Automatic Sleep Stage Classification using Marginal Hilbert Spectrum Features and a Convolutional Neural Network.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Accurate sleep stage classification is crucial for diagnosing sleep disorders.
• Traditional methods often rely on multi-channel EEG, which can be cumbersome.
• Automated analysis of sleep patterns using single-channel EEG is an active research area.

Purpose of the Study:

• To develop a novel, automated method for sleep stage classification using single-channel electroencephalography (EEG).
• To leverage time-frequency domain features and deep learning for improved classification accuracy.
• To address the challenge of class imbalance in sleep EEG data.

Main Methods:

• Utilized marginal Hilbert spectrum (MHS) to extract time-frequency features from 30-second EEG epochs.
• Employed a convolutional neural network (CNN) to process MHS features as multi-channel sequences.
• Integrated a focal loss function within the CNN to mitigate issues arising from imbalanced sleep stage data.

Main Results:

• The proposed method achieved an overall accuracy of 86.14% on the public Sleep-EDF dataset.
• The MHS features effectively captured discriminative information for different sleep stages.
• The CNN model demonstrated robust performance in classifying sleep stages.

Conclusions:

• The novel method based on MHS and CNN provides a competitive and effective approach for automatic sleep stage classification.
• This technique shows promise for sleep analysis using less invasive single-channel EEG recordings.
• Further exploration of this deep learning model is warranted for advancing sleep medicine research.