The effect of preprocessing in dynamic functional network connectivity used to classify mild traumatic brain injury

Abstract

IntroductionDynamic functional network connectivity (dFNC), derived from magnetic resonance imaging (fMRI), is an important technique in the search for biomarkers of brain diseases such as mild traumatic brain injury (mTBI). At the individual level, mTBI can affect cognitive functions and change personality traits. Previous research aimed at detecting significant changes in the dFNC of mTBI subjects. However, one of the main concerns in dFNC analysis is the appropriateness of methods used to correct for subject movement. In this work, we focus on the effect that rearranging movement correction at different points of the processing pipeline has in dFNC analysis utilizing mTBI data.MethodsThe sample cohort consists of 50 mTBI patients and matched healthy controls. A 5‐min resting‐state run was completed by each participant. Data were preprocessed using different pipeline alternatives varying with the place where motion‐related variance was removed. In all pipelines, group‐independent component analysis (gICA) followed by dFNC analysis was performed. Additional tests were performed varying the detection of temporal spikes, the number of gICA components, and the sliding‐window size. A linear support vector machine was used to test how each pipeline affects classification accuracy.ResultsResults suggest that correction for motion variance before spatial smoothing, but leaving correction for spiky time courses after gICA produced the best mean classification performance. The number of gICA components and the sliding‐window size were also important in determining classification performance. Variance in spikes correction affected some pipelines more than others with fewer significant differences than the other parameters.ConclusionThe sequence of preprocessing steps motion regression, smoothing, gICA, and despiking produced data most suitable for differentiating mTBI from healthy subjects. However, the selection of optimal preprocessing parameters strongly affected the final results.

Changing brain connectivity dynamics: From early childhood to adulthood

“…However, multiple studies have demonstrated that various aspects of connectivity between different parts of the brain can change within a single scanning session (Calhoun, Miller, Pearlson, & Adali, ; Chang & Glover, ; Hutchison, Womelsdorf, Gati, Everling, & Menon, ; Preti, Bolton, Van, & Ville, ; Sakoglu et al, ). Various studies have now shown that time‐varying estimated within short scan sessions is highly replicable (Abrol et al, ) and also predictive of individual subject brain disorders (Rashid et al, ; Vergara, Mayer, Damaraju, & Calhoun, ). Of the studies that have used dynamic connectivity to examine brain maturation, one (Qin et al, ) showed that an individual's chronological age can be predicted accurately using brain dynamics variability.…”

On the nature of time-varying functional connectivity in resting fMRI

“…The most widely used approach in this class estimates pairwise correlations within a sliding window, resulting in time-resolved correlation matrices (one per window) (Hutchison, et al 2013;Sakoglu, et al 2010). There are many variations on this theme, including the type of window used (square (Allen, et al 2014), tapered (Allen, et al 2014), or exponentially decaying (Lindquist, et al 2014)), the flexibility of the window (fixed (Allen, et al 2014) or adaptive (Lindquist, et al 2014;Yaesoubi, et al 2015)), as well as the length of the window (Leonardi and Van De Ville 2015;Liegeois, et al 2016;Sakoglu, et al 2010;Vergara, et al 2017;Zalesky and Breakspear 2015). Other (windowless) methods estimate connectivity without assuming locality of the neighboring timepoints (Yaesoubi, et al 2018), or utilize time-frequency methods to estimate instantaneous connectivity using phase synchrony (Chang and Glover 2010;Pedersen, et al 2018;Yaesoubi, et al 2015).…”

“…In addition, if overlapping windows are used, an autocorrelation (beyond that already present due to the smoothness of the BOLD signal) is induced in the estimated TVC values, which can make changes in connectivity appear artificially smooth (Lindquist, et al 2014). Recent work (Sakoglu, et al 2010;Vergara, et al in press;Vergara, et al 2017) suggests that the optimal window length to minimize these concerns may be shorter than the ~60 seconds which has been previously recommended (Leonardi and Van De Ville 2015;Zalesky and Breakspear 2015), and one can consider the choice of window size to be a tunable filter which can be optimized based on the question of interest (Lindquist, et al 2014;Vergara, et al in press).…”

“…It has recently been observed that although nuisance correction occurs prior to TVC estimation, correlations with nuisance regressors can reemerge once TVC has been calculated (Nalci, et al 2018). For sliding-window analyses, given a window of length , it has been recommended to remove frequency components below 1/ (Leonardi and Van De Ville 2015) (although recent work suggests shorter windows may provide improved accuracy when using TVC estimates to predict phenotypic characteristics (Vergara, et al 2017;Zalesky and Breakspear 2015)). Finally, there is evidence that time series autocorrelation may influence TVC estimates (Lehmann, et al 2017), and that removing the effects of autocorrelation through pre-whitening the data may decrease sampling variability (Liegeois, et al 2017).…”

Kernel machine tests of association between brain networks and phenotypes

“…The 492 difference between simulation and real datasets could be due to a variety of factors, 493 either working in isolation or compounded on one another. Several recently-published 494 studies [8] [29] [30] have noted that choice of pre-processing pipeline can impact the 495 results of an inferential analysis involving graph theoretic measures, especially in resting 496 state fMRI. We have not studied the impact of different parameters within the same 497 pre-processing pipeline nor the impact of an entirely different manner of pre-processing 498 on the RPD.…”