HFSleepNet: An Automatic Sleep Staging Model Based on Hierarchical Feedback Fusion Aided Full Sequence Encoder

Sleep staging is critical to assess sleep quality and diagnose sleep disorders. Traditional sleep staging methods usually ignore the interactions between different-scale features and lack concurrent attention to both long-term and short-term characteristics of EEG signals. Thus, distinctive features cannot be extracted and modeled exactly. To bridge these gaps, an automatic sleep staging model based on hierarchical feedback fusion aided full sequence encoder called HFSleepNet is proposed in this paper, using EEG sequence. The backbone module with multiple network layers is used to extract feature sequences at multiple time scales from single-channel EEG signals, capturing both long-term and short-term EEG characteristics. The hierarchical feedback fusion module reinforces and integrates EEG features through downward feedback fusion between convolutional layers and auxiliary feature fusion mechanism to extract multi-scale features and their interaction patterns. The full sequence encoder module outputs sleep staging results by capturing concurrent static and dynamic sequence dependencies in location coding. The model is tested on three public datasets, namely Sleep-EDF-2013, Sleep-EDF-2018, and SHHS. For the first two datasets, the classification accuracy of single-channel Fpz-Cz is 87.1% and 84.7%, respectively. For SHHS, the overall accuracy of single-channel C4-A1 is 85.3%. The experimental results show superiority of the proposed method over the state-of-the-art methods for automatic sleep staging.