A three-branch 3D convolutional neural network for EEG-based different hand movement stages classification.

Area of Science:

• Neuroscience
• Computer Science
• Machine Learning

Background:

• Motor Imagery is a key Brain Computer Interaction (BCI) method using electroencephalogram (EEG) signals.
• Deep learning approaches are exploring effective EEG representations preserving spatial and temporal information.

Purpose of the Study:

• To propose a novel 3D representation of EEG signals.
• To develop an end-to-end three-branch 3D convolutional neural network (3D-CNN) for motor imagery classification.
• To address challenges of class imbalance and varying classification difficulty in EEG datasets.

Main Methods:

• A 3D EEG representation was developed to capture spatial and temporal features.
• An end-to-end three-branch 3D-CNN architecture was designed.
• A class balance cropped strategy and focal loss were implemented to handle dataset imbalances and classification difficulty.

Main Results:

• The proposed 3D-CNN with focal loss demonstrated improved classification performance on the WAY-EEG-GAL dataset.
• The method achieved more balanced classification accuracy across different motor stages.
• Experimental results validate the effectiveness of the 3D representation and network architecture.

Conclusions:

• The proposed 3D EEG representation and three-branch 3D-CNN effectively enhance motor imagery classification.
• The integration of focal loss successfully mitigates issues related to class imbalance and varying classification difficulty.
• This approach offers a promising advancement for BCI applications utilizing EEG data.