A Study of Breathing Sound-based Classification of Patients With Breathing Disorders

Overview

Sleep-disordered breathing can damage the cardiovascular system, and may also lead to dysregulation of the autonomic nervous system, endocrine disorders, and hemodynamic changes, causing multi-system and multi-organ damage. Screening for potential central-type patients among patients with respiratory disorders can help provide scientific diagnosis and treatment decisions, thus achieving precise treatment. Currently, research on the identification of sleep-disordered breathing phenotypes is in its infancy. Sleep-disordered breathing phenotypes, such as obstructive and central respiratory events, vary widely among individuals. Compared to indirect methods such as RIP and SpO2, changes in breathing sounds and snoring during sleep can more directly reflect airway obstruction. Different types of sleep-disordered breathing exhibit different characteristics in terms of snoring. Patients with obstructive sleep apnea experience narrowing or blockage of the airway due to relaxation of the throat muscles during sleep, which leads to breathing pauses and hypopnea events, resulting in decreased blood oxygen levels, arousal, and snoring. Central sleep apnea is caused by problems with the brainstem or respiratory control center, leading to breathing pauses. Snoring is usually not very prominent in patients with central sleep apnea. This study aims to screen for potential central-type patients by analyzing upper airway sounds of patients with sleep-disordered breathing, in order to achieve precise treatment.

Screening for central apnea from obstructive apnea is important for the precise treatment of respiratory disorders. Based on the above assumptions that the time domain and acoustic variability of respiratory sound signals contain key information about the degree of upper respiratory tract obstruction and the role of respiratory effort, this study proposes a sleep breathing disorder category identification model based on respiratory sound analysis. A microphone device and sound card are used to capture the patient's audio signal overnight and transmit it to the Raspberry Pi for processing and storage. The microphone device is worn at the neckline of the patient to collect the sound signal of breathing, which ensures that the sound signal is less affected by the sleeping position. Sleep and wakefulness are then separated from breathing sound signals throughout the night and the patient's sleep period is analyzed individually. The apnea location is determined in 30s frames, and in apnea event detection, if the sound stops and lasts for more than 10 seconds, it may be a apnea event. Taking the sound signal of 20s to 30s before apnea as the analysis object, the OpenSmile and Tsfresh feature extraction tools are used to extract acoustic features and envelope features, respectively. The acoustic signature reflects the frequency domain information of apnea, and the envelope feature reflects the time domain signature of apnea. Fusion of acoustic and envelope features enables analysis of airway obstruction and respiratory effort in patients with respiratory disorders. Finally, a machine learning model is established using acoustic features and envelope features as inputs, and each apnea event is classified one by one. In this study, two centers are included, namely the Sleep Therapy Center of the First People's Hospital of Huai'an and the Sleep Therapy Center of the Jiangsu Provincial People's Hospital. Sleep audio data for 167 and 62 cases are expected to be included. The training and validation sets used for modeling are 90 cases, using ten-fold cross-validation, the internal test set is expected to include 77 sleep audio data, and the audio data of 62 patients collected from Jiangsu Provincial People's Hospital are used as the external test set.