Motor Imagery, Execution, and Observation Classification using Small Amount of EEG Data with Multiple Two-Class CNNs

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) data is crucial for understanding brain activity.
• Classifying complex brain states like motor imagery and execution is challenging.
• Convolutional Neural Networks (CNNs) show promise in analyzing EEG data.

Purpose of the Study:

• To compare the classification accuracy of a single five-class CNN versus multiple two-class CNNs for EEG data.
• To evaluate the impact of using 19-channel EEG versus 4-channel EEG on classification performance.
• To determine the effectiveness of different CNN approaches for classifying motor-related brain states.

Main Methods:

• Utilized electroencephalogram (EEG) data from five states: four motor tasks and resting.
• Implemented two CNN architectures: a single five-class CNN and multiple two-class CNNs.
• Compared classification accuracy using 19-channel EEG versus 4-channel EEG data.

Main Results:

• Multiple two-class CNNs achieved higher classification accuracies (52.8 ± 9.7% with 19-channel EEG) compared to the single five-class CNN (48.2 ± 5.9% with 19-channel EEG).
• The approach using all 19 scalp electrodes generally yielded better results than using only 4 channels.
• The multiple two-class CNNs demonstrated effectiveness even with a small dataset.

Conclusions:

• Multiple two-class CNNs are more effective for classifying motor imagery, execution, and observation states from EEG data.
• Utilizing all 19 scalp electrodes provides a more comprehensive input for CNN-based EEG classification.
• This approach offers a viable method for brain-computer interfaces and neurological studies, especially with limited data.