Deep-Learning-Based Classification of Cyclic-Alternating-Pattern Sleep Phases.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Cyclic-alternating pattern (CAP) phases in sleep are vital for sleep quality assessment.
• Current CAP classification methods predominantly use classical machine learning and require manual feature extraction.
• Deep learning approaches for CAP classification are underutilized.

Purpose of the Study:

• To develop a fully automatic deep learning algorithm for classifying electroencephalography (EEG) signals for CAP detection.
• To investigate the efficacy of various time-frequency representations for CAP classification.
• To enhance CAP identification by leveraging contextual information and specialized data augmentation.

Main Methods:

• Utilized convolutional neural network architectures for EEG signal classification.
• Employed time-frequency representations, specifically Wigner-based methods, for analysis.
• Incorporated contextual information and data augmentation preserving time-frequency structure.
• Trained the model on the publicly available CAP sleep database (CAPSLPDB).

Main Results:

• Wigner-based time-frequency representations demonstrated superior performance over the short-time Fourier transform for CAP classification.
• The proposed deep learning algorithm achieved 77.5% accuracy on a balanced test set and 81.8% on an unbalanced test set.
• The algorithm surpasses existing machine learning-based methods in CAP classification accuracy.

Conclusions:

• The developed deep learning algorithm offers an efficient and scalable solution for automatic CAP phase identification.
• The approach is suitable for on-device implementation, potentially improving sleep quality assessment tools.
• This study highlights the potential of deep learning and advanced time-frequency analysis in sleep research.