Analysis and automatic identification of sleep stages using higher order spectra

Area of Science:

• Neuroscience
• Signal Processing
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) signals are crucial for studying brain activity and sleep staging.
• EEG signals exhibit nonlinear and non-stationary characteristics, challenging traditional linear analysis methods.
• Visual interpretation and linear techniques are often insufficient for accurate sleep staging.

Purpose of the Study:

• To apply nonlinear techniques, specifically higher order spectra (HOS), to extract hidden information from sleep EEG signals.
• To develop unique bispectrum and bicoherence plots as visual aids for diagnosing sleep stages.
• To enable automatic sleep stage identification using HOS-based features and machine learning.

Main Methods:

• Utilized higher order spectra (HOS) analysis, including bispectrum and bicoherence, for sleep EEG signal processing.
• Extracted statistically significant HOS-based features across different sleep stages (Wakefulness, REM, Non-REM Stages 1-4) using ANOVA.
• Employed a Gaussian mixture model (GMM) classifier for automated sleep stage identification based on extracted features.

Main Results:

• Developed novel bispectrum and bicoherence plots for visualizing sleep EEG characteristics.
• Identified statistically significant HOS-based features (p < 0.001) differentiating various sleep stages.
• Achieved an 88.7% accuracy rate in automatic sleep stage identification using the proposed GMM classifier.

Conclusions:

• Higher order spectra (HOS) analysis is an effective nonlinear technique for extracting valuable information from sleep EEG signals.
• The proposed HOS-based features and visualization plots can serve as effective diagnostic tools for sleep staging.
• The developed automated system demonstrates high accuracy in identifying sleep stages, offering a promising approach for clinical applications.