MVF-SleepNet: Multi-View Fusion Network for Sleep Stage Classification

Area of Science:

• Biomedical Engineering
• Neuroscience
• Artificial Intelligence

Background:

• Automated sleep stage classification is crucial for health monitoring but current models lack clinical applicability.
• Manual sleep scoring is time-consuming and requires expert knowledge.

Purpose of the Study:

• To develop a novel multi-view fusion network (MVF-SleepNet) for accurate sleep stage classification.
• To leverage multi-modal physiological signals including EEG, ECG, EOG, and EMG.

Main Methods:

• Constructed two views: Time-frequency (TF) images and Graph-learned (GL) graphs from multi-modal signals.
• Employed VGG-16 and GRU for spectral-temporal representation from TF images.
• Utilized Chebyshev graph convolution and temporal convolution for spatial-temporal representation from GL graphs.
• Fused these representations to enhance classification performance.

Main Results:

• MVF-SleepNet achieved 82.1% accuracy, 0.802 F1-score, and 0.768 Kappa on the ISRUC-S1 dataset.
• On the ISRUC-S3 dataset, it reached 84.1% accuracy, 0.828 F1-score, and 0.795 Kappa.
• Demonstrated competitive performance against state-of-the-art baselines.

Conclusions:

• The proposed MVF-SleepNet effectively classifies sleep stages using multi-modal signals.
• Fusion of spectral-temporal and spatial-temporal representations significantly improves classification accuracy.
• MVF-SleepNet shows promise for clinical application in sleep monitoring.