Decoding motor imagery tasks using ESI and hybrid feature CNN.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) using motor imagery electroencephalogram (MI-EEG) offer natural interaction.
• MI-EEG signal analysis faces challenges with low spatial resolution and limited data.
• Volume conduction effects and overfitting are significant hurdles in MI-EEG classification.

Purpose of the Study:

• To propose a new framework for MI-EEG binary classification.
• To enhance spatial resolution and mitigate volume conduction effects in EEG signals.
• To address small dataset issues and reduce network overfitting in MI-EEG analysis.

Main Methods:

• Electrophysiological source imaging (ESI) was employed to improve spatial resolution.
• Continuous wavelet transform and optimal time-of-interest (TOI) selection were used for feature extraction.
• A seven-convolution layer neural network was utilized for classification, alongside data augmentation techniques.

Main Results:

• The proposed model achieved high classification accuracies of 93.2% and 95.4% on benchmark datasets.
• Selecting subject-specific best TOIs improved classification accuracy by approximately 2%.
• Noise addition and overlap data augmentation methods enhanced accuracy by at least 4%, while rotation and flip methods decreased it.

Conclusions:

• Combining ESI and data augmentation effectively addresses low spatial resolution and small sample size issues in EEG.
• The developed framework demonstrates superior accuracy and practical potential for MI-EEG binary classification tasks.
• This approach advances the decoding of motor imagery tasks for improved BCI applications.