Snoring Recognition And Classification Of Obstructive Sleep Apnea Hypopnea

Obstructive sleep apnea hypopnea syndrome(OSAHS)is a common sleep breathing disorder.At present,the medical community mainly diagnoses and analyzes it based on polysomnography(PSG).Snoring is the most significant clinical features of OSAHS.Snoring analysis has the characteristics of simple operation and non-invasive.In recent years,it has been widely used for the study of auxiliary diagnosis of OSAHS.In this paper,the snoring and non-snoring segments were separated through the analysis of sleep recording signals.The acoustic features were used to analyze OSAHS patients and simple snorers.The pathological snoring and simple snoring were classified,and the sleep related apnea hypopnea index(AHI)was estimated.For the automatic recognition of snoring segments,this paper proposes a recognition algorithm based on the combination of acoustic features and XGBoost model.After the sound segment is detected by the adaptive threshold method,the bark sub-band features,Mel cepstrum coefficient,linear predictive coding,pitch frequency,800 Hz power ratio,crest factor,maximum power ratio and other features are extracted.The Fisher ratio method is used for feature selection to reduce the computational complexity.The results show that after feature selection,the accuracy of feature collection classification is 87.22%-94.34%,and the bark sub-band feature and Mel cepstrum coefficient contribute the most to the classification.The results show that the classification accuracy of the feature set after feature selection is 87.22%-94.34%,and the bark sub-band feature and Mel cepstrum coefficient contribute the most to the classification,indicating that the bark sub-band feature has a good effect on snore recognition.The proposed algorithm takes about 348 s to process the 6-hour recording data all night,and has good operation efficiency and performance.Due to the difference of snoring between OSAHS patients and simple snorers,the Mel cepstrum coefficient,perceptual linear predictive,bark sub-band featuers,spectral entropy,formant,800 Hz power ratio and other characteristics of subjects’ snoring were extracted.The Gaussian mixture model was used to model the snoring characteristic distribution of subjects to realize OSAHS patient screening.Fisher ratio was used for feature selection to compare the performance of models in different dimensions.Experiments show that the overall accuracy of using the features of the Top100 dimension for the subjects reaches 90.00%,in which the features of spectral entropy and gamma tone cepstrum coefficient(dimension 9)contribute the most,and the prediction time is 0.134 ± 0.005 s respectively.This method has good computational efficiency and performance.On the basis of the above research,the subjects’ snoring was divided into respiratory disorder event-related snores and simple snores.Three models were established for classification: 1 Based on acoustic characteristics and XGBoost;2.Based on Mel map and convolution neural network;3.Based on Mel map and residual neural network,the model is fused.The results showed that the highest accuracy of model fusion was 83.44%,and the AHI of subjects was estimated.The Pearson correlation coefficient was 0.913 and the consistency correlation coefficient was 0.9076.This paper discusses an efficient snore detection algorithm through the Fisher ratio feature dimension reduction method and acoustic features;Using the Gaussian mixture model to screen patients with OSAHS has achieved good results and computational efficiency;Furthermore,xgboost,convolutional neural network,and residual neural network are fused to classify ordinary snoring and pathological snoring,so as to estimate AHI index,which provides a basis for snoring assisted diagnosis of OSAHS and prediction of AHI.