Extracting Multi-Scale and Salient Features by MSE Based U-Structure and CBAM for Sleep Staging

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Sleep disorders (somnipathy) are a growing global health concern, affecting millions worldwide.
• Accurate sleep staging is crucial for assessing sleep quality and diagnosing related neurological and psychiatric conditions.
• Current deep learning methods for sleep staging face challenges in modeling salient wave characteristics and inter-stage transitions, alongside issues with imbalanced datasets.

Purpose of the Study:

• To develop an effective deep learning model for automatic sleep staging from single-channel EEG signals.
• To address limitations in modeling intrinsic wave characteristics and transition rules between sleep stages.
• To overcome the class imbalance problem in sleep staging datasets.

Main Methods:

• A novel deep neural network combining Multi-Scale Extraction (MSE)-based U-structure and Convolutional Block Attention Module (CBAM) was constructed.
• The MSE U-structure extracts multi-scale features from raw EEG signals.
• CBAM enhances focus on salient variations and learns transition rules, while a class-adaptive weight cross-entropy loss function addresses data imbalance.

Main Results:

• The proposed model demonstrated superior performance across three public datasets (Sleep-EDF-39, Sleep-EDF-153, SHHS).
• Achieved high overall accuracy (up to 90.3%) and macro F1-scores (up to 86.2%), outperforming existing state-of-the-art methods.
• The model effectively captured multi-scale salient waves and transition rules, mitigating class imbalance issues.

Conclusions:

• The developed deep learning model shows significant promise for automatic sleep staging.
• It offers a robust and accurate alternative to human experts in sleep stage classification.
• The approach effectively models complex EEG signal characteristics and transitions, advancing the field of sleep analysis.