15600
Novel Network Estimation Tools Extract Common ASD Features from Abide Dataset

Thursday, May 15, 2014
Atrium Ballroom (Marriott Marquis Atlanta)
S. N. Tomson1,2,3, M. Narayan4, G. I. Allen5,6, S. Y. Bookheimer1,3 and M. Dapretto1,2, (1)Psychiatry and Biobehavioral Sciences, UCLA, Los Angeles, CA, (2)Ahmanson-Lovelace Brain Mapping Center, UCLA, Los Angeles, CA, (3)Center for Cognitive Neuroscience, UCLA, Los Angeles, CA, (4)Electrical and Computer Engineering, Rice University, Houston, TX, (5)Statistics, Rice University, Houston, TX, (6)Jan and Dan Duncan Neurological Research Institute, Houston, TX
Background:   Functional connectivity studies have recently gained momentum in autism spectrum disorder (ASD) research, but the phenotypic diversity and inherent inter-subject variability in ASD populations have produced widely varying results in network metrics. This incongruity might be solved by exploring two primary methodological challenges: what is the most reliable way to measure functional connections, and how do we quantify differences between group-level networks with hundreds of nodes and thousands of edges?  We seek answers to these methodological questions in ASD and neurotypical populations, with the goal of working toward a diagnostic biomarker. 

Objectives: The goal of this work is to develop new statistical models to address the heterogeneity in high dimensional ASD neuroimaging data sets.  We propose novel methods of quantifying resting state networks in functional MRI data, and we compare our results to existing findings in the ASD literature.  

Methods: To address this challenge, we studied resting state fMRI data from the Autism Brain Imaging Data Exchange (ABIDE).  All individuals with ASD (N=35, Controls N=43) had an ASD diagnosis confirmed using the Autism Diagnostic Observation Schedule (ADOS). We employed Gaussian Markov Networks to estimate direct connections between nodes using partial correlation (Smith et al. 2002).  Partial correlations offer estimates of "direct" functional connections, eliminating correlations from indirect influences. To determine the sparsity of the network, we adapted a statistical technique called stability selection, which enforces the sparsity level that retains the most stable edges (Liu et al. 2010). We used block-bootstrapping and resampling to compare the networks, then performed a resampling and randomized regularization procedure to determine how often individual edges were present in each group. We used Storey’s method (Storey 2002) to estimate the false discovery rate, and we report here 7 of the most strongly differential edges of the 235 edge level hypotheses tested. 

Results:   Our results support three common trends in the ASD literature: decreased inter-hemispheric communication, reduced long-range connectivity, and fewer connections with the inferior frontal and fusiform gyri in ASD.  These findings were derived using statistical techniques specifically designed for high dimensional data to 1) account for inter-subject variability and 2) appropriately address multiple testing corrections.  Our methods are now freely available online as the Markov Network Toolbox (MoNeT) for functional connectivity estimation.

Conclusions:   We suggest novel methods for estimation of functional connectivity networks, and we provide evidence that these methods produce results in line with prior findings in the ASD literature.  Future goals include analyzing several ABIDE datasets and comparing across samples to identify common network features in ASD.