Coordinated multivoxel coding beyond univariate effects is not likely to be observable in fMRI data

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Data Analysis

Background:

• Simultaneous brain region recordings offer more information than isolated ones.
• Multivariate pattern analysis (MVPA) in fMRI is often interpreted as evidence for distributed neural coding.
• Univariate effects (UVE) may not capture the full picture of brain activity.

Purpose of the Study:

• To challenge the interpretation of conventional MVPA in fMRI.
• To propose that 'real' multivoxel coding should manifest as changes in higher-order statistics (e.g., second-order multivariate effects, sMVE).
• To investigate the limitations of fMRI for studying coordinated coding across voxels.

Main Methods:

• Applied MVPA to fMRI data from humans and macaques.
• Analyzed for second-order multivariate effects (sMVE) in conditions with robust MVPA findings.
• Compared fMRI results with analyses of neuronal population spiking activity.

Main Results:

• Conventional MVPA primarily acts as a sensitive decoding tool by combining many 'weak classifiers' (voxels).
• Analysis of conditions with significant MVPA did not reveal significant sMVE in humans or macaques.
• Higher-order statistics (sMVE) were observable in neuronal spiking data but unlikely to be detected in fMRI due to hemodynamic coupling and low spatial resolution.

Conclusions:

• fMRI signals have inherent limitations for studying coordinated coding across voxels.
• Significant MVPA results should be interpreted cautiously, as they may represent a collection of univariate effects rather than true distributed coding.
• The findings highlight the need for careful consideration when interpreting MVPA in neuroimaging studies.