Transformer-based hybrid systems to combat BCI illiteracy.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Brain-Computer Interfaces (BCIs) face challenges with low signal-to-noise ratios and user-specific
• BCI illiteracy
• affecting up to 20% of users.
• Transformer models show potential but are under-explored in BCI research.

Purpose of the Study:

• To enhance Brain-Computer Interface (BCI) performance by developing and evaluating novel hybrid architectures.
• To address limitations of low signal-to-noise ratios and improve classification accuracy for both strong and weak BCI learners.
• To investigate the efficacy of integrating Convolutional Neural Networks (CNNs), Transformer blocks, and noise-infusion techniques.

Main Methods:

• Experiment A: Assessed hybrid Convolutional and Transformer Block architectures for binary Motor Imagery (MI) classification.
• Experiment B: Introduced a hybrid system with refined blocks and a Noise Focus Block for robust multi-class MI classification.
• Experiment C: Evaluated architectures on 106 subjects, focusing on robustness across diverse user learning capabilities.

Main Results:

• Experiment A achieved a validation accuracy of 0.914.
• Experiment B's architecture improved multi-class MI classification to 84.5%, notably aiding BCI-illiterate users.
• Experiment C demonstrated high robustness with Kappa >83% and a peak validation accuracy of 88.69% across all subjects.

Conclusions:

• Hybrid integration of Transformers, CNNs, and noise-resonance layers significantly boosts BCI classification performance.
• The proposed methods show particular benefit for weak BCI learners and users with low signal quality.
• Further research is recommended for optimizing hybrid BCI architectures and hyperparameters to overcome existing performance limitations.