Hierarchical vector auto-regressive models and their applications to multi-subject effective connectivity

Area of Science:

• Neuroscience
• Computational Neuroscience
• Biostatistics

Background:

• Standard Vector Auto-Regressive (VAR) models are foundational for brain network connectivity analysis using brain regions of interest (ROIs).
• Limitations of standard VAR models include a high number of parameters, leading to estimation problems, and failure to account for inter-subject variability in connectivity.
• These limitations hinder accurate investigation of effective connectivity in neuroimaging studies.

Purpose of the Study:

• To develop a generalized Vector Auto-Regressive (VAR) model that overcomes the limitations of standard VAR approaches.
• To propose a Bayesian hierarchical framework to manage high-dimensional parameter spaces and account for within-subject and between-subject variations.
• To investigate differences in effective brain connectivity between healthy controls and stroke patients during a hand grasp experiment.

Main Methods:

• Developed a novel Bayesian hierarchical framework for Vector Auto-Regressive (VAR) models.
• The framework incorporates prior distributions that yield estimates equivalent to penalized least squares with an elastic net penalty.
• Applied the generalized VAR model to functional neuroimaging data from a hand grasp experiment.

Main Results:

• The proposed Bayesian hierarchical VAR model effectively addresses the high dimensionality and parameter estimation challenges inherent in standard VAR models.
• The model successfully accounts for both temporal correlations within subjects and variations in connectivity structure across subjects.
• Differences in effective connectivity were identified between healthy controls and stroke patients with residual motor deficits.

Conclusions:

• The novel Bayesian hierarchical VAR framework provides a robust method for analyzing brain network connectivity, overcoming limitations of traditional VAR models.
• This approach enhances the investigation of effective connectivity in neuroimaging, particularly in the presence of high dimensionality and inter-subject variability.
• The model is effective in identifying group differences in brain connectivity, as demonstrated in the comparison between healthy controls and stroke patients.