Missing data estimation in fMRI dynamic causal modeling.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Cognitive Science

Background:

• Dynamic Causal Modeling (DCM) quantifies cognitive function via effective connectivity.
• Subject-specific ambiguities in active brain regions limit DCM's use in individual cognitive phenotyping.

Purpose of the Study:

• To develop a preprocessing method for handling missing brain regions in DCM.
• To enable comprehensive model comparison across subjects for individual cognitive phenotyping.

Main Methods:

• Missing brain regions were treated as missing data and time courses were estimated using zero-filling, average-filling, noise-filling, or expectation-maximization.
• The impact of these estimation methods was evaluated as a preprocessing step for DCM, analyzing effects on model evidence.
• Simulations and real data (Go/No-Go, Simon tasks) were used to assess performance.

Main Results:

• Expectation-maximization yielded the highest classification accuracy and model evidence in simulations.
• This method improved DCM analysis across various dataset sizes and model choices.
• Real-data application enabled signal computation for missing nodes, allowing model evidence calculation in 100% of subjects.

Conclusions:

• The proposed preprocessing scheme effectively handles missing brain regions in DCM.
• This approach enhances the feasibility of using single-subject DCM for individual cognitive phenotyping.