Adaptive Time-Frequency Attention Network for Sleep Stage Classification Using Respiratory Signals

Sleep stage classification typically requires the uncomfortable and expensive polysomnography (PSG) test, which limits its widespread use in long-term monitoring and home-based environments. In this paper, we propose SleepTFANet, a novel deep learning model designed for automatic sleep stage classification using respiratory signals. SleepTFANet introduces two key modules: the Time Attention Module (TAM) and the Frequency Attention Module (FAM), to concurrently capture local and global dependencies in time-series data. TAM utilizes Transformer-based patch segmentation to capture temporal dependencies, while FAM leverages spectral analysis to extract frequency-based features. To further enhance performance, we propose the Enhanced Harmonic Energy Allocation (EHEA) method that dynamically adjusts the weighting between these two modules based on the periodicity of the input time series, allowing the model to better adapt to varying signal dynamics. Experimental results demonstrate that SleepTFANet consistently outperforms benchmark models and current state-of-the-art methods, showcasing superior robustness and generalization across multiple real-world datasets. Furthermore, ablation studies confirm the indispensability of each component in our proposed model, offering a promising approach for automatic sleep stage classification using respiratory signals.