Breaking the Depth Barrier in Motor Imagery Classification via a Residual Depthwise-Separable Network

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) electroencephalogram (EEG) decoding is crucial for brain-computer interfaces (BCIs).
• Existing lightweight networks using depthwise-separable convolutions are often shallow, limiting hierarchical feature extraction.
• Challenges in integrating residual structures with deeper networks hinder performance improvements.

Purpose of the Study:

• To propose a novel residual depthwise-separable deep neural network (ResDSNet) for enhanced MI-EEG decoding.
• To address the depth barrier in current BCI networks by optimizing residual connections.
• To improve the compatibility of data preprocessing with residual modules for MI tasks.

Main Methods:

• Developed ResDSNet based on an analysis of residual connection structures, optimizing layer distribution.
• Designed a novel residual depthwise-separable convolution module.
• Created a tailored data-preprocessing module to filter noise and retain MI features.

Main Results:

• ResDSNet achieved high accuracies on three public datasets: 79.36% (BCI Competition IV IIa), 84.95% (BCI Competition IV IIb), and 64.13% (PhysioNet).
• Outperformed state-of-the-art methods by significant margins (3.16%, 1.59%, and 8.40%).
• Demonstrated effective unlocking of hierarchical representation capabilities for robust MI-EEG decoding.

Conclusions:

• ResDSNet effectively overcomes the depth barrier in MI-EEG decoding networks.
• The proposed method significantly enhances BCI performance by leveraging deep hierarchical features.
• ResDSNet shows substantial potential for advancing BCI technology through improved EEG signal processing.