The effect of model order selection in group PICA

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Brain Network Analysis

Background:

• Independent Component Analysis (ICA) is crucial for analyzing functional MRI (fMRI) data, particularly resting-state networks (RSNs).
• Model order selection in ICA is known to affect the characteristics of identified independent components (ICs).
• Limited understanding exists regarding how varying model orders influence RSNs derived from group probabilistic ICA (group PICA).

Purpose of the Study:

• To investigate the effect of varying model orders on the characteristics of specific RSNs identified via group PICA.
• To determine optimal model order ranges for reliable RSN detection and characterization in resting-state fMRI.

Main Methods:

• Performed group PICA on resting-state fMRI data from 55 healthy subjects.
• Utilized ICASSO for repeatability assessment and component clustering.
• Analyzed specific RSNs (VSS, DMN, S(1), S(2), M(1), striatum, preC) across model orders from 10 to 200.

Main Results:

• Lower model orders (e.g., 10) fused distinct RSNs, while higher orders resolved them into multiple components.
• Significant changes in IC volume and mean z-score were observed as a function of model order (P < 0.05).
• Model orders 70 ± 10 provided detailed RSN evaluation, whereas orders > 100 reduced repeatability without improving results.

Conclusions:

• Model order is a critical parameter influencing IC characteristics and RSN detection in group PICA.
• Model orders ≤ 20 offer a general overview, but higher orders are necessary for specific components like S(1), S(2), and striatum.
• A model order range of 70 ± 10 is recommended for detailed RSN analysis in group PICA settings.