A two-stage transformer based network for motor imagery classification.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Brain-computer interfaces (BCIs) are vital for understanding brain function and treating neurological disorders.
• BCIs aid in motor dysfunction rehabilitation and advance motor imagery applications.
• Electroencephalogram (EEG) signals are crucial for classifying motor intentions in BCIs.

Purpose of the Study:

• To enhance classification performance on benchmarked EEG signals for motor imagery.
• To present a novel two-stage transformer-based architecture for improved BCI accuracy.
• To address the challenge of limited training data in deep learning architectures for BCIs.

Main Methods:

• A two-stage transformer architecture combining handcrafted features and deep learning.
• Stage-1: Spatiotemporal feature extraction using EEGNet, multi-head attention, and separable temporal convolution networks.
• Stage-2: Classification using TabNet with features and embeddings from Stage-1, plus a novel channel cluster swapping data augmentation technique.

Main Results:

• Achieved an average classification accuracy of 88.5% on the BCI Competition IV-2a dataset.
• Achieved an average classification accuracy of 88.3% on the BCI Competition IV-2b dataset.
• Demonstrated a 3.0% improvement in accuracy compared to recent reported works.

Conclusions:

• The developed two-stage architecture significantly enhances EEG signal classification for BCIs.
• The novel data augmentation technique effectively handles limited training samples.
• The proposed method offers superior performance for motor imagery applications in BCIs.