Benchmarking brain-computer interface algorithms: Riemannian approaches vs convolutional neural networks.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) enable communication and control via neural signals.
• Comparing deep convolutional neural networks (CNNs) and Riemannian decoding methods for BCIs is crucial for advancing the field.
• Existing research lacks a comprehensive comparison across diverse BCI paradigms.

Purpose of the Study:

• To benchmark novel CNNs against state-of-the-art Riemannian decoding methods for BCIs.
• To evaluate performance across motor imagery, P300, and steady-state visual evoked potentials (SSVEP) paradigms.
• To assess effectiveness in within-session, cross-session, and cross-subject BCI settings.

Main Methods:

• Utilized MOABB (The Mother Of All BCI Benchmarks) for systematic evaluation.
• Compared EEGNet, shallow ConvNet, and deep ConvNet against established Riemannian decoding techniques.
• Analyzed decoding performance using within-session, cross-session, and cross-subject experimental designs.

Main Results:

• No significant differences in decoding performance were observed between CNNs and Riemannian methods.
• Performance was comparable across within-session, cross-session, and cross-subject analyses.
• Both approaches demonstrated effectiveness in traditional BCI paradigms.

Conclusions:

• The choice between CNNs and Riemannian methods may not critically impact BCI decoding performance in many scenarios.
• Researchers can select decoding approaches based on implementation ease, computational cost, or preference.
• Findings support flexibility and practical considerations in BCI system development.