MHFNet: A Multimodal Hybrid-Embedding Fusion Network for Automatic Sleep Staging.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Automated sleep scoring is crucial for assessing sleep continuity and structure.
• Existing methods face challenges in fusing temporal information, utilizing signal correlations, and incorporating adjacent epoch logic.

Purpose of the Study:

• To introduce a multimodal hybrid-embedding fusion network (MHFNet) for automated sleep stage scoring.
• To address limitations in current sleep scoring models by integrating local and global temporal information, signal correlations, and scoring rule logic.

Main Methods:

• MHFNet utilizes multi-stream Xception blocks for wave characteristic extraction.
• A hybrid time-embedding module combines local and global temporal data.
• A dual-path gate transformer fuses and enhances attention features.
• A refined output header reconstructs sleep scoring.

Main Results:

• MHFNet demonstrated superior performance over baseline approaches in cross-validation on public datasets (SleepEDF-ST, SleepEDF-SC, SHHS).
• Individual-level testing showed a 9% average R² score improvement compared to state-of-the-art models.

Conclusions:

• MHFNet effectively tackles challenges in automated sleep scoring.
• The model shows significant improvements in accuracy and potential for real-world sleep medicine applications.