An Time-Frequency Graph Attention Network for Obstructive Sleep Apnea Detection

Obstructive sleep apnea (OSA), characterized by interruptions in breathing during sleep, is among the diseases with a high incidence and long-term harm. With the development of wearable devices and advancements in deep learning techniques, methods utilizing deep learning to process photoplethysmography (PPG) and oxygen saturation (SpO2) signals have shown significant potential in OSA detection. However, recent work fails to explicitly capture the intricate relationships within and between single modalities, the long-range dependencies of signals from different modalities, and the lack of model interpretability. Recently, graph attention neural networks have progressed in tasks related to time series, enabling better extraction of spatial features and in-terpretability. This paper proposes a time-frequency graph attention network for OSA detection to address this challenge and research gap. The proposed method uses a time-frequency feature extraction module to extract contextual dependencies between signals. It introduces a graph attention neural network to explore spatio-temporal dependencies among multi-channel signals. Experimental results on the MESA benchmark dataset demonstrate that the proposed model outperforms existing deep learning-based methods in OSA detection tasks.