A Study on Acoustic Characteristics of Snoring in Patients with Obstructive Sleep Apnea Hypopnea Syndrome

蒋燕梅

彭健新

硕士

华南理工大学

阻塞性睡眠呼吸暂停低通气综合征;;鼾声;;特征提取;;神经网络;;随机森林

阻塞性睡眠呼吸暂停低通气综合征(OSAHS)是一种常见的与睡眠相关的呼吸疾病,打鼾是OSAHS患者最直接、最典型的特征。近年来,国内外许多研究者都尝试利用鼾声分析技术辅助诊断OSAHS患者,试图探索一种低廉、便捷、有效的OSAHS患者检测系统。本文通过分析鼾声的声学特征,识别出OSAHS患者和普通打鼾者,并对OSAHS患者的七类鼾声进行分类,进而预测OSAHS患者的AHI值。针对鼾声片段的自动识别,文中提出一种基于声音图谱和神经网络的方法识别鼾声。潜在鼾声片段经子带谱熵法检测出后,提取潜在鼾声片段的时域图、频谱图、语谱图、Mel时频图和CQT时频图,并分别采用模型CNNs-DNNs和模型CNNs-LSTMs-DNNs分类鼾声和非鼾声。结果表明,鼾声与非鼾声在频域特征上存在显著性差异,尤其是低频特征。在本文提取的5种图谱中,Mel时频图能更好地反映鼾声与非鼾声的区别。当使用Mel时频图和模型CNNs-LSTMs-DNNs分类鼾声时,分类效果最佳,能达到95.07%的准确率,95.42%的灵敏度和95.82%的特异性。由于OSAHS患者和普通打鼾者的鼾声有所不同,论文探讨比较了鼾声的常见频域特征和不同分类器在区分OSAHS患者和普通打鼾者的表现能力。在识别出鼾声片段后,从鼾声片段中提取梅尔倒谱系数、800Hz功率比、谱熵等10种声学特征,再利用基于随机森林的特征选择算法筛选出Top-6特征,并用5种机器学习模型验证Top-6特征的有效性。结果表明,在综合考虑分类性能和计算效率的情况下,逻辑回归模型与Top-6特征的组合表现最好,可以成功区分OSAHS患者和普通打鼾者。该方法计算复杂度低、对OSAHS患者鼾声识别率较高,在识别出OSAHS鼾声的基础上,能够评估出患者是否患有OSAHS。OSAHS患者整晚的鼾声特性存在差异,文中将其分为呼吸暂停前鼾声、呼吸暂停中鼾声、呼吸暂停后鼾声、低通气前鼾声、低通气中鼾声、低通气后鼾声和普通鼾声七类。论文提取七类鼾声的梅尔倒谱系数、谱熵、800Hz功率比等声学特征,利用Relef F算法对提取的特征进行筛选,并分别用支持向量机、逻辑回归和随机森林三个机器学习模型对鼾声进行七分类研究。实验结果表明,在分类模型中随机森林的分类能力最强,其与Top-20特征在鼾声的七分类中能够达到80.36%的整体分类准确率,这为利用七类鼾声预测患者的AHI提供基础。

Obstructive sleep apnea hypopnea syndrome(OSAHS)is a common sleep-related respiratory disease and snoring is the most direct and typical characteristic of OSAHS patients.In recent years,many researchers at home and abroad have tried to assist the diagnosis of OSAHS patients using the snoring analysis technology,attempting to explore a cheap,portable and effective OSAHS monitoring system.In this paper,acoustic features of snore sounds were analyzed to identify the patients with OSAHS and simple snorers,and then seven types of snore sounds from OSAHS patients were classified to predict the AHI values of OSAHS patients.To automatically identify snore episodes,a method based on sound images and neural network is proposed in this paper.After detecting the potential snore episodes using the subband spectral entropy method,the time-domain waveform,spectrum,spectrogram,Mel-spectrogram and CQT-spectrogram of potential snore episodes were extracted,and then these sound images were fed into model CNNs-DNNs and model CNNs-LSTMs-DNNs,respectively.The results show that significant difference between snores and non-snores is found in frequency-domain characteristics,especially in the low frequency.Among the five sound images extracted in this paper,Mel-spectrogram can better reflect the difference between snores and non-snores.The best performance with 95.07% accuracy,95.42% sensitivity and 95.82% specificity is achieved by the combination of Mel-spectrogram and model CNNs-LSTMs-DNNs.As the snore of OSAHS patients is different from that of simple snorers,this paper discusses the common frequency-domain characteristics of snores and the performance ability of different classifiers in the classification of OSAHS patients and simple snorers.After identifying the snore episodes,Mel-frequency cepstral coefficients,800 Hz power ratio,spectral entropy and other 10 kinds of acoustic features were extracted from snore episodes and a feature selection algorithm based on random forest was applied to these features to screen Top-6 features.The effectiveness of Top-6 features was verified by 5 kinds of machine learning model.The results show that the combination of logistic regression model and Top-6 features performs best which can successfully distinguish the OSAHS patients from simple snorers considering the classification performance and calculation efficiency comprehensively.The proposed method has low computational complexity and high recognition rate for OSAHS snores,which can assess the patients whether suffer OSAHS based on the identified OSAHS snore sounds.The whole night snore sounds of OSAHS patients are different.In this paper,the snore sounds were classified as snore before apnea,snore during apnea,snore after apnea,snore before hypopnea,snore during hypopnea,snore after hypopnea and simple snore.This paper extracted the acoustic features of seven types of snore sounds,such as Mel-frequency cepstral coefficients,spectral entropy,800 Hz power ratio and other features.Then,a Relief F algorithm was used to screen some important features and the support vector machine,logistic regression and random forest were used to classify seven types of snores,respectively.The experimental results show that the classification ability of random forest is stronger than other classification models.Furthermore,the combination of random forest and Top-20 features perform best in the classification of seven types of snores which can achieve overall accuracy of 80.36%.These results provide the basis for predicting AHI values of patients using seven types of snores.