Domain - Aware Spatial-Temporal Graph Convolutional Network for Sleep Apnea Detection via Multivariant BCG Signals

• Published in: IEEE International Conference on Communications (2003) pp. 5515 - 5520
• Main Authors: Huang, Yongfeng, Chen, Kuiyou, Zhang, Zhiming
• Format: Conference Proceeding
• Language: English
• Published: IEEE 28.05.2023
• Subjects: Convolution; Domain adaptation; Feature extraction; Graph neural networks; IOT application; Multivariant time series classification; Representation learning; Robustness; Sleep apnea; Sleep apnea detection; Thorax; Time series analysis
• ISSN: 1938-1883
• DOI: 10.1109/ICC45041.2023.10278967

Sleep apnea is a common respiratory disorder that affects up to one billion people globally. It is shown to be an independent risk factor for the cardiovascular diseases and even mortality. Sleep apnea detection via ballistocardiogram (BCG) signals is still a challenging task due to poor signal quality and signal-to-noise ratio. In order to achieve higher exactitude, convolution networks are most frequently-used to capture temporal features. However, the second-order information (movements of the thorax and the diaphragm) of the Multivariant BCG signals, which is naturally more informative and discriminative for sleep apnea detection, is rarely investigated in existing methods. Additionally, it is rarely investigated in existing methods that BCG signals from different subjects are in heterogenous distribution. This may not be optimal for extracting respiratory-relevant features and excluding subject-specific patterns. In this paper, we propose the Domain-aware Spatial-Temporal Graph Convolutional Network (DAST-GCN) to explicitly capture inter-sensor dependencies. Dynamic graph connection and attention mechanism are implemented to fully utilize such dependencies. We further employ an adversarial domain adaptation module to extract domain-invariant features. Experiments on a BCG dataset validate the effectiveness of the proposed method. Furthermore, we illustrate that DAST-GCN captures crucial respiratory patterns and improves the robustness against the domain shift issue.