LMCSleepNet: A Lightweight Multi-Channel Sleep Staging Model Based on Wavelet Transform and Muli-Scale Convolutions

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Sleep staging is vital for sleep quality assessment, sleep monitoring, and diagnosing sleep disorders.
• Current methods face challenges in extracting salient features from multi-channel sleep data and have excessive parameters, hindering efficiency.
• Developing efficient and accurate sleep staging models is crucial for clinical applications.

Purpose of the Study:

• To propose a lightweight multi-channel sleep staging network (LMCSleepNet) addressing feature extraction and parameter efficiency limitations.
• To enhance feature extraction using continuous wavelet transform and multi-scale convolutions.
• To optimize model parameters using depthwise separable convolutions and attention mechanisms.

Main Methods:

• LMCSleepNet utilizes a four-module architecture: continuous wavelet transform for frequency enhancement, multi-scale convolutions for time-frequency feature extraction, optimized ResNet18 with depthwise separable convolutions, and Convolutional Block Attention Module (CBAM) for spatial correlation.
• The model was evaluated on public datasets (SleepEDF-20, SleepEDF-78).
• Experiments analyzed the impact of temporal sampling points and multi-scale dilated convolution fusion methods.

Main Results:

• LMCSleepNet achieved high classification accuracies of 88.2% (κ = 0.84, MF1 = 82.4%) on SleepEDF-20 and 84.1% (κ = 0.77, MF1 = 77.7%) on SleepEDF-78.
• The model significantly reduced parameters to 1.49 M, demonstrating improved efficiency.
• Experimental validation confirmed the influence of wavelet transform parameters and convolution fusion methods on performance.

Conclusions:

• LMCSleepNet is an efficient and lightweight model for multi-channel sleep staging.
• The network effectively extracts and integrates multimodal features from Polysomnography (PSG) data.
• Its efficiency makes LMCSleepNet suitable for resource-constrained environments, facilitating broader application in sleep monitoring and disorder diagnosis.