Improving Cross-State and Cross-Subject Visual ERP-Based BCI With Temporal Modeling and Adversarial Training

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) offer vital communication for individuals with neuromuscular impairments.
• Event-related potential (ERP)-based BCIs face performance degradation in real-world scenarios, particularly with varying mental states and across different users.

Purpose of the Study:

• To enhance the performance of visual ERP-based BCIs under diverse mental workload conditions.
• To address challenges in cross-state and cross-subject BCI applications.

Main Methods:

• Utilized a hierarchical recurrent network for temporal modeling of electroencephalography (EEG) signals.
• Incorporated adversarial training with dynamic adversarial perturbations to improve model robustness and generalization.
• Applied the method to a visual ERP-based BCI task with 15 subjects across 3 auditory workload states.

Main Results:

• The hierarchical recurrent network effectively modeled long-sequence EEG data.
• The proposed method demonstrated superior performance compared to baseline approaches, especially in cross-state and cross-subject conditions.
• Deep learning with adversarial training significantly improved ERP-based BCI accuracy using limited EEG data.

Conclusions:

• The developed temporal modeling and adversarial training approach offers a robust solution for improving ERP-based BCIs.
• This method shows promise for real-world BCI applications by enhancing reliability across different users and mental states.
• The study highlights the potential of deep learning techniques, particularly adversarial training, in advancing BCI technology with limited data.