High-dimensional multivariate autoregressive model estimation of human electrophysiological data using fMRI priors

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Multivariate autoregressive (MVAR) models are crucial for assessing causal interactions in brain networks.
• Estimating MVAR models for high-dimensional electrophysiological data is data-intensive, limiting applications in large-scale brain studies.
• Previous methods focused on coefficient selection to reduce data requirements for MVAR estimation.

Purpose of the Study:

• To develop a more data-efficient method for MVAR model estimation in high-dimensional brain recordings.
• To incorporate prior neuroimaging information into MVAR estimation to improve accuracy and reduce data demands.
• To facilitate effective connectivity analyses over short timescales for understanding brain function.

Main Methods:

• Proposed a weighted group least absolute shrinkage and selection operator (LASSO) regularization strategy.
• Incorporated prior information, such as resting-state functional connectivity from fMRI, into MVAR model estimation.
• Validated the method using simulation studies with physiologically realistic MVAR models from intracranial electroencephalography (iEEG) data.

Main Results:

• The proposed method reduces data requirements by a factor of two compared to existing group LASSO methods.
• Achieved more parsimonious and accurate MVAR models.
• Demonstrated robustness to variations in prior information and iEEG data acquisition conditions (e.g., different sleep stages).

Conclusions:

• The novel weighted group LASSO approach enhances MVAR model estimation efficiency and accuracy.
• This method significantly lowers data requirements for analyzing brain network causality.
• Enables precise effective connectivity analysis over short time scales, aiding the study of perception and cognition during rapid behavioral state transitions.