Research And Application Of Sleep Apnea And Hypopnea Detection Based On Audio Signal

Sleep Apnea Hypopnea Syndrome(SAHS)is a common sleep disorder that is closely related to a variety of cardiovascular and neurological disorders and affects approximately 7% of adults worldwide,but is often unrecognized when it occurs.Polysomnography(PSG),the gold standard used clinically to diagnose apnea,is an examination procedure that integrates multiple signal recordings but is cumbersome and expensive,and is not conducive to early detection and long-term monitoring.In recent years,home monitoring methods based on sensor and signal processing technologies have gradually become popular,among which,sound signals,as a noninvasive and easily accessible biosignal,can be collected by mobile recording devices without contacting the human body,and have potential application advantages and this thesis takes this as the research object and land on the practical application of sound event detection,aiming to achieve the examination and assessment of PSG under non-invasive monitoring close to effect,the main research work of this thesis is as follows:First,an acoustic signal-based apnea hypopnea detection model SAHLoc is proposed for non-invasive monitoring during sleep.To address the problem of misclassification and omission that may be caused by the coarse granularity of conventional respiratory event detection,the model converts the task from the fragment cutting and classification process of onedimensional signals to the classification of semantic labels per frame of two-dimensional spectrum to achieve the discovery and localization of respiratory events from the perspective of semantic segmentation through operations such as sound feature extraction.In order to better improve the detection accuracy and address the problems of scale differences and detection efficiency among sound events,this thesis compares and analyzes the performance of multiple semantic segmentation and sound event detection models under this task,and improves the Transformer-based model to modify the sound signal to learn the potential patterns of apnea and hypopnea,avoiding the traditional methods in dealing with the long-time dependencies in the time-frequency domain limitations;for the problems of low feature utilization and poor segmentation edge effect,this thesis proposes a CNN decoder combining the multi-scale feature fusion method of up-hopping connection and attention gating mechanism to output the boundary and time information of respiratory events,and finally obtain more complete and accurate detection effect.Then,a speech enhancement method for breath sound acquisition in home environment is proposed to reduce the impact of unavoidable complex background noise on the detection effect during acquisition.A deep learning speech enhancement model based on improved DTLN is proposed to suppress non-stationary noise by generating a mask for dual signal transformation in the time and frequency domains.To address the local characteristics of noise and the low computational efficiency of RNN,an dilated causal convolutional modeling sequence of TCN is introduced to improve the robustness and speed of the method while improving the detection process of respiratory events to achieve a purer respiratory sound input for the SAHS detection model.Finally,a sleep apnea hypopnea monitoring applet is developed for daily home scenarios.The applet can be used with smartphones to collect and detect breathing sounds during the whole night sleep,provide AHI and other indicators and output sleep breathing event reports,creating conditions for daily monitoring at home and assisting doctors’ treatment.In summary,this thesis proposes a sound signal-based sleep apnea hypopnea detection method that enables frame-level event localization and a noise-reduction method that helps to capture breathing sounds in the home environment,both of which are conducive to promoting early diagnosis and treatment of sleep breathing problems and improving the quality of life of patients.