A Simplified CNN Classification Method for MI-EEG via the Electrode Pairs Signals

Area of Science:

• Biomedical Engineering
• Neuroscience
• Machine Learning

Background:

• Brain-computer interfaces (BCI) leverage electroencephalography (EEG) for brain-computer communication.
• Selecting optimal electrodes and features is crucial for improving BCI classification performance.
• Existing methods often require extensive preprocessing and feature engineering.

Purpose of the Study:

• To develop a deep convolutional neural network (CNN) for efficient EEG signal processing in BCI applications.
• To investigate the effectiveness of a CNN with separated temporal and spatial filters for motor imagery classification.
• To evaluate the performance of the proposed method using minimal electrode selection.

Main Methods:

• A 5-layer CNN architecture with separated temporal and spatial filters was designed.
• Raw EEG signals from electrode pairs over the motor cortex were used as hybrid samples.
• Max pooling, dropout, and batch normalization were employed for dimensionality reduction and overfitting prevention.
• The model was trained and validated on motor imagery tasks from the Physionet database.

Main Results:

• The proposed CNN achieved a global averaged accuracy of 97.28% on group-level classification.
• The area under the receiver operating characteristic (ROC) curve reached 0.997.
• High accuracy (98.61%) was obtained using only two electrodes (FC3-FC4) for classification.

Conclusions:

• The CNN approach with minimal electrode selection offers a simplified and effective BCI system design.
• This method demonstrates superior performance compared to other approaches on the same dataset.
• The findings accelerate the potential for clinical applications of BCI technology.