Fine-Grained Detection of Apnea-Hypopnea Events Based on Transformer Network in Audio Recordings

Sleep Apnea Hypopnea Syndrome (SAHS) is a common sleep disorder that affects approximately 7% of adults worldwide. The clinical gold standard for diagnosis, polysomnography (PSG), is expensive and unsuitable for long-term monitoring. In recent years, home sleep monitoring methods based on sensors and signal processing technologies have become popular. This paper aims to develop a method for detecting apnea and hypopnea events using only audio signals without causing discomfort to the user. To overcome the limitations of traditional methods in dealing with time-frequency domain dependencies and the problems of low accuracy and ineffectiveness of audio event segmentation at a single scale, a segmentation model based on Transformer and multi-scale feature fusion is proposed to effectively utilize global information and multi-scale features to achieve localization of event start and end times. Experiments on the PSG-AUDIO public dataset achieve a Dice similarity coefficient of 0.98 and the segment-based f1-score of 0.95, demonstrating the great potential of the method in detecting SAHS and providing a new way for non-invasive home sleep monitoring.