Relevance-based channel selection in motor imagery brain-computer interface.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Channel selection is crucial for optimizing electroencephalogram (EEG)-based brain-computer interface (BCI) decoding efficacy.
• Deep learning (DL) models in BCIs necessitate novel channel selection techniques, particularly for subject-independent applications.
• Understanding inter-subject variability in EEG is key for effective subject-independent DL training.

Purpose of the Study:

• To propose and evaluate a novel subject-independent channel selection methodology for DL-based motor imagery (MI)-BCI.
• To leverage Layer-wise Relevance Propagation (LRP) and neural network pruning for identifying optimal EEG channels.
• To assess the impact of LRP-based channel selection on classification accuracy and neurophysiological plausibility.

Main Methods:

• Implemented a subject-independent channel selection approach using LRP and neural network pruning.
• Utilized Deep ConvNet and 62-channel MI EEG data from the Korea University dataset for experiments.
• Compared LRP relevance-based channel selection with conventional weight-based methods.

Main Results:

• Achieved a 61% channel reduction with no significant drop in subject-independent classification accuracy (p=0.09).
• LRP-based selection outperformed weight-based methods, using <40% channels with accuracy improvements of 1.72%-5.96%.
• Sparse-LRP models demonstrated comparable or superior performance to the baseline, even with minimal channels (16%-35%).

Conclusions:

• The proposed LRP-based method effectively reduces channels in DL-based MI-BCI without compromising accuracy.
• This approach offers a significant advancement in addressing traditional EEG-BCI channel selection challenges.
• The study highlights the utility of model interpretability (LRP) as a powerful technique for BCI problem-solving.