Bayesian estimation of dynamical systems: an application to fMRI

Area of Science:

• Computational neuroscience
• Statistical modeling
• Neuroimaging analysis

Background:

• Current functional magnetic resonance imaging (fMRI) analysis models often lack explicit physical interpretations for parameters and struggle with nonlinearities.
• Classical statistical inference in fMRI relies on hypothesis testing, which may not provide the most plausible inferences about model parameters given the data.

Purpose of the Study:

• To develop a novel Bayesian method for estimating the conditional distribution of parameters in deterministic dynamical systems.
• To extend fMRI analysis by accommodating nonlinearities and enabling physically interpretable parameters.
• To facilitate more plausible inferences about model parameters using Bayesian approaches.

Main Methods:

• The proposed method employs an Expectation-Maximization (EM) algorithm combined with a Gauss-Newton search to find the maximum posterior density.
• Incorporation of prior information into the estimation process ensures robust and rapid convergence.
• The method is demonstrated using an input-state-output model of hemodynamic coupling in fMRI, analyzing the relationship between experimental factors and the Blood-Oxygen-Level-Dependent (BOLD) response.

Main Results:

• The developed method successfully estimates the posterior distribution of model parameters for deterministic dynamical systems.
• The application to fMRI data provides a generalized model that handles nonlinearities and yields parameters with explicit physical interpretations.
• The Bayesian inference framework allows for confidence intervals derived from the conditional density, offering a more nuanced understanding than traditional hypothesis testing.

Conclusions:

• This Bayesian inference approach offers a powerful and flexible framework for analyzing complex dynamical systems, particularly in neuroimaging.
• The method enhances fMRI analysis by providing physically meaningful parameters and enabling more robust statistical inference.
• The generalized model represents a significant advancement over current fMRI analysis techniques, paving the way for deeper insights into brain function.