Comparison between gradients and parcellations for functional connectivity prediction of behavior

Area of Science:

• Neuroscience
• Brain Imaging
• Connectomics

Background:

• Resting-state functional connectivity (RSFC) is a key neuroimaging measure.
• Predicting behavior from RSFC is a major goal in neuroscience.
• Parcellation and gradient approaches are dominant methods for RSFC representation.

Purpose of the Study:

• To compare the efficacy of different parcellation and gradient approaches for predicting behavioral measures.
• To evaluate these methods across two large-scale datasets: the Human Connectome Project (HCP) and the Adolescent Brain Cognitive Development (ABCD) study.
• To determine the optimal number of principal gradients for behavioral prediction.

Main Methods:

• Compared group-average hard parcellations, individual-specific hard parcellations, and individual-specific soft parcellations (spatial independent component analysis with dual regression).
• Evaluated principal gradients and local gradient approaches for RSFC representation.
• Utilized two regression algorithms to predict a wide range of behavioral measures in HCP and ABCD datasets.

Main Results:

• Individual-specific hard parcellations showed superior performance in the HCP dataset.
• Principal gradients, spatial independent component analysis, and group-average hard parcellations demonstrated comparable performance across datasets.
• Local gradients consistently performed the worst across both datasets.
• The principal gradient approach requires 40-60 gradients to match parcellation performance.

Conclusions:

• The choice of RSFC representation method impacts behavioral prediction accuracy.
• Individual-specific parcellations offer strong predictive power, particularly in the HCP.
• Higher-order principal gradients are crucial for capturing behaviorally relevant information, challenging the common practice of using single gradients.