Online continual decoding of streaming EEG signal with a balanced and informative memory buffer

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain Computer Interface (BCI) systems are crucial for individuals with neurological impairments.
• Online learning is essential for real-world BCI applications in clinical assistance and rehabilitation.
• Catastrophic forgetting is a major challenge in sequential EEG decoding due to inter-subject variability.

Purpose of the Study:

• To address catastrophic forgetting in online sequential EEG decoding for BCI systems.
• To develop a memory-based approach that handles sequentially arriving subjects with imbalanced data volumes.
• To enable long-term continual classification in subject-agnostic BCI scenarios.

Main Methods:

• A memory-based approach with a dynamically selected, balanced memory buffer.
• A kernel-based subject shift detection method for subject-agnostic decoding.
• Development of challenging benchmarks for streaming EEG data.

Main Results:

• The proposed model effectively mitigates catastrophic forgetting in sequential EEG decoding.
• Maintained performance across previously seen subjects over extended periods.
• Demonstrated the model's potential for real-world BCI applications.

Conclusions:

• The memory-based approach successfully enables continual classification in online BCI systems.
• The subject shift detection method enhances robustness in subject-agnostic scenarios.
• The findings support the practical deployment of BCI systems for neurological impairment management.