A multi-branch, multi-scale, and multi-view CNN with lightweight temporal attention mechanism for EEG-based motor...

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) are prevalent for decoding motor imagery (MI) from electroencephalogram (EEG) signals.
• Extracting robust spatial-temporal-spectral features from noisy EEG data presents a significant challenge for current methods.

Purpose of the Study:

• To propose MBMSNet, a novel multi-branch, multi-scale, multi-view CNN with temporal attention for enhanced EEG-based MI decoding.
• To improve the accuracy and robustness of brain-computer interfaces (BCIs) by addressing feature extraction limitations in low signal-to-noise ratio EEG.

Main Methods:

• MBMSNet employs multi-view representations of raw EEG signals.
• Independent branches capture distinct spatial, spectral, temporal-spatial, and temporal-spectral features, each with specialized layers and temporal attention.
• Features are fused via weighted concatenation for final classification using a fully connected layer.

Main Results:

• MBMSNet achieved superior performance compared to state-of-the-art models across multiple datasets.
• Specific accuracies include 84.60% on BCI Competition IV 2a, 87.80% on 2b, and 74.58% on OpenBMI.
• The model demonstrates significant potential for reliable and effective brain-computer interface applications.

Conclusions:

• MBMSNet offers a powerful and effective approach for decoding motor imagery from EEG signals.
• The proposed architecture successfully extracts discriminative features, overcoming challenges associated with low signal-to-noise ratios.
• This work advances the development of robust and high-performance BCIs.