Impact of methodological variables on functional connectivity findings in autism spectrum disorders

Abstract

Growing evidence suggests that Autism Spectrum Disorder (ASD) involves abnormalities of multiple functional networks. Neuroimaging studies of ASD have therefore increasingly focused on connectivity. Many functional connectivity (fcMRI) studies have reported network underconnectivity in children and adults with ASD. However, there are notable inconsistencies, with some studies reporting overconnectivity. A previous literature survey suggested that a few methodological factors play a crucial role in differential fcMRI outcomes. Using three ASD data sets (two task-related, one resting state) from 54 ASD and 51 typically developing (TD) participants (ages 9-18 years), we examined the impact of four methodological factors: type of pipeline (co-activation vs. intrinsic analysis, related to temporal filtering and removal of task-related effects), seed selection, field of view (whole brain vs. limited ROIs), and dataset. Significant effects were found for type of pipeline, field of view, and dataset. Notably, for each dataset results ranging from robust underconnectivity to robust overconnectivity were detected, depending on the type of pipeline, with intrinsic fcMRI analyses (low bandpass filter and task regressor) predominantly yielding overconnectivity in ASD, but co-activation analyses (no low bandpass filter or task removal) mostly generating underconnectivity findings. These results suggest that methodological variables have dramatic impact on group differences reported in fcMRI studies. Improved awareness of their implications appears indispensible in fcMRI studies when inferences about "underconnectivity" or "overconnectivity" in ASD are made. In the absence of a gold standard for functional connectivity, the combination of different methodological approaches promises a more comprehensive understanding of connectivity in ASD.

Keywords: autism; fMRI; functional connectivity; region of interest; resting state; task regression; temporal filtering.

Figure 1

A: Correlation matrices for dataset‐specific ROIs. Note that for all matrices time series were averaged across all voxels within each ROI. VS dataset (top): Correlation matrices for intrinsic versus co‐activation pipelines and ROIs selected based on activation in the TD group—left and right middle occipital gyrus (RMO and LMO), right superior parietal lobule (RSPL), left supplementary motor area (LSMA), and left precuneus (LP); based on activation in the ASD group—right and left superior parietal lobule (RSPL and LSPL), right and left middle occipital gyrus (RMO and LMO), and left supplementary motor area (LSMA); and based on activation for both groups combined—right superior parietal lobule (RSPL), left middle occipital gyrus (LMO), left supplementary motor area (LSMA), left precuneus (LP), and left lingual gyrus (LLG). RSVP dataset (middle): Correlation matrices for ROIs selected based on activation in the TD group—left fusiform gyrus (LFG), right lingual gyrus (RLG), right middle occipital gyrus (RMOG), left inferior frontal gyrus (LIFG), and left precentral gyrus (LPG); based on activation in the ASD group—left precuneus (LP), right middle frontal gyrus (RMFG), right and left precentral gyrus (RPG and LPG), right supramarginal gyrus (RSG); and based on activation in both groups combined—left fusiform gyrus (LFG), left middle occipital gyrus (LMOG), right precentral gyrus (RPG), left inferior frontal gyrus (LIFG), and right lingual gyrus (RLG). RS dataset (bottom): Correlation matrices for DMN ROIs—posterior cingulate cortex/precuneus (PCC), retrosplenial cortex (RET‐SP), left and right lateral parietal cortex (LLP and RLP), left and right medial prefrontal cortex (LMPF and RMPF), left and right superior frontal gyrus (LSFG and RSFG), left and right parahippocampal gyrus (LPAR‐HIP and RPAR‐HIP)—for low band‐pass filtering versus no low‐pass filtering. +/−, significantly higher/lower r value on direct between‐group comparison; white solid squares indicate significantly greater connectivity than for comparison pipeline (dashed squares represent significantly reduced connectivity in comparison matrix). B: Correlation matrices from intrinsic FC analysis for each data set and for ROIs derived from Just et al. [2004]: calcarine fissure (CALC), left dorsolateral prefrontal cortex (LDLPFC), left inferior extrastriate (LIES), left inferior frontal gyrus (LIFG), left intraparietal sulcus (LIPS), and left superior temporal gyrus (LSTG). White squares indicate ROI pairs showing significant main effect of dataset in univariate ANOVA; solid squares represents highest r value, dashed squares medium r value, dotted squares lowest r value. C: Examples of divergent between‐group effects for co‐activation versus intrinsic pipeline, for VS dataset with seed in left middle occipital gyrus (based on TD activation); for RSVP dataset with seed in left fusiform gyrus (based on TD activation); for RS dataset with seed in posterior cingulate cortex (DMN). Negative x‐coordinate indicates sagittal slice through left hemisphere.

Figure 2

Underconnectivity ratios averaged across all pipelines (as shown in Table 2) for each variable for which main effects were detected in ANOVA. Underconnectivity ratios are higher for co‐activation type of analysis and for field of view limited to activation‐derived ROIs. Differences related to data set can also be seen, that is, more overconnectivity for VS than for RSVP, with RS falling in the middle. P values are based on main effects from ANOVA. **p < 0.01; *p < 0.05. [Color figure can be viewed in the online issue, which is available at http://wileyonlinelibrary.com.]