Channel Selection for Stereo- Electroencephalography (SEEG)-Based Invasive Brain-Computer Interfaces Using Deep...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• High-throughput brain-computer interfaces (BCIs) utilize numerous electrodes for enhanced performance, leading to signal redundancy and invasiveness.
• Effective channel selection is crucial for invasive BCIs, particularly those employing deep learning, yet remains under-explored.

Purpose of the Study:

• To propose and evaluate two novel deep learning-based channel selection methods, Gumbel and STG, for invasive BCIs.
• To compare the performance of Gumbel and STG against traditional methods (manual, mutual information) and no selection.

Main Methods:

• Developed two deep learning algorithms, Gumbel and STG, for channel selection in BCIs.
• Evaluated methods using Stereo-electroencephalography (SEEG) signals for classifying five distinct movements.
• Compared classification accuracy with manual selection, mutual information (MI), and all-channel approaches.

Main Results:

• With 10 selected channels, Gumbel achieved 65% accuracy, outperforming STG (60%), manual selection (60%), MI (47%), and all channels (59%).
• Gumbel and STG successfully identified key motor control areas (pre-central and post-central) in SEEG recordings.
• Both deep learning methods maintained high decoding accuracy while selecting informative channels.

Conclusions:

• Gumbel and STG are effective deep learning-based channel selection strategies for high-throughput invasive BCIs.
• These methods enhance decoding performance and reduce computational and transmission burdens in BCIs.
• The study paves the way for more efficient and practical future BCI applications.