# Test-retest reliability of dynamic functional connectivity parameters for a two-state model

## TL;DR

The study assesses how reliably dynamic brain states can be measured across different scan settings and finds that state prevalence is the most reliable parameter.

## Contribution

The work provides a systematic evaluation of dFC parameter reliability across scan lengths, data centering, and atlas resolutions.

## Key findings

- State prevalence showed the highest reliability with ICC ≈ 0.5 for uncentered data.
- Shorter scans and within-subject centering reduced reliability of dFC parameters.
- Atlas choice had no significant effect on reliability of dFC parameters.

## Abstract

Reliability of imaging parameters is of pivotal importance for further correlation analyses. Here, we investigated the test-retest reliability of two dynamic functional connectivity (dFC) brain states and related parameters for different scan length, atlases with 116 versus 442 regions, and data centering in 23 participants and reproduced the findings in 501 subjects of the Human Connectome Project. Results showed an integrated and a segregated brain state with high intraclass correlation coefficient (ICC) values of the states between sessions (0.67 ≥ ICC ≥ 0.99). The most reliable dFC parameter was state prevalence with an ICC ≈ 0.5 for ∼15 min of uncentered data, while other parameters, such as mean dwell time, were much less reliable. While shorter scans and within-subject data centering further reduce reliability, the atlas choice had no effects. Spearman’s correlations among dFC parameters strongly depend on data centering. The effect of global signal regression and a higher number of states is discussed. In conclusion, we recommend formulating hypotheses on cross-sectional differences and correlations between dFC measures of brain integration and other subject-specific measures in terms of state prevalence, especially in small-scale studies.

We investigated the reliability of popular parameters characterizing dynamic brain states in two datasets. We found high reliability of clustering results with a two-state model but only medium to low reliability of the parameters, of which state prevalence across different strategies was the most reliable.