CoMTAnet: An Enhanced Framework for Sleep Apnea Detection via Contrastive Multi-Temporal Respiratory Signal Analysis

Sleep Apnea Syndrome (SAS) is a prevalent sleep-related disorder that poses multiple health risks. Traditional detection methods, relying on polysomnography (PSG), are cumbersome and time-consuming, creating a pressing need for more convenient automated detection approaches. This paper presents CoMTAnet, a deep learning framework based on supervised contrastive learning, designed for automatic detection of SAS using multi-channel respiratory signals. The CoMTAnet framework employs an encoder-classifier architecture, where the encoder leverages a TimesBlock to extract deep temporal features from single-channel signals and Residual Blocks to derive high-level information across multi-channel features, trained using Supervised Contrastive Learning (SCL). The classifier categorizes the signal features into normal, insufficient breathing, and apnea classes. Tested on two extensive public datasets, MESA and SHHS, CoMTAnet demonstrates superior performance in terms of accuracy, precision, recall, and F1 score compared to other state-of-the-art models. This highlights the exceptional effectiveness of the supervised contrastive learning approach in the detection of SAS.