Single trial classification of magnetoencephalographic recordings using Granger causality

Area of Science:

• Neuroscience
• Computational Neuroscience
• Signal Processing

Background:

• Granger causality (GC) is increasingly used for analyzing dependencies in neuroimaging data.
• Existing GC frameworks often depend on strict autoregressive (AR) model assumptions, which are challenging in practice.

Purpose of the Study:

• To propose an alternative statistical methodology for GC analysis in neuroimaging data.
• To test hypotheses by classifying individual data trials rather than relying on AR model statistics.
• To evaluate the effectiveness of GC features for distinguishing experimental conditions and optimizing analysis parameters.

Main Methods:

• Developed a novel statistical approach based on classifying individual data trials.
• Utilized features derived from autoregressive (AR) and Granger causality (GC) concepts for classification.
• Applied the methodology to magnetoencephalography (MEG) data to differentiate between forced and free button presses.

Main Results:

• Demonstrated that bivariate temporal GC features can successfully classify button presses as forced or free.
• Achieved a mean classification accuracy of 79.2% across 6 subjects.
• Gained insights into AR and GC analysis by determining optimal parameter settings.

Conclusions:

• Classification using GC features is a viable approach for studying MEG signals.
• The proposed methodology offers a robust alternative for hypothesis testing in GC analysis.
• This approach provides a novel way to evaluate parameter variations in GC analysis.