Orthogonal convolutional neural networks for automatic sleep stage classification based on single-channel EEG

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) are increasingly used for automatic sleep stage classification from EEG data.
• Current CNN methods are often complex, posing challenges for portable sleep devices.
• Developing simpler CNNs that learn rich EEG representations is crucial.

Purpose of the Study:

• Propose a novel and simpler CNN model for accurate sleep stage classification.
• Enhance the ability of CNNs to learn rich and diverse feature representations from EEG data.
• Improve the feasibility of automatic sleep analysis for wearable technology.

Main Methods:

• Converted EEG signals to time-frequency representations using Hilbert-Huang transform.
• Introduced an Orthogonal Convolutional Neural Network (OCNN) with orthogonal weight initialization.
• Employed orthogonality regularizations and Squeeze-and-Excitation (SE) blocks to maintain weight orthogonality and recalibrate features.

Main Results:

• Achieved high classification accuracies of 88.4% and 87.6% on two public datasets.
• Demonstrated superior performance compared to other CNN models.
• Validated the effectiveness of the proposed OCNN for sleep stage classification.

Conclusions:

• The OCNN effectively learns diverse feature representations from EEG time-frequency images.
• Orthogonality regularization is a simple yet powerful technique adaptable to other architectures.
• The proposed method offers a promising solution for accurate and efficient sleep stage classification.