Hierarchical Transformer for Motor Imagery-Based Brain Computer Interface.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Brain computer interfaces (BCI) enable communication and control through brain signals.
• Motor imagery (MI) electroencephalogram (EEG) signals are crucial for BCI applications.
• Existing BCI algorithms face challenges in accurately classifying complex MI signals.

Purpose of the Study:

• To propose a novel transformer-based classification algorithm for MI-EEG signals.
• To enhance BCI performance by accurately identifying relevant motor imagery periods.
• To introduce a hierarchical transformer architecture for improved feature extraction and attention.

Main Methods:

• Utilized a deep learning transformer model, adapted for EEG signal processing.
• Developed a hierarchical transformer architecture with a high-level transformer (HLT) and low-level transformer (LLT).
• LLT processes short-term intervals, while HLT uses self-attention to focus on relevant features.

Main Results:

• The proposed hierarchical transformer algorithm demonstrated superior performance on four open MI datasets.
• Achieved excellent results in both subject-dependent and subject-independent BCI classification tests.
• The model effectively focuses on relevant time periods within long MI trials, ignoring artifacts.

Conclusions:

• The hierarchical transformer architecture is a promising approach for MI-EEG signal classification.
• This novel algorithm significantly advances the capabilities of brain computer interfaces.
• The method shows potential for real-world BCI applications requiring robust signal interpretation.