Exploring individual and group differences in latent brain networks using cross-validated simultaneous component...

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Statistical Modeling

Background:

• Component models like PCA and ICA reduce neuroimaging data into latent brain networks.
• Simultaneous estimation (tensor ICA, group ICA) is common for multi-subject data.
• Existing methods struggle with cross-validation due to model limitations.

Purpose of the Study:

• To address limitations of tensor ICA and group ICA for multi-subject component analysis.
• To propose a flexible Simultaneous Component Analysis (SCA) model hierarchy.
• To introduce cross-validation methods for robust parameter tuning and overfitting reduction.

Main Methods:

• Comparison of a hierarchy of Simultaneous Component Analysis (SCA) models.
• Implementation of the Parafac2 model, a flexible SCA variant.
• Application of cross-validation for model parameter tuning.
• Validation using simulated and real neuroimaging data.

Main Results:

• Tensor ICA is too rigid; group ICA is too flexible for reliable latent brain network identification.
• The proposed Parafac2 model uniquely identifies latent brain networks while modeling heterogeneity.
• Cross-validation reduces overfitting and enhances data interpretability.
• The approach yields credible components revealing individual and group differences.

Conclusions:

• The Parafac2 model offers a superior approach for multi-subject component analysis in neuroimaging.
• This method enhances the identification and cross-validation of latent brain networks.
• The findings support the use of flexible SCA models and cross-validation for robust neuroimaging research.