Separating obstructive and central respiratory events during sleep using breathing sounds: Utilizing transfer learning on deep convolutional networks

Abstract

Sleep apnea diagnosis relies on polysomnography (PSG), which is resource-intensive and requires manual analysis to differentiate obstructive sleep apnea (OSA) from central sleep apnea (CSA). Existing portable devices, while valuable in detecting sleep apnea, often do not distinguish between the two types of apnea. Such differentiation is critical because OSA and CSA have distinct underlying causes and treatment approaches. This study addresses this gap by leveraging tracheal breathing sounds as a non-invasive and cost-effective method to classify central and obstructive events. We employed a transfer learning strategy on six pre-trained deep convolutional neural networks (CNNs), including Alexnet, Resnet18, Resnet50, Densenet161, VGG16, and VGG19. These networks were fine-tuned using spectrograms of tracheal sound signals recorded during PSG. The dataset, comprising 50 participants with a combination of central and obstructive events, was used to train and validate the model. Results showed high accuracy in differentiating central from obstructive respiratory events, with the combined CNN architecture achieving an overall accuracy of 83.66 % and a sensitivity and specificity above 83 %. The findings suggest that tracheal breathing sounds can effectively distinguish between OSA and CSA, providing a less invasive and more accessible alternative to traditional PSG. This methodology could be implemented in portable devices to enhance the diagnosis of sleep apnea, enabling targeted treatment. By facilitating earlier and more accurate diagnoses, this method supports personalized treatment strategies, optimizing therapy selection (e.g., CPAP for OSA, ASV for CSA) and ultimately enhancing clinical outcomes.