Single-trial connectivity estimation for classification of motor imagery data

Area of Science:

• Neuroscience
• Biomedical Engineering
• Signal Processing

Background:

• Brain-computer interfaces (BCIs) commonly use electroencephalogram (EEG) band power (BP) for motor imagery (MI) classification.
• Current BCIs often overlook brain area connectivity, limiting performance.

Purpose of the Study:

• Introduce and evaluate single-trial connectivity features for enhanced MI classification in BCIs.
• Investigate the utility of vector autoregressive (VAR) models for extracting these connectivity features.

Main Methods:

• Extract single-trial connectivity estimates using VAR models applied to independent components of EEG data.
• Compare the performance of connectivity features (e.g., Directed Transfer Function - DTF) against traditional BP measures in a simulated BCI.

Main Results:

• Connectivity features, specifically DTF with full-frequency normalization, achieved classification performance comparable to BP.
• Other connectivity measures, like partial directed coherence, performed significantly worse than BP.

Conclusions:

• Single-trial motor imagery classification is feasible using connectivity measures derived from VAR models.
• These connectivity features hold promise for future BCI applications, potentially improving performance.