Motor Imagery Classification for Brain Computer Interface Using Deep Convolutional Neural Networks and Mixup...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Training deep learning (DL) models typically requires large datasets, which is challenging for electroencephalography (EEG) due to long data collection times and inter-individual variability.
• Existing methods often combine data from multiple subjects, potentially reducing classification performance for individual analyses.

Purpose of the Study:

• To develop a DL model trainable on small EEG datasets from a single subject.
• To avoid lengthy EEG data collection and the negative impact of multi-subject data aggregation.

Main Methods:

• A customized Convolutional Neural Network (CNN) architecture was employed.
• Mixup augmentation technique was integrated to enhance model robustness.
• Models were trained using approximately 120 EEG trials per subject.

Main Results:

• Modified ResNet18 and DenseNet121 models with mixup augmentation demonstrated high classification accuracy.
• ResNet18 achieved 0.920 accuracy (95% CI: 0.908, 0.933).
• DenseNet121 achieved 0.933 accuracy (95% CI: 0.922, 0.945).

Conclusions:

• The designed DL classifiers outperformed previous methods on the same dataset.
• Effective classification performance was achieved despite using a limited single-subject training dataset.