Functional MRI connectivity of children with autism and low verbal and cognitive performance

Functional connectivity in AS

Multiple researchers have shown altered functional connectivity (both overconnectivity and underconnectivity) in adolescents, adults, and some children with HVCP. A review of this research [8] reports most commonly decreased long-range functional connectivity during various cognitive tasks as well as no-task or “resting state” conditions. However, studies in younger children with autism have also shown brain hyperconnectivity, possibly suggesting developmental evolution of brain connectivity with age or development [9–11] No prior connectivity studies have included comparisons between LVCP, HVCP and neurotypical (NT) groups, although some suggest a continuum of functional connectivity associated with symptom severity exists within HVCP and NT samples [12].

Recent attention has focused on connectivity within default and salience brain networks as indicators of autism that can predict prognosis. The default network, including medial prefrontal, precuneus, inferior and medial temporal, and temporoparietal junction regions, is associated with internal narrative and stimuli [13] while the salience network involves anterior cingulate and frontoinsular regions and is associated with response to novel stimuli [14]. Several studies reported increased functional connectivity in autism in some neural subsystems [4, 11, 15–21], although some are in discrete neural subsystems such as inhibitory connections with corticostriatal connectivity. These findings may represent decreased inhibition in autism [15, 22–24] or may represent aggregate effects of large networks due to increased local connectivity [25] rather than long-range connectivity between discrete regions [11].

Some differences in connectivity abnormalities in autism across the literature may be explained by methodological variations across experimental conditions [26]. Nonetheless, multisite data sharing such as the National Database for Autism Research (NDAR) and Autism Brain Imaging Data Exchange (ABIDE) initiatives [27, 28] has shown convergent results from 1000+ participants, demonstrating predominantly decreased interhemispheric and within-network functional connectivity in adolescents and adults, particularly involving default mode and salience networks [27, 29, 30], increased corticostriatal connectivity and variable short-range connectivity on the scale of millimeters [27].

An attractive unifying hypothesis is that increased synchrony across different networks (as a result of impaired network segregation) and decreased connectivity within excitatory connections from distant regions within networks (representing impaired network integration) may arise from abnormal excitatory/inhibitory balance within cortical circuits [10]. This could result in altered network integration and segregation that impairs the brain’s ability to effectively integrate information from disparate brain regions within cohesive networks [31]. These symptoms may be seen behaviorally in impaired social abilities and increased restricted and repetitive behaviors [10, 32].

Individuals with LVCP have been systematically excluded from fMRI studies because of logistical limitations. No prior studies report how brain functional connectivity may be altered in LVCP, an important cohort where treatment decisions (including early intervention) are particularly impactful and could be informed by neurophysiological data. In particular, it is not well understood whether LVCP brain pathophysiology is similar to, but more severe than what is seen in HVCP, or whether brain processes are distinctly different.

A couple of recent studies have included samples with IQ score ranges that include some individuals in what we have considered the LVCP range [33, 34]. The focus of these studies has not been on the LVCP population results, however. For example, Reiter et al. [34] analyzed samples taken from existing data, including data from ABIDE sites. Within this sample, a lowerperformance group with autism included a mean Full Scale IQ score (77 ± 6) matched to a very high-performance group with autism (mean IQ 123 ± 8). Their results between these two autism groups showed significant underconnectivity within the default mode network and the visual ventral stream. Their high-performance group differed from typically developing controls in decreased anticorrelations among default mode, salience, and task-positive regions.

We hypothesized that behavioral and anxiety-reduction techniques may facilitate collection of functional MRI data in LVCP individuals to better inform the study of brain differences and similarities across the autism spectrum. We looked at a wide range of ages (ages 7-17) with LVCP, alongside HVCP and neurotypical (NT) children for comparison. Since resources such as a behavior analyst, mock scanner, and motion sensors may not be universally available, we also adapted the Nordahl et al. procedures to provide moderate levels of preparation and support that may be accessible to most centers. Our goal was to scan 80% of LVCP participants successfully, with a minimum of 200 motion-free volumes defining a usable scan.

Our second hypothesis about possible differences between LVCP, HVCP, and NT populations is guided somewhat by existing data collected from HVCP samples, but since so little is known about LVCP individuals, we used a discovery approach without restriction to specific brain regions.

Participants

Participants responded to social media and email invitations to local autism groups. Families completed a phone intake to predict comfort level with scanning procedures (see Additional file 1: Table S1). Specifically, we invited potential participants who are typically able to tolerate dental visits and haircuts, although two participants with a history of needing sedation at the dentist also attempted scans. MRI safety checklists were completed in advance of arrival for both parents and participants. Ten families completed intake but did not attempt scanning, because parents predicted their children would not be able to hold still for more than a few minutes in the scanning environment.

We recruited a total of 62 children and adolescents (ages 7–17). Twenty of these were NT children with no history of psychological concern or diagnosis. The other 42 had been previously diagnosed with AS. We verified diagnosis based on observations from the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2) administered by research-reliable psychologists.

Preparation and support procedures

Our procedures are adaptable to most research and clinical settings. We utilized video modeling, an evidence-based practice for increasing skills in individuals with AS across levels of verbal or cognitive ability. Details of video modeling and other preparation and support procedures are shown in Additional file 1: Table S2. Participants were instructed to watch the video and listen to recordings of scanner sounds at least 3–5 times at home, before coming to the MRI facility. All participants watched the video again in the waiting room prior to giving assent. See Additional file 1: Table S3 for average reported viewing/listening times for video and audio files.

Image acquisition

We mitigated effects of participant head motion by employing a high temporal resolution scanning protocol involving multiband echoplanar fMRI imaging (Siemens Trio 3T MRI scanner, 32 channel Siemens Head Coil, TR = 800 ms, TE = 33 ms, multiband factor = 8, flip angle = 52°, 2 × 2 × 2-mm resolution, whole brain coverage 72 slices, 1240 volumes per subject divided into two 8.5-min sequences). Structural MRI imaging consisted of 3D T1-weighted magnetization prepared rapid gradient echo imaging (MPRAGE) acquisition (TR = 20 ms, TE = 4.92 ms, flip angle = 25°, 1 × 1 × 1-mm resolution). All MPRAGE images were manually inspected prior to analysis, and two subjects’ data were discarded due to unacceptable motion artifacts on structural imaging that precluded accurate coregistration with functional imaging data. Using a relatively high temporal resolution allowed better temporal control of motion-free data segments that could be used for analysis and using relatively long acquisition periods (2 × 8.5 min = 17 min total) also facilitated identification of motion-free segments.

Functional MRI imaging was performed during presentation of audio-visual stimuli with low cognitive load to help mitigate head motion and anxiety during scan sessions and reduce drowsiness that might affect functional connectivity results. Stimuli consisted of the Inscapes movie paradigm, which depicts abstract shapes and slowly moving artificial scenes without social references or any narrative story and set to a soothing audio track. Previous results have shown that reproducible functional connectivity results similar to the standard resting state connectivity may be obtained using the Inscapes paradigm. In addition, head motion and drowsiness are decreased using Inscapes.

A sequential process for preparing raw imaging data for analysis was followed prior to data analysis. The sequence is shown in Additional file 1: Table S4 and consisted of motion realignment, coregistration to MPRAGE image, segmentation of MPRAGE image, and normalization of MPRAGE image to MNI template in SPM 12 software (Wellcome Trust, London) for MATLAB (Natick, MA). Subsequently, voxelwise nuisance regression included 6 subject motion parameters, degraded white matter, degraded CSF, and soft tissues of the face time series with bandpass filtering from 0.001 to 0.1 Hz. Degraded white matter and CSF time series were obtained by retaining all voxels at least one voxel removed from the boundaries of subject-specific restriction masks for white matter and CSF components from the segmented MPRAGE image. Because the use of global signal regression remains controversial in functional connectivity analysis (i.e., [41, 42]) we performed analysis both with and without the use of global signal regression. An additional processing stream also included the global signal as a regressor, which was obtained from the mean BOLD signal from all in-brain voxels determined by the union of tpm.nii images for gray matter, white matter, and CSF distributed with the SPM 12 software package.

For functional connectivity analyses, time series were extracted from concatenated volumes for each subject after removing frames before and after a head motion of greater than 0.2 mm, from 333 regions of interest (ROIs) comprising a parcellation of cortical gray matter from a published network parcellation of the brain [41]. This parcellation, http://www.nil.wustl.edu/labs/petersen/Resources.html, includes 333 Gy matter regions covering the cortex and 286 nodes ascribed to a functional brain network (47 regions ascribed to “none” including regions in temporopolar and orbitofrontal cortex, detailed in Parcels.xlsx at the above resource). For functional network analysis, we used 8 networks combining the twelve Gordon et al. [43] networks as follows: auditory, salience (Gordon et al. cingulo-opercular and salience networks), frontoparietal (Gordon et al. frontoparietal and cinguloparietal networks), default (Gordon et al. default and retrosplenial temporal networks), dorsal attention, ventral attention, sensorimotor (Gordon et al. sensorimotor hand and sensorimotor mouth networks), and visual, as previously described. Two levels of analysis were performed, one at the 8-network level and one at the 333 ROI level, to balance multiple comparisons and granularity of spatial information. Functional connectivity was estimated as the Pearson correlation coefficient between mean time series from each ROI, with connectivity estimates Fisher-transformed prior to group analyses. Network-level time series were estimated as the mean of all ROI time series within each network, which were Fisher-transformed prior to group analyses. Left-right homotopic connectivity was estimated for each ROI by the pairwise functional connectivity between each region and the corresponding region in the opposite hemisphere using the centroid closest to the mirror image of the region’s centroid.

Statistical analyses were corrected for multiple comparisons using false discovery rate correction across all ROI or network pairs used for respective analyses. Significant results were evaluated for relationship to age, sex, and number of motion-free volumes using a general linear model with age and motion-free volumes as continuous variables and sex as a discrete variable.

Head motion

As anticipated, participants in the LVCP group showed greater head motion during scanning than the HVCP or NT participants (3-group ANOVA: p = 0.024, mean number of motion-free volumes: LVCP 630 +/− 341; HVCP 854 +/− 244; NT 884 +/− 286). Mean framewise displacement prior to volume censoring differed significantly for the LVCP group (p = 0.00002, ANOVA) with mean framewise displacement: LVCP 0.31, HVCP 0.15, typical 0.13. These results are shown in Fig. 1, along with summary plots showing the relationship of head motion with age (r = 0.26, p = 0.047), ADOS-2 comparison severity score (r = 0.014, p = 0.93), and full-scale IQ (r = 0.24, p = 0.064). Two NT participants and one LVCP participant were able to tolerate only one 8.5-min BOLD sequence during scanning, and one NT and one LVCP participant had partial second scans. All participants were able to achieve at least 200 volumes of motion-free BOLD scanning. In particularly high motion subjects in our sample, head movements were interspersed relatively uniformly throughout the acquisition, with short epochs of motion-free frames. For example, in the most extreme case of a subject with 236 motion-free frames, the longest 2 epochs of unusable data were 37 and 28 s in duration with greater than 80% of unusable epochs less than 10-s duration.

In previous analyses, randomly censoring (scrubbing) greater than 50% of the volumes of a multiband acquisition sequence had negligible effects on reproducibility of functional connectivity results, although truncating the scan to shorter duration caused decreased reproducibility, indicating that the most important factor in single-subject reliability is the aggregate time in the scanner, allowing sampling of more brain microstates [44]. Since the underlying signals comprising functional connectivity measurements are slow, typically less than 0.1 Hz, interspersed measurements throughout the time in the scanner may allow meaningful sampling of these slow time-varying signals and estimation of functional connectivity measurements.

Within-network functional connectivity

To assess differences in functional connectivity within intrinsic connectivity networks, we examined the set of pairwise functional connectivity measurements between ROIs within each of the eight functional networks studied. For each participant, we calculated the mean functional connectivity between ROIs for each of these eight networks and used a three-group ANOVA to assess for group differences, with age, sex, and number of motion-free volumes as covariates. Significant between-group differences were observed for four of eight networks, shown in Fig. 2. Specifically, decreased within-network connectivity was observed (LVCP < HVCP) in participants for the default network (F = 3.45, p = 0.039), salience network (F = 3.93, p = 0.026), auditory network (F = 4.69, p = 0.014), and frontoparietal network (F = 3.83, p = 0.028). In this sample, neurotypical individuals showed intermediate within-network connectivity between LVCP and HVCP cohorts, without significant differences from either group.

Between-network functional connectivity

Mean time series for eight functional networks were obtained in each participant, and between-network functional connectivity (i.e., synchrony of the mean time series for each pair of networks) was evaluated for differences between LVCP and NT and between LVCP and HVCP participants. Results are shown in Fig. 3a. Connectivity was higher for LVCP participants than for NT participants between the default network and the dorsal attention network (two-tailed t test, t = 3.74, p = 0.0007), and between the default network and frontoparietal network (two-tailed t test, t = 3.67, p = 0.0008), with false discovery rate (FDR) q < 0.05 across all network pairs. These differences could not be explained by age, sex, or head motion in a general linear model including group, and these 3 covariates and showing persistent increased connectivity between default network and dorsal attention network (F = 5.40, p = 0.0075) and default network and frontoparietal network (F = 6.28, p = 0.0037). Significant differences in between-network connectivity were not observed between HVCP and LVCP cohorts.

Given the extensive literature characterizing functional connectivity abnormalities in the default network in autism, we specifically evaluated functional connectivity between this network and each of the 333 cortical ROIs analyzed, shown in Fig. 3b. Significantly higher functional connectivity was observed in LVCP compared to NT participants between the default network and three regions after false discovery rate multiple comparison correction. These regions were in the right temporo-occipital (two ROIs: MNI x = 47, y = − 52, z = − 12; x = 57, y = − 54, z = − 1) and right frontoopercular (MNI: x = 47, y = 8, z = 19) areas.

Homotopic functional connectivity

Prior research on individuals with autism has also demonstrated decreased connectivity between left-right homotopic region pairs [27, 45], so we analyzed specifically whether similar connectivity differences were observed in our LVCP sample. Across the 333 ROIs compared to their interhemispheric homologs, we observed that 49 regions (FDR q < 0.05) exhibited decreased interhemispheric connectivity relative to HVCP participants, and 28 regions (FDR q < 0.05) exhibited decreased interhemispheric connectivity relative to NT participants, as rendered in Fig. 4. Widespread decreases in interhemispheric connectivity across the brain are also illustrated in histograms of T-statistics between LVCP and HVCP groups and between LVCP and NT groups. Significant differences in homotopic connectivity were not observed between HVCP and NT groups.

GSR results

Data from the GSR analyses show that there is a characteristic pattern of atypical connectivity in autism that is modulated by function, as shown in Fig. 5. In these analyses, individuals with autism show decreased within-network connectivity (especially homotopic and within the default and salience networks) but increased connectivity between the default and attentional networks (salience, dorsal attention, frontoparietal). Individuals with very low cognitive function when compared to those with high function show predominantly decreased connectivity across the brain (both within-network and between-network connections). When GSR is applied, which effectively normalizes the median connection to be zero, the effect of generalized decreased connectivity is removed, and the same pattern is seen for LVCP vs. HVCP as for LVCP vs. neurotypical individuals. In other words, if you adjust for the overall higher connectivity in HVCP individuals, you see the same pattern between LVCP and HVCP, suggesting that the “signature” of abnormal connectivity in autism is preserved and is worse in the LVCP group, implying this signature is more pronounced with greater disease severity.

Effects of IQ on connectivity

For LVCP, we observed increased brain synchrony throughout corticocortical connections to be associated with lower full-scale IQ. Figure 6a shows Pearson correlation coefficients between full-scale IQ and functional connectivity between each pair of cortical ROIs for the LVCP group only. Almost all pairwise connections show higher functional connectivity for participants with lower IQ. These relationships could not be explained by head motion, as there was no relationship between number of motion-free volumes and full-scale IQ (r = − 0.061, p = 0.82) within the LVCP participants.

To explore these effects across the range of IQ performance in our sample, we examined between-group differences for two of the results with largest effect size by averaging functional connectivity for each subject for all pairs of ROIs in these two sets of connections: decreased within-network connectivity in the default network (Fig. 6b) and increased between-network connectivity between the default and dorsal attention networks (Fig. 6c),. Here, we observe the approximation of a U-shaped curve between IQ and connectivity: IQ is negatively correlated with connectivity for the LVCP group (especially true for the lowest IQ scores). However, these are positively correlated for HVCP and NT groups.

Limitations

The ultimate goal of brain function research in autism is to discover predictors of individual performance to inform intervention. Nonetheless, considerable heterogeneity across the range of social, communicative, and intellectual functioning complicates the search for reliable individual difference factors. With our small sample, we are not able to determine predictors of individual performance with a high degree of confidence, and it is possible that the reason why LVCP vs. neurotypical between-network connectivity differences were observed while LVCP-HVCP differences were not is a combination of the small sample size, relatively higher effect size of connectivity differences in LVCP individuals, and greater heterogeneity of the HVCP sample.

Nevertheless, our study represents proof of concept that scanning LVCP children is possible and can add critical understanding of brain function in autism, although data loss from movement should be anticipated in LVCP groups. Studies involving the reproducibility of functional connectivity reassure that concatenated motion-free epochs of high temporal resolution fMRI data produce similar results to sustained datasets for an individual [44]. Nevertheless, effects of small head movements on functional connectivity data are complex, and our results should be interpreted in the context of data obtained from other methodologies such as sedated diffusion tensor imaging, optical, and electrophysiological techniques. The wide age range in our study could also be problematic in combination with the small sample size because connectivity likely changes across development [58]. A larger multisite study may improve sample size and allow for improved modeling of brain development associated with age.

Four individuals with extremely low functional language did not successfully scan. Scanning with individuals who had a history of difficulty with the dentist or haircuts was not generally attempted because of low ability to hold still and to avoid introducing aversion to the MRI environment (potentially affecting future health care) without the availability of more extensive supports. Thus, a degree of selection bias was present. More time in a mock scanner environment, with repeated attempts and behavioral support, have been shown to be effective [6, 7], but repeated trials for acclimatization and behavioral expertise may not be universally available. Some of our procedures, i.e., the multiband scan protocol and the active noise-canceling headphones, are also not universally available, although the scan sequences can be run on many machines and the headphones represent a one-time investment with large cost-benefit ratio.