Automatic Snoring Classification Based On Deep Learning For Different Pathological Mechanisms

There are two completely different sleep apnea disorders,one is central sleep apnea(CSA);the other is obstructive sleep apnea(OSA).Due to the intermittent stop of the motor signals controlled by the central nervous system,apnea is accompanied by the decrease of blood oxygen saturation.The snoring after the end of the pause is defined as CSA snoring;Due to upper airway stenosis or obstruction,resulting in apnea accompanied by decreased blood oxygen saturation,this kind of snoring after the end of apnea,we define as OSA snoring.This paper classifies two kinds of snoring from the aspect of audio signal,which promotes the development and application of acoustic signal processing and medical diagnosis.Our recording data were synchronized with clinical PSG diagnosis,and the experimental data of 90 patients were manually labeled with CSA snoring and OSA snoring.The upper airway impulse response(UAIR)is separated by homomorphic deconvolution,from which five artificial features are extracted.The ROC results show that UAIR can distinguish two kinds of snoring.In this paper,a one-dimensional convolution neural network(1D CNN)with one-dimensional UAIR as network input is designed.The structure of 1D CNN includes three convolution layers and three full connection layers.In view of the imbalance in the number of two types of snoring data,this paper uses a variety of data augmentation methods for CSA snoring,and uses four fold cross validation to get the recall rate of CSA is 72.27%,and that of OSA is 86.4%,and compares with the original audio signal,complex cepstrum,real Cepstrum and logarithmic Mel spectrum,which is higher than the results of these classic features.Considering that the deeper the network is,the more effective features can be extracted.One dimensional deep residual networks(1D ResNet)is designed.On the basis of 1D ResNet,the squeeze-and-excitation model and dropout method are discussed.The results are better than 1D CNN.The recall rate of CSA and OSA is 76.21% and 83.51%,respectively.In this paper,the method of deep learning is superior to the traditional artificial characteristics,which has high academic research value and is worthy of further discussion.Our research is a simple and non-contact new auxiliary diagnosis method,which provides a more reliable basis for medical diagnosis service.