A novel decoding method for motor imagery tasks with 4D data representation and 3D convolutional neural networks.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery electroencephalography (MI-EEG) is crucial for intelligent rehabilitation systems.
• Current MI-EEG data representations often lack complete cortical activity information.
• 3D convolutional neural networks (3DCNNs) show promise for recognizing MI tasks.

Purpose of the Study:

• To develop a novel data representation for MI-EEG using raw spatiotemporal dipole information.
• To create a matching 3DCNN architecture for improved MI task recognition.
• To enhance the accuracy and reliability of brain-computer interfaces for rehabilitation.

Main Methods:

• Developed the Electroencephalography Source Imaging with Cascaded Convolutional Neural Network Decoding (ESICNND) method.
• Mapped MI-EEG data to the cerebral cortex using standardized low-resolution electromagnetic tomography.
• Generated 4D dipole feature matrices (4DDFMs) by selecting optimal dipole sampling points and performing interpolation and down-sampling.
• Utilized a 3DCNN with a three-module cascading architecture (3M3DCNN) for feature extraction and classification.

Main Results:

• Achieved average ten-fold cross-validation classification accuracies of 88.73% and 96.25% on two public datasets.
• Demonstrated outstanding consistency and stability in classification performance through statistical analysis.
• The 4DDFMs effectively capture spatiotemporal variations in cortical activation.

Conclusions:

• The proposed ESICNND method, utilizing 4DDFMs and a 3M3DCNN, significantly improves MI task recognition.
• This approach maximizes the use of high-resolution spatiotemporal information from all dipoles.
• The findings offer a more comprehensive data representation for MI-EEG analysis in intelligent rehabilitation.