Searchlight-based multi-voxel pattern analysis of fMRI by cross-validated MANOVA

Area of Science:

• Neuroimaging analysis
• Cognitive neuroscience
• Machine learning in neuroscience

Background:

• Multi-voxel pattern analysis (MVPA) is increasingly used alongside traditional univariate methods in neuroimaging.
• Current searchlight-based MVPA typically uses a 'decoding' approach measuring classifier accuracy.
• This approach has limitations in fully characterizing complex data structures.

Purpose of the Study:

• To propose a new method for searchlight-based MVPA using the multivariate general linear model.
• To introduce 'pattern distinctness' (D) as a measure to characterize multi-voxel data structure.
• To enable the application of complex factorial designs within MVPA.

Main Methods:

• Developed a measure based on the multivariate general linear model, analogous to multivariate analysis of variance (MANOVA).
• Applied cross-validation to obtain an unbiased estimate (D^) of the population value of pattern distinctness.
• Generalized pattern distinctness to accommodate arbitrary numbers of classes and parametric regressors, including complex factorial designs.

Main Results:

• The proposed D^ serves as both a test statistic and an interpretable measure of multivariate effect size.
• Pattern distinctness generalizes existing multivariate statistics like Mahalanobis distance.
• The method allows for the analysis of main effects and interactions, mirroring univariate fMRI analyses.

Conclusions:

• The novel pattern distinctness measure enhances MVPA by providing richer information and statistical power.
• This approach integrates the analytical capabilities of complex factorial designs into MVPA.
• The method also yields a measure of pattern stability, analogous to cross-decoding.