Influence of EEG Signal Augmentation Methods on Classification Accuracy of Motor Imagery Events

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Electroencephalography (EEG) is crucial for brain-computer interfaces (BCIs).
• Motor imagery classification using EEG data presents challenges in generalization.
• Data augmentation can potentially improve model robustness and accuracy.

Purpose of the Study:

• To evaluate the impact of diverse data augmentation techniques on EEG-based motor imagery classification.
• To assess the effectiveness of individual and combined augmentation strategies.
• To optimize neural network performance for BCIs.

Main Methods:

• Utilized EEG data from a public dataset for three-class motor imagery event classification (left vs. right hand imagination).
• Employed EEGNet, a convolutional neural network tailored for EEG signal processing.
• Tested multiple augmentation techniques across time, frequency, and spatial domains, including generative approaches.
• Investigated individual, cascaded (2-3 methods), and varied data ratios (1:0.25 to 1:1) with different data-splitting strategies.

Main Results:

• Data augmentation techniques significantly influenced classification accuracy.
• Combined augmentation strategies demonstrated a notable impact on performance.
• The choice and combination of augmentation methods are critical for enhancing generalization.

Conclusions:

• Appropriate selection and combination of data augmentation are vital for improving EEG-based motor imagery classification.
• These findings are essential for advancing the development of effective EEG-based BCIs.
• Further research into optimized augmentation strategies can lead to more robust BCI systems.