声学特性作为诊断OSAHS的一种潜在的成本效益高且可靠的替代方法已被广泛研究。本文提出了一种新的降噪方法—基于噪声协方差矩阵估计的广义子空间降噪方法，与常规子空间降噪算法相比，该方法对噪声协方差矩阵的估计更好，并且在不同的非平稳噪声环境下，鼾声信号的客观质量测量和频谱图在噪声降低和信号失真方面都有明显改善。接着，为了识别整晚录音中的鼾声与非鼾声，提出了一种基于声学特征和XGBoost的鼾声检测算法，该方法在计算成本和性能之间取得了很好的平衡，为快速、实时地识别鼾声与非鼾声提供了保障。通过预先训练的VGG19和长短期记忆(LSTM)融合模型，分析患者的鼾声声学特征、识别OSAHS患者并预测呼吸暂停低通气指数（AHI），分类准确率达到99.31%；模型估计的AHI与PSG AHI高度相关，Pearson相关系数为0.966 (p < 0.001)。利用决策树模型对鼾声进行7分类，最终的准确率为91%；决策树模型中F1-F5、PTmean、63的倍频带能量、鼾声时长这8个鼾声声学特征组合在区分7类鼾声中起到的作用最大。接着通过决策树模型用7类鼾声分别预测单纯打鼾患者、轻、中、重度OSAHS患者，结果发现普通鼾声、低通气中鼾声、呼吸暂停前鼾声、呼吸暂停中鼾声、呼吸暂停后鼾声预测准确率均高于95%。总之，本项目提出了一套基于噪声协方差矩阵估计的广义子空间降噪方法改善了鼾声信号的质量；提出了一种计算成本和性能之间取得了很好平衡的基于声学特征和XGBoost的鼾声检测算法；利用决策树模型发现F1-F5、PTmean、63的倍频带能量、鼾声时长8个鼾声声学特征组合能很好地区分7类鼾声；进一步通过决策树模型发现普通鼾声、低通气中鼾声、呼吸暂停前鼾声、呼吸暂停中鼾声、呼吸暂停后鼾声能很好地预测打鼾患者的严重程度；提出的VGG19 + LSTM模型能准确预测AHI。本项目的研究为鼾声准确诊断OSAHS奠定了坚实基础。本项目发表相关 SCI 论文5篇，中文核心期刊论文1篇。本项目培养了硕士研究生3名，博士研究生1名。