Cross-subject motor imagery EEG signal classification based on meta-transfer learning

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Motor imagery (MI) electroencephalography (EEG) classification is crucial for brain-computer interfaces (BCIs).
• EEG signals suffer from low signal-to-noise ratio and non-stationarity, hindering traditional machine learning approaches.
• Existing methods show limitations in cross-subject and small-sample classification tasks.

Purpose of the Study:

• To develop an optimized deep learning model for MI-EEG classification.
• To address the challenges of low signal quality and limited data in BCI research.
• To enhance the performance of BCIs in practical, real-world, and clinical settings.

Main Methods:

• Proposed CMHA-Net, a Convolutional Neural Network (CNN) specifically designed for MI-EEG.
• Integrated depthwise separable convolution, deep convolution, and multi-head attention mechanisms.
• Employed a Meta-SGD-based meta-transfer learning framework for improved generalization.

Main Results:

• Achieved 81.61% accuracy on the BCI-IV-2a dataset and 88.15% on the High Gamma dataset.
• Outperformed existing models by a significant margin of 4-15% in accuracy.
• Demonstrated superior performance in small-sample classification scenarios.

Conclusions:

• CMHA-Net offers a robust solution for MI-EEG classification, overcoming key limitations of current methods.
• The proposed model shows significant potential for advancing clinical and real-world BCI applications.
• Meta-transfer learning combined with specialized CNN architecture enhances BCI system usability and effectiveness.