Machine Learning-based Snoring Monitoring And Acoustic Characterization

Obstructive sleep apnea syndrome(OSAHS)is a common disorder that occurs during breathing due to the presence of partial or complete obstruction of the upper airway.According to the American Academy of Sleep Medicine,the number of people with OSAHS grew from 980 million to 1.05 billion worldwide between 2015 and 2019 and is expected to grow to about 1.15 billion by 2024,making the disease a serious clinical contributor to various cardiovascular diseases.Currently,OSAHS is diagnosed mainly by drug-induced sleep endoscopy(DISE),but DISE is not only expensive,but there are large differences between drug-induced sleep and natural sleep,which can affect the final pathological diagnosis.Therefore,there is an urgent need to investigate a non-invasive detection method for pathological diagnosis of OSAHS that can accurately collect user data without user sensitization.To address this goal,this thesis investigates an identification and diagnosis method based on patient snoring sound(SNS)acquisition and acoustic pathology feature construction,which combines subtle features in sound to establish correlations with different pathologies of OSAHS to achieve noninvasive condition tracking and diagnosis analysis of OSAHS patients.The specific work of this thesis includes.To address the problem that traditional acoustic features as well as low-level descriptors perform poorly in mining the deeper features of snoring,a feature extraction method based on multi-level discrete wavelet transform(DWT)and local patterns is investigated in this thesis.First,DWT is used to decompose the audio signal,and local eight-valued patterns are used to extract features at each decomposition level.According to the experimental results,the best classification of the feature vector is achieved by choosing a 7-level DWT.7-level DWT and local octet pattern together construct a feature generation network with 4096 features.In order to reduce the dimensionality of the features to ensure the lightweight of the model,an iterable feature selection method based on ReliefF is investigated,and the KNN loss function is used in the process to calculate the loss values of different feature subsets and select the best feature subset.The results show that the UAR of the model is improved by 14.1%compared to the local eight-valued pattern feature extraction method only,and by 9.6%compared to MFCC,especially the recognition accuracy of T-class snoring is improved by 37.5%,which verifies the universality of the method.To further improve the accuracy of recognition,a deep neural network feature extraction method based on time-frequency map is studied in this thesis.In the data preparation stage,the short-time Fourier transform(STFT)is used to convert the audio signal into a timefrequency map,which solves the problem of varying audio length in the corpus.The fully connected layer outputs of AlexNet and VGG19-based neural networks are used as features to investigate the effects of color mapping of time-frequency maps and neural network architecture on the classification effect,and support vector machines(SVMs)are used as classifiers according to the experimental results.The results show that the unweighted average recall(UAR)of the proposed method is improved by 26.8%,3.3%,and 8.6%compared to CNN+LSTM,DualConvGRU,and baseline statistical function methods.The generalization and accuracy of the model are improved compared with other deep learning methods,and it is more adaptive in the face of new data.In summary,in order to achieve non-invasive condition tracking monitoring and diagnostic analysis of OSAHS patients,this thesis analyzes the acoustic characteristics of snoring,proposes a classification model with higher recognition accuracy and stronger generalization,and lays the foundation for developing a snoring-based OSAHS auxiliary diagnosis method.