Non-contact Detection of Sleep Apnea Event based on Respiratory Signal using Deep Learning

Sleep apnea is a common sleep disorder. In recent years, numerous studies have concentrated on developing automated diagnostic methods for sleep apnea based on machine learning and deep learning. However, most of these methods use the signals from PSG sensors, which usually obstruct the patients' breathing and cause discomfort. Radar-based contactless detection techniques overcome these limitations, but there are still some drawbacks. For example, most focus on sleep apnea classification and cannot localize events precisely. In this study, we propose a fine-grained, non-contact sleep apnea detection method based on respiratory signals. Our method combines convolutional neural networks and multi-head self-attention mechanisms. We evaluate the method on clinical datasets, and the results demonstrate that this method achieves an accuracy of 83.7% in 1-second-level sleep apnea segment detection. Additionally, the Intraclass Correlation Coefficient (ICC) for the Apnea Index (AI) is 0.915, and the Mean Absolute Error (MAE) is 6.34. It has the potential to serve as an at-home monitoring method.