Staging study of single-channel sleep EEG signals based on data augmentation

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Accurate sleep staging is crucial for assessing sleep quality and conducting sleep research.
• Electroencephalogram (EEG) datasets often exhibit class imbalance due to unequal distribution of sleep stages, hindering automatic sleep staging.
• This imbalance negatively impacts the performance of machine learning models for sleep stage classification.

Purpose of the Study:

• To develop and evaluate a data augmentation model to overcome class imbalance in sleep EEG datasets.
• To improve the accuracy of automatic sleep staging by enhancing minority sleep stage data.
• To validate the proposed model's effectiveness using a public sleep dataset.

Main Methods:

• A novel data augmentation model, Residual Dense Block and Deep Convolutional Generative Adversarial Network (RDB-DCGAN), was proposed.
• The model utilizes two-dimensional continuous wavelet time-frequency maps as input.
• The RDB-DCGAN expands minority sleep EEG data classes before sleep staging by Convolutional Neural Network (CNN).

Main Results:

• The RDB-DCGAN model improved overall sleep staging accuracy by 6% on the Sleep-EDF dataset.
• Classification accuracy for the N1 sleep stage, typically low due to limited data, saw a significant improvement of 19%.
• The results demonstrate the effectiveness of the proposed data augmentation strategy in addressing class imbalance.

Conclusions:

• Data augmentation using improved DCGAN models can effectively resolve class imbalance issues in sleep datasets.
• The RDB-DCGAN approach offers a promising solution for enhancing the reliability of automatic sleep staging.
• This methodology contributes to more accurate sleep quality assessment and research.