Distance-Based Metrics of Time-Varying Functional Connectivity
Poster No:
1702
Submission Type:
Abstract Submission
Authors:
Omar Shafik1, Myar Sayed1, Hussien Al Asi2, Ansam El Shikh1
Institutions:
1British University in Egypt, Cairo,Egypt, 2Cairo University, Cairo,Egypt
First Author:
Co-Author(s):
Introduction:
Sliding-window methods of estimating time-varying functional connectivity (tvFC) are limited by their recommended window size (Leonardi, 2015), and their ability to distinguish null from significant functional connectivity dynamics (Laumann, 2017; Hindriks, 2016). The sum of absolute distances between BOLD signals have been previously shown to produce similar statistics to static correlation-based FC (Minati, 2014). Here, we introduce a distance-based sliding-window method for computing tvFC estimates with high temporal resolution, and argue using complementary test statistics for the method's ability to distinguish significant from null-modeled FC estimates.
Methods:
To obtain tvFC estimates, we compute absolute distances between z-scored signals and between z-scored first-order derivatives of the signals. Then, distances are averaged for each time point, inverted, log-transformed, and sequentially sampled using a sliding-average window. We use window sizes spanning from 9 to 69 time points.
To evaluate the interpretability of tvFC estimates, we employ 3 null hypotheses: H0,1 interrogating the estimates' variability from the global (null-hypothesized) mean using their z-scores, H0,2 interrogating the variability of static estimates using spectrally constrained Gaussian noise, and H0,3 interrogating the variability of edges' variances using spectrally and covariance constrained Gaussian noise. We use empirical resting-state fMRI data from the HCP database (Van Essen, 2013). Significance levels of 0.05 were used.
To evaluate the interpretability of tvFC estimates, we employ 3 null hypotheses: H0,1 interrogating the estimates' variability from the global (null-hypothesized) mean using their z-scores, H0,2 interrogating the variability of static estimates using spectrally constrained Gaussian noise, and H0,3 interrogating the variability of edges' variances using spectrally and covariance constrained Gaussian noise. We use empirical resting-state fMRI data from the HCP database (Van Essen, 2013). Significance levels of 0.05 were used.
Results:
We found that the sampling distributions for all used window sizes are consistently normal around a mean that corresponds to computed estimates from white noise signals. We found that the global mean and variance are stationary for all analyzed time points using Augmented Dickey–Fuller test. Moreover, Inter-subject variations were found insignificant.
Although distribution tails of empirical estimates were similar to null-modeled surrogates, empirical edge statistics diverged from nulls for H0,2 and H0,3. We found that more than 35% of empirical edges are statistically significant to H0,2, and more than 14% of empirical edges are statistically significant to H0,3. In total, more than 38% of empirical edges are significant to H0,2 and H0,3. Third, we found that between 1.4% and 4% of tvFC estimates have significant z-scores to H0,1 and belong to significant edges.
Although distribution tails of empirical estimates were similar to null-modeled surrogates, empirical edge statistics diverged from nulls for H0,2 and H0,3. We found that more than 35% of empirical edges are statistically significant to H0,2, and more than 14% of empirical edges are statistically significant to H0,3. In total, more than 38% of empirical edges are significant to H0,2 and H0,3. Third, we found that between 1.4% and 4% of tvFC estimates have significant z-scores to H0,1 and belong to significant edges.
Conclusions:
The proposed method offers a sliding-window alternative that overcomes limitations of temporal resolution and interpretability. Estimates are computed with window sizes as short as 9 time points while maintaining within-session and intersubject stability of the ensemble parameters. Although empirical estimates were replicated in randomly-driven estimates, we were able to discern interesting time-resolved observations and dynamics by investigating mechanistic effects in edgewise statistics. Diverging from the null space of H0,1 and H0,2 suggests that time-resolved estimates from a specific edge are driven by an underlying mechanism. Diverging from the null space of H0,1 and H0,3 suggests that estimates from a specific edge are driven by a dynamic underlying mechanism. These interpretations do not infer anything about the weight, directness, or directionality of hypothesized connectivity. They merely suggest that an observed effect is mechanistic rather than random. All procedures are implemented in Python and will be available on GitHub.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural)
fMRI Connectivity and Network Modeling 1
Methods Development 2
Task-Independent and Resting-State Analysis
Keywords:
FUNCTIONAL MRI
Open-Source Code
Statistical Methods
Other - Functional Connectivity; Time-Varying Functional Connectivity; Dynamic Functional Connectivity
1|2Indicates the priority used for review
Provide references using author date format
Laumann, Timothy O. (2017). "On the stability of BOLD fMRI correlations." Cerebral cortex 27, no. 10, 4719-4732.
Leonardi, Nora. (2015). "On spurious and real fluctuations of dynamic functional connectivity during rest." Neuroimage 104, 430-436.
Minati, Ludovico. (2014). "Fast computation of voxel-level brain connectivity maps from resting-state functional mri using l1-norm as approximation of pearson's temporal correlation: Proof-of-concept and example vector hardware implementation." Medical Engineering & Physics 36, no. 9.
David C. Van Essen. (2013). "The WU-Minn Human Connectome Project: An overview." NeuroImage 80(2013):62-79.