ICA model order selection of task co-activation networks.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Cognitive Neuroscience

Background:

• Independent Component Analysis (ICA) is crucial for identifying functional brain networks during rest and tasks.
• Varied ICA dimensionality across studies complicates cross-study comparisons due to its impact on network topology.
• Applying ICA to neuroimaging databases like BrainMap enables task-based co-activation network extraction.

Purpose of the Study:

• Investigate the influence of model order (dimensionality) on functional properties of ICA-derived networks from BrainMap data.
• Determine the most informative ICA decompositions for BrainMap-based co-activation networks.
• Assess how different dimensionalities reveal large-scale networks versus their sub-network fractionation.

Main Methods:

• Utilized Independent Component Analysis (ICA) on peak activation coordinates from the BrainMap Database.
• Systematically varied the model order (dimensionality) of ICA decomposition.
• Analyzed the functional properties and organizational changes of resulting co-activation networks across dimensionalities.

Main Results:

• A dimensionality of 20 is suggested for low-order ICA to identify large-scale brain networks.
• A dimensionality of 70 provides insights into the fractionation of large-scale networks into sub-networks.
• Functional and organizational assessments were performed on visual, motor, emotion, and interoceptive networks across model orders.

Conclusions:

• Model order significantly impacts the functional properties and network topology derived from ICA on neuroimaging data.
• Specific dimensionalities (20 and 70) are recommended for distinct levels of network analysis (large-scale vs. sub-networks).
• This approach offers a quantitative method for assessing cognitive processes within brain networks using large databases.