Multi-scale ResNet and BiGRU automatic sleep staging based on attention mechanism

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Automatic sleep staging is crucial for diagnosing sleep disorders.
• Current methods face limitations in feature extraction and temporal pattern recognition.
• There's a need for more accurate and robust sleep staging models.

Purpose of the Study:

• To propose an automatic sleep staging model using Electroencephalogram (EEG) signals.
• To improve accuracy and overcome limitations of existing sleep staging techniques.
• To validate the model's effectiveness and generalization performance.

Main Methods:

• Developed a deep learning model combining Multi-scale Attention Residual Nets (MAResnet) and Bidirectional Gated Recurrent Unit (BiGRU).
• Incorporated improved channel and spatial feature attention units and multi-scale convolution kernels for enhanced feature extraction.
• Utilized BiGRU to capture temporal dependencies between sleep stages for automatic staging and sleep cycle extraction.

Main Results:

• Achieved 84.24% classification accuracy and a 0.78 kappa coefficient on the sleep-EDF dataset, outperforming traditional residual networks.
• Demonstrated strong performance on UCD (79.34%) and SHHS (81.6%) datasets, indicating good generalization.
• Significantly improved recognition accuracy compared to existing related studies.

Conclusions:

• The proposed MAResnet-BiGRU model offers a significant advancement in automatic EEG sleep staging.
• The model effectively addresses limitations of previous methods by enhancing feature extraction and temporal analysis.
• The validated effectiveness and generalization performance support its potential for clinical application in sleep disorder diagnosis.