RimeSleepNet: A hybrid deep learning network for s-EEG sleep stage classification

Abstract

Sleep stage classification is essential for sleep research and clinical diagnostics. However, frequency aliasing in sleep electroencephalogram (s-EEG) signals remains a significant challenge, existing methods have yet to effectively address this issue. This study proposes a hybrid deep-learning model, RimeSleepNet, comprising four key components. First, the rime optimization algorithm adaptively tunes variational mode decomposition (VMD) to reduce frequency aliasing by generating intrinsic mode functions (IMFs). Second, a convolutional neural network (CNN) automatically extracts stage-specific features from IMFs. A multi-head self-attention (MHSA) mechanism then dynamically weights these features to prioritize stage-specific patterns, followed by long short-term memory (LSTM) networks that model temporal dynamics for robust classification of NREM, REM, and WAKE stages. Evaluated on the Chengdu People's Hospital and Sleep-EDF datasets, RimeSleepNet achieves the highest F1 scores of 0.94, 0.89, and 0.92 for NREM, REM, and WAKE stages, respectively, with an AUC of 0.92, outperforming baseline models like CNN and LSTM. Cross-dataset validation confirms its robust generalization (Cohen's κ = 0.90), and it reduces validation loss by 53 % compared to LSTM, providing an advanced tool for automated sleep stage analysis in sleep disorder diagnosis and personalized monitoring.