Multivariate linear regression of high-dimensional fMRI data with multiple target variables

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Machine Learning in Neuroscience

Background:

• Multivariate linear regression (MLR) is a common tool for analyzing fMRI data, linking brain activity to stimuli or behavior.
• Current MLR methods may produce inaccurate interpretations by analyzing targets separately, especially when neural activity overlaps.
• This limitation hinders the precise identification of brain regions associated with specific cognitive processes.

Purpose of the Study:

• To address the limitations of standard MLR in fMRI analysis.
• To develop a novel model for accurately identifying spatial patterns of brain activation related to specific experimental targets.
• To improve the generalization and interpretability of fMRI-MLR models.

Main Methods:

• Proposed a new multivariate linear regression formulation that trains on an augmented dataset including all experimental targets.
• Incorporated interaction coefficients to disentangle specific neural effects from overlapping predictive maps.
• Validated the model using simulated fMRI data and a publicly available dataset.

Main Results:

• The proposed method accurately identifies spatial patterns associated with specific targets.
• Demonstrated high spatial sensitivity and improved generalization compared to standard MLR.
• Successfully disentangled specific neural effects from interactions with other predictive maps.

Conclusions:

• The novel MLR approach offers a more reliable method for analyzing fMRI data.
• This advancement aids in correctly interpreting the link between brain activity and cognitive/behavioral variables.
• The formulation enhances the precision of neuroimaging studies by accounting for target interactions.