Systematically comparing properties of local dynamics and pairwise coupling in the brain

1740 

Abstract Submission 

Annie Bryant¹, Kevin Aquino^1,2, Linden Parkes^3,4, Alex Fornito⁴, Ben Fulcher¹

¹The University of Sydney, Sydney, NSW, ²Brain Key Incorporated, San Francisco, CA, ³Rutgers University, Piscataway, NJ, ⁴Monash University, Clayton, VIC

Annie Bryant  
The University of Sydney
Sydney, NSW

Kevin Aquino, PhD  
The University of Sydney|Brain Key Incorporated
Sydney, NSW|San Francisco, CA

Linden Parkes  
Rutgers University|Monash University
Piscataway, NJ|Clayton, VIC

Alex Fornito  
Monash University
Clayton, VIC

Ben Fulcher, PhD  
The University of Sydney
Sydney, NSW

Dynamical structures of brain activity can be quantified from functional magnetic resonance imaging (fMRI), like local regional activity and functional connectivity (FC) between pairs of regions (Fig 1A). To date, most studies use only one of the two representations with a limited set of statistics, like the fractional amplitude of low-frequency fluctuations (fALFF) for regional dynamics and the Pearson correlation coefficient for FC [1]. Emerging work using comprehensive libraries of interdisciplinary time-series features [2,3] suggests that alternative statistics may be more suitable for a given application [3,4], though there is currently no unifying framework for comparing across multiple features and representation types simultaneously. Here, we introduce a systematic approach to quantify diverse types of local and pairwise dynamical structure from fMRI data, comparing the ability of multiple feature-based representations to capture meaningful differences in neuropsychiatric case–control datasets.

We formulated five analyses (depicted in Fig 1B) to investigate different representations of fMRI temporal structure: i) multiple time-series features measured within one region; ii) one property of local dynamics measured across the entire brain; iii) brain-wide maps of all local properties; iv) individual FC metrics across all brain pairs; and (v) each FC metric plus all brain-wide local dynamics. We investigated how well each of these representations captured salient case–control differences using cross-validated linear support vector machine (SVM) classification applied to resting-state fMRI time series from participants with schizophrenia (SCZ, N=48), bipolar I disorder (BP, N=49), attention-deficit hyperactivity disorder (ADHD, N=39), and autism spectrum disorder (ASD, N=513; Fig 1C) [5–7]. For each participant, we computed 25 univariate time-series features and 14 pairwise interaction statistics (representing subsets of broad interdisciplinary libraries [2,3]), which formed the basis for all classifiers (Fig 1D).

Strikingly, dynamical signatures of many brain regions significantly distinguished cases from controls across all four disorders, visualized as spatial maps of disorder-specific localized alterations (Fig 2A). Combining the brain-wide spatial maps from all 25 univariate time-series features generally improved classification performance relative to either representation on their own in SCZ and BP (Fig 2B), suggesting there are complex alterations to brain activity in these disorders that are incompletely captured by a singular brain region or singular time-series feature. Many individual FC metrics also significantly distinguished cases from controls, supporting the continued use of coupling statistics to quantify brain dynamics. However, classification performance was improved when FC was combined with brain-wide maps of local, region-specific dynamical properties (Fig 2C). Importantly, this demonstrates that regional activity and functional connectivity provide complementary and relevant information about disorder-specific brain dynamics. Moreover, we found that time-series features tailored to stochastic, linear, Gaussian processes are well suited for resting-state fMRI time series (Fig 2D), which are typically noisy with low temporal resolution.

Here, we present the first comprehensive, data-driven comparison of five statistical representations of dynamical structures in resting-state fMRI time series, using interdisciplinary libraries of time-series analysis methods to quantify both univariate time-series structure and pairwise coupling strengths. Our results identified multiple types of disruptions to localized activity and pairwise coupling across multiple neuropsychiatric disorders. The systematic approach presented herein is highly generalizable to myriad applications and imaging modalities, in which complex dynamical systems like the brain are analyzed.

Neurodevelopmental/ Early Life (eg. ADHD, autism)

Psychiatric (eg. Depression, Anxiety, Schizophrenia)

fMRI Connectivity and Network Modeling ¹

Methods Development

Task-Independent and Resting-State Analysis ²

Computational Neuroscience

Data analysis

FUNCTIONAL MRI

Multivariate

Open-Source Code

Psychiatric Disorders

Statistical Methods

Univariate

Other - Time-series analysis