BiTS-SleepNet: An Attention-Based Two Stage Temporal-Spectral Fusion Model for Sleep Staging With Single-Channel EEG

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Automated sleep staging is vital for sleep quality assessment and diagnosing sleep disorders.
• Single-channel electroencephalography (EEG) offers portability and accessibility for sleep monitoring.
• Existing methods often overlook spectral information and inter-stage transition rules.

Purpose of the Study:

• To propose an attention-based two-stage temporal-spectral fusion model (BiTS-SleepNet) for enhanced automated sleep staging.
• To overcome limitations of existing methods by integrating temporal, spectral, and contextual sleep stage information.

Main Methods:

• Developed BiTS-SleepNet with two stages: Stage 1 for temporal-spectral feature extraction and fusion using cross-attention, and Stage 2 for feature context and transition rule learning (Bi-GRU and CRF).
• Employed a dual-stream architecture to extract and fuse temporal and spectral EEG features.
• Utilized Conditional Random Field (CRF) to incorporate transition rules between sleep stages.

Main Results:

• BiTS-SleepNet achieved high accuracies on public datasets: 88.50% (Sleep-EDF-20), 85.09% (Sleep-EDF-78), and 87.01% (SHHS).
• The model demonstrated competitive performance compared to recent state-of-the-art methods.
• The fusion of temporal and spectral information, along with context and transition rules, significantly improved staging accuracy.

Conclusions:

• BiTS-SleepNet offers a promising approach for accurate automated sleep staging using single-channel EEG.
• The model's ability to integrate diverse features enhances its potential for practical clinical applications.
• This method advances the field of automated sleep analysis by addressing key limitations in current techniques.