How to Handle Multiple Resting-state Runs with XCP-D Derivatives

Dear all,

I am currently working with resting-state functional connectivity data from the Healthy Brain Network (HBN) study, using the publicly available XCP-D derivatives: https://github.com/ReproBrainChart/HBN_XCP-D.

I plan to use the Gordon 333-parcel atlas to calculate within-network and between-network resting-state functional connectivity (RSFC) measures from the 333 × 333 parcel-wise Pearson correlation matrices.

Specifically, my approach is to first apply Fisher’s r-to-z transformation to the Pearson correlation coefficients between each pair of ROIs and then average the Fisher-transformed values within or between networks to obtain summary measures of network connectivity strength.

Within the Gordon parcellation, ROIs are assigned to different functional networks (e.g., default mode, frontoparietal, dorsal attention, etc.).

  • Within-network RSFC: The average of Fisher-transformed correlations across all unique pairs of ROIs belonging to the same network.

  • Between-network RSFC: The average of Fisher-transformed correlations across all pairs of ROIs belonging to two different networks.

However, I noticed that some participants have two resting-state runs, with separate Pearson correlation matrices provided for each run. For example:

sub-XXXXX_ses-1_task-rest_run-1_space-fsLR_seg-Gordon_stat-pearsoncorrelation_relmat.tsv

sub-XXXXX_ses-1_task-rest_run-2_space-fsLR_seg-Gordon_stat-pearsoncorrelation_relmat.tsv

My question is: How should I handle these two runs when calculating network-level RSFC measures for each participant?

For participants with two runs, should I combine the two 333 × 333 correlation matrices before calculating within-network and between-network RSFC, or should I calculate the network-level RSFC measures separately for each run and then average them?

Is there a recommended approach for handling multiple runs when using XCP-D-derived connectivity matrices?

Any advice or references would be greatly appreciated!

Thank you in advance for your help!

Huiying Deng

i asked a similar question a few years ago.​ What is the best way to concatenate 2 rsfMRI volumes within the same session?

i guess there is no consensus. Computing the average FC across both runs would be fine. Though my preferred approach would be to concatenate the Z-standardized parcellated timeseries data of both runs, before computing the FC matrix

Huiying, with the matrices you already have, the two orders give exactly the same answer if you apply the Fisher transform to each edge in each run first, give both runs equal weight, and use the same edges in both. Averaging over runs and then edges just swaps the order of two sums. Averaging the raw correlations first and then applying Fisher’s transform gives a different result.

I’d calculate and keep a network summary for each run, then take their mean for the participant. That leaves you a simple check: plot run 1 against run 2 before averaging. Exclude the diagonal and check whether either run has missing edges. If the sets differ, even the two averaging orders on the transformed values can disagree. Gerard’s suggestion about concatenating standardised time series is a separate analysis that needs the time series themselves.

One small but important nuance: if you concatenate run-level time series, standardising each run separately and then computing one correlation is, for equal run lengths, equivalent to averaging the raw per-run correlations, not averaging Fisher z values. So it answers a slightly different question. With only the XCP-D matrices, per-run Fisher z and then averaging network summaries with equal weights and the same edges is the clean route. If runs differ in usable frames, consider weighting by frames or reporting run agreement.