Nonlinear Structural Vector Autoregressive Models with Application to Directed Brain Networks.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Structural Equation Models (SEMs) and Vector Autoregressive Models (VARMs) are established methods for studying brain connectivity.
• SEMs focus on instantaneous effects, while VARMs emphasize time-lagged influences.
• Linear Structural Vector Autoregressive Models (SVARMs) integrate both perspectives but are limited to linear dependencies.

Purpose of the Study:

• To extend SVARMs by incorporating nonlinear dependencies using kernel methods.
• To improve the modeling of effective brain connectivity by capturing complex neuronal interactions.
• To develop an efficient method for identifying directed links in brain networks.

Main Methods:

• Kernel-based nonlinear modeling framework applied to SVARMs.
• Development of an efficient regularized estimator to exploit network sparsity.
• Data-driven approach for selecting appropriate kernels from a predefined dictionary.
• Application to electrocorticography (ECoG) data from epilepsy patients.

Main Results:

• Demonstrated capability of kernelized SVARMs to model nonlinear neuronal dependencies.
• Successful identification of previously unknown directed connections between brain regions.
• Validation of the method's efficacy on real-world ECoG data from epilepsy studies.

Conclusions:

• Kernel-based SVARMs offer a powerful extension to linear models for brain connectivity analysis.
• The proposed method effectively uncovers complex, nonlinear directed relationships in neuronal activity.
• This approach holds significant potential for advancing our understanding of brain networks, particularly in conditions like epilepsy.