3DSleepNet: A Multi-Channel Bio-Signal Based Sleep Stages Classification Method Using Deep Learning.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate sleep stage classification is crucial for diagnosing sleep disorders.
• Traditional methods often struggle with the complexity and multi-channel nature of biosignals.

Purpose of the Study:

• To develop a novel multi-channel 3D Convolutional Neural Network (3D-CNN) for enhanced sleep stage classification.
• To improve the accuracy and efficiency of sleep stage prediction using electroencephalography (EEG), electromyogram (EMG), and electrooculogram (EOG) data.

Main Methods:

• Extracted time, frequency, and time-frequency domain features from EEG, EMG, and EOG signals.
• Employed 3D-CNN layers to learn inter-signal and inter-frequency band relationships, and 2D-CNN layers for frequency relations.
• Integrated partial dot-product attention for channel and frequency band importance, and LSTM for epoch transition learning.

Main Results:

• Achieved 0.832 overall accuracy, 0.814 F1-score, and 0.783 Cohen's kappa on the ISRUC-S3 dataset.
• Demonstrated strong performance on the ISRUC-S1 sleep-disorder dataset with 0.820 accuracy, 0.797 F1-score, and 0.768 Cohen's kappa.
• Exhibited superior training speed compared to state-of-the-art graph convolutional networks and ResNet architectures.

Conclusions:

• The proposed 3D-CNN model offers a competitive and efficient approach for sleep stage classification.
• The model's ability to learn complex relationships within multi-channel biosignals contributes to its high accuracy.
• The model demonstrates generalizability across healthy and sleep-disordered subjects and offers significant computational advantages.