Constructing Time-Varying Directed EEG Network by Multivariate Nonparametric Dynamical Granger Causality.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Time-varying directed electroencephalography (EEG) networks are crucial for understanding brain dynamics at millisecond resolution.
• Current methods, like adaptive directed transfer function (ADTF), rely on parametric models (e.g., MVAAR) that struggle with complex spectral features.
• This limitation hinders accurate analysis of dynamic causality in brain networks.

Purpose of the Study:

• To propose and evaluate a novel nonparametric method for constructing time-varying directed EEG networks.
• To overcome the limitations of parametric approaches in capturing complex spectral features and dynamic causality.
• To enhance the understanding of brain adaptability in information processing and inspire brain-inspired machine learning.

Main Methods:

• Developed and applied the multivariate nonparametric dynamical Granger causality (mndGC) method.
• Inferred network causality in a data-driven manner, avoiding model-dependent assumptions.
• Compared mndGC performance against the established adaptive directed transfer function (ADTF) using simulations and real motor imagery data.

Main Results:

• Simulation studies showed mndGC's superiority in noise resistance and capturing instantaneous directed network changes.
• Application to motor imagery (MI) data revealed mndGC better distinguished network characteristics between left- and right-hand MI stages.
• mndGC demonstrated enhanced ability to reveal subtle, time-varying network dynamics.

Conclusions:

• The mndGC method provides a robust, data-driven alternative for time-varying directed EEG network analysis.
• This nonparametric approach overcomes limitations of parametric methods, offering improved accuracy and insight.
• The findings extend nonparametric causality exploration and offer practical guidance for EEG network analysis.