Identification And Evaluation Of Obstructive Sleep Apnea Hypopnea Syndrome Based On Snoring Sound

Obstructive sleep apnea hypopnea syndrome(OSAHS)is a common chronic disease that caused by continuous partial or complete collapse of the upper airway during sleep.It not only badly influences the sleep quality of OSAHS patients,but also causes some cardiovascular disease and even stroke.The clinically diagnose of OSAHS is Polysomnography(PSG),which is expensive and difficult to popularize.It causes the untreated OSAHS be a substantial but underappreciated public health threat.Snoring sound is the most typical symptom of OSAHS,which is easy to obtain.With the growing interest about sleep health,researchers have extensively explored the relationship between snoring sounds and OSAHS,developing a portable and effective diagnosis system of OSAHS patients based on snoring sounds.In this dissertation,snoring sounds are firstly automatically extracted from the whole night recordings.Then,simple snorers and OSAHS patients are recognized respectively based on the analysis of snoring sounds during the whole night.The severities and the obstruction site of OSAHS are furtherly analyzed.The snoring sounds identification algorithm is proposed based on wavelet packet transform.Firstly,the generation subspace snoring sound enhancement algorithm based on noise covariance matrix estimation is applied to reduce the nonstationary noise of the sleep environment.Then,the potential snoring segments are detected by the adaptive threshold method.Considering the different distribution of sounds,the wavelet packet transform is applied to divide the potential snoring segments into multi-scale and multi-resolution.A serious of acoustic features are extracted for different sub-bands and classified by integrated classifier.Results indicate that snoring sounds and non-snoring sounds have different distribution on different frequency bands.The proposed classifier is of great performance and robustness.For the limitations and instability of OSAHS patients’ identification result based on single classifier,a fused OSAHS patients identification model based on transfer learning is proposed.Three independent models are trained respectively including two transfer learning model and a acoustic model.The modes are fused based on hard voting strategy to finally classify simple snorers and OSAHS patients.Feature visualization results indicate that snoring sounds of simple snorers are stable with relatively concentrated feature distribution.The frequent apneahypopnea causes large fluctuation of snoring sounds of OSAHS patients.The proposed method could well fuse the advantage of classifier and features from different domains to correctly identify OSAHS patients based on snoring sounds extracted by ENT doctor and wavelet packet transform respectively.The work classifies snoring sounds of OSAHS patients into apnea-hypopnea snoring sounds and normal snoring sounds based on VGG19+LSTM fused model.The apnea-hypopnea event further identified based on these snoring sounds to estimate the AHI of OSAHS patients.Subject independence-based experiments show that the AHI estimated based on the proposed model have high correlation with its AHI obtained by PSG.For the problem of limited and unbalanced dataset of snoring sounds from different obstruction site,the dissertation treats it as few-shot classification task and proposed a model based on prototypical network to classify snoring sounds of different obstruction site.The CNN is applied to represent features of snoring sounds with low dimension to construct prototypical network in which each type of snoring sound could be represented as a prototype.The complement cross entropy is applied to solve the problem of unbalanced distribution of snoring sounds.The proposed method could well classify each type of snoring sounds which is better than traditional classification strategy.The analysis results of snoring sounds of OSAHS patients with different severities demonstrate the precision and reliability of the OSAHS identification and evaluation model proposed by the dissertation.It provides reliable experiment for the development of the portable OSAHS diagnosis system based on snoring sounds.