LWSleepNet: A lightweight attention-based deep learning model for sleep staging with singlechannel EEG

Area of Science:

• Biomedical Engineering
• Computer Science
• Sleep Medicine

Background:

• Sleep staging is crucial for health assessment but manual classification is inefficient.
• There is a growing need for portable devices for sleep quality detection.
• Lightweight automatic sleep staging models are required for these devices.

Purpose of the Study:

• To propose a novel attention-based lightweight deep learning model for automatic sleep staging.
• To develop a model that is accurate and computationally efficient for portable applications.
• To evaluate the model's performance on public sleep datasets.

Main Methods:

• Developed LWSleepNet, a lightweight deep learning model utilizing depthwise separable multi-resolution convolutional neural networks.
• Incorporated a temporal feature extraction module with multi-head attention for time-dependent information.
• Replaced standard convolutions with depthwise separable convolutions to reduce parameters and computational cost.

Main Results:

• LWSleepNet achieved 86.6% accuracy and 79.2% Macro-F1 on Sleep-EDF-20.
• The model attained 81.5% accuracy and 74.3% Macro-F1 on Sleep-EDF-78.
• The model has only 180K parameters and 55.3 million floating-point operations per second.

Conclusions:

• LWSleepNet demonstrates excellent prediction performance with significantly reduced parameters.
• The proposed model modules offer portability and accuracy for integration into sleep monitoring devices.
• The study highlights the potential of lightweight deep learning for accessible sleep assessment.