A Deep Learning Method Approach for Sleep Stage Classification with EEG Spectrogram.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Automated sleep stage classification is crucial but faces challenges with time-consuming, subjective, and error-prone manual methods.
• Existing electroencephalogram (EEG)-based automated methods struggle with accuracy, particularly for the N1 sleep stage, due to data imbalance.

Purpose of the Study:

• To develop an accurate and robust automated sleep stage classification method using EEG spectrograms.
• To address the limitations of current methods in classifying the N1 sleep stage.

Main Methods:

• A deep learning model, EEGSNet, was designed utilizing multi-layer convolutional neural networks (CNNs) for feature extraction from EEG spectrograms.
• Two-layer bi-directional long short-term memory networks (Bi-LSTMs) were employed to learn transition rules between epochs and classify sleep stages.
• Gaussian error linear units (GELUs) were incorporated as activation functions in CNNs to enhance model generalization.

Main Results:

• The EEGSNet model achieved high accuracy across four public datasets (Sleep-EDFX-8, Sleep-EDFX-20, Sleep-EDFX-78, and SHHS), with accuracies ranging from 83.02% to 94.17%.
• The method demonstrated strong performance with MF1 scores between 77.26% and 87.78% and Kappa values from 0.77 to 0.91.
• Significantly improved classification results were observed for the N1 sleep stage, with F1-scores ranging from 47.26% to 70.16%.

Conclusions:

• The proposed EEGSNet method offers a promising approach for accurate and reliable automated sleep stage classification.
• The model's architecture effectively extracts temporal and frequency features and learns epoch transitions, outperforming existing methods, especially for N1 classification.
• This advancement has the potential to streamline sleep analysis and improve diagnostic capabilities in sleep medicine.