Test-retest reliability of dynamic causal modeling for fMRI

Area of Science:

• Neuroimaging analysis
• Computational neuroscience
• Brain connectivity research

Background:

• Dynamic Causal Modeling (DCM) is a Bayesian framework for inferring effective connectivity from neuroimaging data.
• Previous studies have investigated DCM's validity, but systematic examination of its parameter estimate reliability across sessions is less common.
• Assessing test-retest reliability is crucial for the clinical and research applications of DCM in functional Magnetic Resonance Imaging (fMRI).

Purpose of the Study:

• To systematically compare the test-retest reliability of different Dynamic Causal Modeling (DCM) implementations for fMRI data.
• To investigate the influence of software versions and prior distributions on the reliability of DCM parameter estimates.
• To evaluate the reliability of DCM in a challenging scenario with complex models and limited data points.

Main Methods:

• Functional Magnetic Resonance Imaging (fMRI) data were acquired from 35 subjects performing a motor task across two sessions, one month apart.
• Dynamic Causal Modeling (DCM) was applied to motor regions using classical DCM (cDCM) in SPM5 and a newer version (DCM10) in SPM8.
• Test-retest reliability was quantified using the intra-class correlation coefficient (ICC) for model evidence and parameter estimates.

Main Results:

• Classical DCM (cDCM) in SPM5 demonstrated high test-retest reliability for model evidence (ICC=0.94) and parameter estimates (median ICC=0.47).
• The newer DCM10 version in SPM8 showed notably reduced test-retest reliability.
• Using cDCM priors within DCM10 restored and even improved reliability compared to cDCM, suggesting the crucial role of prior distributions and regularization.

Conclusions:

• The reliability of Dynamic Causal Modeling (DCM) for fMRI effective connectivity is sensitive to software versions and prior distributions.
• Tighter regularization from cDCM priors appears to enhance reliability by reducing the likelihood of local extrema in the objective function.
• Future developments focusing on global optimization and empirical Bayesian procedures are needed to overcome software-dependency and improve DCM reliability.