Summary:

Experimental Design:

The study used two independent datasets to test repeatability:

Dataset 1 - Bangor University:

• 42 healthy participants (20 males, mean age 29.6; 22 females, mean age 35.8)
• 5-minute resting state scan with eyes open
• No specific instructions except to remain still

Dataset 2 - NITRC 1000 Functional Connectomes Project:

• 44 healthy participants ages 20-30 (24 males, mean age 23.2; 20 females, mean age 24.6)
• 10-minute resting state scan with eyes open
• Used first 150 time points after excluding first 10 volumes (to match Bangor dataset)

MRI Acquisition:

Bangor:

• 3T Philips Achieva scanner
• Single-shot EPI: TR/TE = 2000/30ms
• 35 contiguous axial slices, 3mm isotropic resolution
• 150 volumes acquired

NITRC:

• 3T Siemens Magnetom Trio Tim scanner
• Single-shot EPI: TR/TE = 2500/30ms
• 38 interleaved axial slices, 3mm isotropic resolution
• 260 volumes acquired (first 10 discarded, next 150 used)

Analysis Methods:

1. Preprocessing (SPM8):

• Slice timing correction
• Realignment to first volume
• Coregistration of anatomical to mean functional
• Segmentation using 'New Segment' method
• DARTEL normalization (group-specific template for each dataset)
• Smoothing
• No motion/physiological regression (unnecessary for ICA as noise separates into components)

2. Constrained Independent Component Analysis (GIFT toolbox):

• Novel approach: Used spatial templates to extract specific networks of interest
• Three templates supplied: DMN, SN, and CEN based on Resting State Network Templates
• Semiblind ICA method incorporating prior spatial information
• Closeness measure = 0.08 (default minimum value)
• Extracted 3 components (L=3) representing the three networks
• ICA time courses used as input for DCM

3. Dynamic Causal Modeling (DCM12 in SPM12b):

• Three competing models tested for each subject:
• Model 1: DMN modulates connections between SN and CEN
• Model 2: SN modulates connections between DMN and CEN (winning model)
• Model 3: CEN modulates connections between DMN and SN
• Model specifications:
• All models had fully connected intrinsic connections
• Nonlinear modulation to test how one network influences connections between the other two
• Stochastic DCM (nonlinear stochastic differential equations)
• No external input specified (resting state)

4. Bayesian Model Selection (BMS):

• Random effects analysis at group level
• Computed expected posterior probability and exceedance probability
• Protected exceedance probability (Rigoux et al., 2014) - probability that model is more frequent than others above chance
• Bayesian parameter averaging for winning model

5. Between-Dataset Comparisons:

• Independent samples t-tests comparing ICA components between datasets
• Threshold: p < 0.05 (FWE), extent threshold = 50 voxels
• Assessed reproducibility across independent samples

Key Methodological Innovations:

1. Combined constrained ICA with DCM to examine network switching
2. Used entire network time courses (from ICA) rather than single ROIs
3. Applied nonlinear stochastic DCM to resting state data
4. Validated findings across two completely independent datasets
5. Avoided bias from ROI selection by using data-driven network extraction

Results Verification:

• Correlation between ICA-derived components and template masks ranged from 32.6% to 58.7%
• Model 2 (SN modulation) won in both datasets
• Bangor: protected exceedance probability = 0.3749
• NITRC: protected exceedance probability = 0.9252

This methodology successfully replicated Sridharan et al. (2008) findings using a different analytical approach (DCM vs. Granger causality) and demonstrated reproducibility across independent datasets.