ReDCM

ReDCM enhances estimation of effective connectivity in brain networks by applying dynamic causal modeling (DCM) with graph-theoretical search strategies to identify models within large model-spaces.

Key Features:

• Graph-theoretical search integration: Integrates graph theoretical search algorithms specifically adapted for DCM applications to explore model-spaces.
• Pre-computed model-space database: Separates model estimation from the search by using a pre-computed database containing parameters for all models within a full model-space to improve computational efficiency.
• Optimized deterministic bilinear DCM: Implements an optimized version of the deterministic bilinear DCM algorithm in an R package to accelerate model estimation.
• Tailored network search algorithms: Provides three network search algorithms modified for DCM that incorporate adjustments based on posterior parameter estimates from DCM analyses.
• Evaluation metrics: Evaluates methods using model evidence, structural similarities among models, and the number of estimations required during the search.
• Comparative performance findings: Reports that topological algorithms often outperform analytical methods in single-subject analyses by recovering common network properties, while Bayesian model reduction (BMR) remains preferred for higher-level analyses with parametric empirical Bayes.

Scientific Applications:

• Single-subject DCM analysis: Recover common network properties and perform model selection in single-subject effective connectivity studies.
• Group-level studies and model-space characterization: Systematically characterize and compare search algorithms across subjects and full model-spaces for group-level inference.
• Bayesian inference with parametric empirical Bayes: Support higher-level statistical analyses using parametric empirical Bayes with Bayesian model reduction (BMR).
• Methodological evaluation of search algorithms: Facilitate comparative assessment of topological versus analytical search methods in DCM model discovery.

Methodology:

Implements an optimized deterministic bilinear DCM algorithm in R; uses a pre-computed database of parameter estimates for all models in a full model-space to decouple estimation from search; applies three network search algorithms adapted using posterior parameter estimates; and evaluates methods via model evidence, structural similarity, and number of estimations, with Bayesian model reduction used for parametric empirical Bayes analyses.