|
Abstract:
|
Analyzing large-scale brain connections and activations has attracted much attention recently, especially utilizing large functional MRI (fMRI) data repositories. Ordinary differential equation (ODE) modeling becomes an important tool to infer causal activations and connections. However, it remains a major statistical challenge to fit ODE models to large-scale data. In this paper, we propose a sparse causal dynamic network (SCDN) model to estimate brain activations and connections simultaneously for large-scale functional magnetic resonance image (fMRI) data. Built on a popular dynamic causal model that was developed mainly for confirmatory analysis of small-scale data, we propose a novel and computationally efficient method that uses data-driven search from a huge number of ODE models. Our estimation approach is based on an optimization-based criterion which balances the data fitting errors, ODE fitting errors and lasso penalty. We propose a block coordinate descent algorithm to efficiently solve the optimization problem. The advantages of this method are illustrated using extensive simulations and data from the human connectome project (HCP).
|
