Quantifying functional connectivity in multi-subject fMRI data using component models.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Brain Connectivity Analysis

Background:

• Functional magnetic resonance imaging (fMRI) is crucial for mapping brain functional connectivity.
• Analyzing high-dimensional fMRI data presents challenges in identifying stable, generalizable group-level connectivity patterns.
• Component models offer interpretable subspace representations of functional connectivity.

Purpose of the Study:

• To compare the effectiveness of various component models for representing multi-subject functional brain networks.
• To introduce a cross-validation framework for evaluating model generalization in predicting voxel-wise covariance.
• To determine the optimal component model for robust functional connectivity analysis in fMRI.

Main Methods:

• Developed and applied a flexible cross-validation approach to assess model generalization.
• Compared a range of component models with varying flexibility on simulated and experimental resting-state fMRI data.
• Evaluated models based on predictability, robustness, subject variability characterization, and parameter interpretability.

Main Results:

• Highly flexible subject-specific and overly constrained average models showed poor predictive performance.
• Models that account for subject variability within a common spatial subspace demonstrated superior generalization.
• Spatial Independent Component Analysis (sICA) yielded interpretable patterns, while PARAFAC2 offered greater stability in connectivity relationships.

Conclusions:

• The proposed evaluation framework quantitatively assesses component model performance for fMRI data.
• Optimal functional connectivity models balance individual differences within a shared representational space.
• Findings highlight critical differences in predictability, robustness, and interpretability among subspace models.