A Novel Sleep Staging Network Based on Data Adaptation and Multimodal Fusion

Area of Science:

• Biomedical Engineering
• Neuroscience
• Artificial Intelligence

Background:

• Sleep staging is vital for diagnosing sleep disorders but is labor-intensive.
• Current methods often rely on single electroencephalogram (EEG) channels and expert knowledge.
• Single-channel EEG may not capture the full complexity of physiological sleep states.

Purpose of the Study:

• To develop an automatic sleep staging model using multimodal data.
• To overcome limitations of hand-engineered features and single-channel EEG.
• To fuse electroencephalogram (EEG) and electrooculogram (EOG) signals for improved accuracy.

Main Methods:

• Proposed a deep learning model integrating 3D-CNN for EEG time-frequency analysis and LSTM for EOG frequency evolution.
• Employed data adaptation and multimodal feature fusion techniques.
• Utilized a deep probabilistic network to model nonlinear relationships between EEG and EOG features.

Main Results:

• Achieved state-of-the-art performance on both the SLEEP-EDF and a private dataset.
• Demonstrated the effectiveness of multimodal feature fusion.
• Model predictions closely matched expert sleep stage diagnoses.

Conclusions:

• The proposed multimodal deep learning approach significantly enhances automatic sleep staging accuracy.
• This method offers a more comprehensive and data-driven alternative to traditional techniques.
• The model shows strong potential for clinical application in sleep disorder diagnosis.