A specific hypoactivation of right temporo-parietal junction/posterior superior temporal sulcus in response to socially awkward situations in autism

Subjects

Each subject provided informed consent to participate in this study, and the institutional review board of the California Institute of Technology approved the fMRI study protocol. A total of 45 subjects participated in the fMRI experiment, including 20 individuals with ASD and 25 NT controls. Diagnoses of an ASD were made by a clinical psychologist according to the Diagnostic and Statistical Manual of Mental Disorders (4th edition, TR), following administration of the autism diagnostic observation schedule (ADOS; Lord et al., 2001) and, when a parent was available (i.e. in all but one case), either the autism diagnostic interview-revised (ADI-R; Lord et al., 1994) or the social communication questionnaire (SCQ; Rutter et al., 2003).

Three subjects (1 ASD and 2 NT) were excluded prior to inspection of the data (1 ASD subject fell asleep during scanning; 1 NT subject quit in the middle of the experiment; 1 NT subject’s behavior was noted as strange by the experimenters). After preprocessing of the fMRI data, four more subjects’ data (2 ASD and 2 NT) were excluded based on high levels of a signal artifact metric associated with head motion (temporal derivative of variance across voxels or DVARS; Power et al., 2012); see fMRI analysis later for more details.

Thus, our fMRI analyses included data from 38 total subjects (17 individuals with ASD and 21 NT controls). Demographic and behavioral/clinical data describing the two samples are provided in Table 1. There were no statistically significant group differences in gender breakdown, age or IQ.

In addition, an independent sample of 46 NT undergraduate subjects participated in a behavioral study (described later, Quantifying Social Awkwardness) for course credit (Mean age = 20.8, s.d. = 1.6; 24 males/22 females). The institutional review board of Indiana University approved this behavioral study.

Stimuli

Each subject watched the same episode of the television show The Office (Season 1, Episode 6, ‘Hot Girl’; Production Code 1003; Copyright NBC Universal), complete and unedited with the exception of the opening title sequence being removed. This episode had a total running time of 20:52 and was divided into two halves of approximately equal length. This particular television program was chosen for several reasons. It conveys many socially awkward moments, utilizes limited cinematic features (e.g. zooming, panning and camera cuts) that can direct viewers' attention, is not accompanied by a laugh track, and features often subtle expressions of emotion compared with other shows. See Supplementary Material for a description of the television series and the specific episode.

Quantifying social awkwardness

An independent sample of 46 NT undergraduate subjects watched the same complete episode in order to provide a moment-by-moment rating of its level of social awkwardness. These subjects were instructed to push a spring-loaded joystick (Pro Logitech Extreme 3D Pro; Model # 963290-0403) forward to the extent that they deemed what was happening within the scene to be awkward. A bar on the right side of the screen provided a real-time indicator of the degree to which they were moving the joystick, corresponding to a level between ‘not at all awkward’ (baseline; resting state of joystick) to ‘extremely awkward’ (pushed all the way forward). Subjects were instructed to continue pressing the joystick forward (to whatever extent they deemed appropriate) for as long as the scene continued to be awkward; thus, their moment-to-moment positioning of the joystick was meant to reflect their perceived degree of awkwardness of the video at any given time. These awkwardness ratings were sampled at the presentation of each frame of the episode (24 samples per second); Figure 1 shows each of the 46 subjects’ responses, along with the mean awkwardness calculated across all subjects. Because the repetition time (TR) of the fMRI scanner was 2.5 s, subjects’ mean awkwardness rating over each 2.5-s interval was calculated to be used as our regressor of interest.

Most of the fMRI subjects (12 of the subjects with ASD and 19 of the NT controls) rewatched the episode after the main experiment was complete, while still in the scanner. During the second half of the episode, these subjects pressed a button when they deemed something socially awkward to be happening. These raw ratings are presented as Supplementary Material. Because we were unable to collect these data for many of the subjects, because this was a fairly coarse behavioral measure taken during a repeated viewing of the episode, and because subjects were explicitly instructed during this repeated viewing to rate the awkwardness, we do not place great emphasis on these awkwardness ratings as a faithful representation of how the fMRI subjects perceived the episode during their first, undirected viewing. Nevertheless, both the ASD and NT groups appear to show good correspondence with the ratings that were provided by the independent sample of 46 NT subjects, with no obvious group differences in the pattern of responses.

Imaging

All MRI data were acquired using a 3 T Tim Trio system (Siemens Medical Solutions) with a 32-channel phased array head coil. High-resolution T1-weighted magnetization prepared rapid acquisition gradient echo (MP-RAGE) volumetric datasets [1 mm isotropic voxel size; TR/Echo time (TE)/Inversion time (TI) = 1500/2.91/800 ms] were acquired as anatomical references for each participant. Functional MRI image acquisition consisted of two sessions of ∼10.5 min of T2*-weighted echo planar images (EPIs) (TR/TE = 2500/30 ms, flip angle = 85°, 3 mm isotropic voxels and 47 slices acquired with ascending ordering, covering the whole brain). These two sessions corresponded to the first and second half of the episode (247 and 258 volumes; see Stimuli) and data from the two halves were collapsed together for the fMRI analyses reported here. The first two volumes (5 s) in each session were discarded prior to preprocessing to minimize magnetization equilibration effects. Gradient echo field mapping data were acquired with identical geometry to the EPI data for EPI off-resonance distortion correction (TR/TE = 500/2.5 ms, 5 ms, flip angle = 60°).

fMRI preprocessing and analysis

Preprocessing and most statistical analyses were performed using fMRI Expert Analysis Tool (FEAT, v6.00) within the FMRIB Software Library (FSL; available at http://fsl.fmrib.ox.ac.uk/; Jenkinson et al., 2012). Preprocessing consisted of rigid-body motion correction, slice-timing correction, field map-based geometric distortion correction, non-brain removal, temporal smoothing (100 s high-pass filter cutoff), and spatial smoothing [5 mm full width at half maximum (FWHM) Gaussian kernel]. Blood oxygenation level dependent (BOLD) EPI data were registered in two steps: affine registration to each subject’s T1-weighted anatomical image using boundary-based registration (BBR; FMRIB’s Linear Image Registration Tool, FLIRT; http://fsl.fmrib.ox.ac.uk/flirt/) and non-linear registration of the T1-weighted anatomical image to the standard Montreal Neurological Institute (MNI152 T1-weighted; FMRIB’s Non-linear Image Registration Tool, FNIRT; http://fsl.fmrib.ox.ac.uk/fnirt/).

For each subject, we calculated the temporal derivative of variance across voxels (DVARS; Power et al., 2012) for each volume in the session as an indicator of gross artifact affecting the overall BOLD signal. Any volume with DVARS (i.e. root mean square intensity difference of volume N to volume N + 1) >.6% was effectively excluded from the analysis via a nuisance ‘spike’ regressor in the general linear model (GLM). There was no significant difference in the mean number of volumes removed for ASD ($M=7.7,SD=8.1$) vs NT control participants ($M=8.0,SD=12.4$; $t[36]=0.10,P=0.92$) nor was there was a significant group difference in mean DVARS for the remaining volumes included in the analyses (ASD: $M=33.3,SD=3.0$; NT: $M=35.1,SD=3.4$; $t[36]=1.75,P=0.09$).

A stability problem with 2 of the 32 elements of the head coil used for this study resulted in intermittent Nyquist ghost artifacts in many subjects’ scans. We created an algorithm for detecting these artifacts within each horizontal slice at each TR (i.e. each volume) based on the measured signal outside of the brain. This detection algorithm yielded a continuous output for each of ∼30 horizontal slices within each volume, generating a set of ∼30 continuous nuisance regressors that were entered into the GLM for each subject. There was no significant group difference in the mean output yielded by the Nyquist ghost detection algorithm across all brain slices ($t[36]=0.63$, P = 0.53) nor was there a significant group difference in the mean output yielded by this algorithm for each subject’s worst slice ($t[36]=1.18$, P = 0.24). Cerebrospinal fluid and white matter signals and their temporal derivatives were also included in the GLM as nuisance regressors, as were six head motion regressors (x, y and z translations and rotations) and their temporal derivatives.

We convolved our social awkwardness measure (see Quantifying Social Awkwardness) with a gamma function (6 s mean lag, 3 s half width, 0 s phase), and included this convolved, continuous regressor (along with its temporal derivative) in the GLM as our variable of interest. The GLM parameters were estimated with FILM (FMRIB’s Improved Linear Modeling; part of FSL), a method which includes a pre-whitening step. The parameter estimates of GLMs fit to each subject were the inputs for higher-level whole brain analyses.

A recent study (Dufour et al., 2013) found typical levels of activity in ASD across several brain regions often connected to social cognition, using a verbal false belief task (presented in text form or auditorily). We modeled region of interest (ROI) analyses after theirs, using a Bayesian approach that, like theirs, allowed us to report evidence in favor of no group differences where the data supported such a claim. We used the ROIs they make available (at saxelab.mit.edu/hypothesis_spaces.zip; see Supplementary Material for additional information about these ROIs), defined with respect to their large and independent dataset. This allowed us to examine the same approximate brain regions thought to be important for social cognition, and explore whether our comparatively naturalistic stimuli elicited a different pattern of brain activation with respect to this network.

Whole brain analysis: main effects of social awkwardness

Using FSL’s-mixed effects algorithm (FMRIB’s Local Analysis of Mixed Effects, 1 + 2), we thresholded voxels positively correlated with our measure of social awkwardness (at $Z>2.33,P<.01$) and applied a cluster-level significance threshold of P < 0.05 (Worsley, 2001). Collapsed across groups, we observed significant clusters of activity in several regions of the brain predicted a priori to be important for social cognition—including RTPJ, RSTS, PC and MPFC—in addition to a number of other brain regions—including orbitofrontal cortices, occipital cortex and bilateral occipital fusiform gyri (Figure 2; see Supplementary Material for a table of information about these clusters and local maxima within them). Excluding brain stem voxels, 11% of the brain was significantly correlated with the social awkwardness regressor. These clusters found to be significantly correlated with social awkwardness bore resemblance to brain areas found by the automated meta-analytic tool Neurosynth to be recruited in social tasks across 800 studies (Figure 2, bottom row; indexed with ‘social: reverse inference’ at www.neurosynth.org; Yarkoni et al., 2011).

Figure 2 (top row) shows the overlap between those clusters we found to be significantly correlated with our social awkwardness regressor via whole brain analysis and a predefined network of seven ROI masks Dufour et al. (2013) derived empirically (RTPJ $\cup$ RSTS $\cup$ LTPJ $\cup$ PC $\cup$ DMPFC $\cup$ MMPFC $\cup$ VMPFC). We hypothesized that socially awkward moments would recruit this predefined brain network; however, given the uncontrolled nature of the stimulus we did not expect that these moments of this particular television program would recruit only these regions. Unsurprisingly, the neural response was not entirely specific to these ROIs: at this statistical threshold, 75% of the voxels that were correlated with socially awkward moments fell outside of the Dufour et al. (2013) predefined network. Some voxel clusters corresponded well with these ROIs, but extended outside of them, some voxels were not contained within the Dufour et al. ROIs, but nevertheless showed excellent correspondence to ‘social’ regions culled from Neurosynth (e.g. orbitofrontal cortices), and some voxels belonged to regions entirely distinct from these regions of a priori interest (e.g. occipital fusiform gyri). Nevertheless, after accounting for volume (these ROIs comprised only 9% of total brain voxels), the Dufour et al. ROIs were correlated with social awkwardness at a rate far higher than the rest of the brain: 33% of voxels within the predefined Dufour et al. network were significantly correlated with social awkwardness with respect to the whole brain cluster analysis, compared with just 9% of voxels lying outside of this network (excluding brain stem).

In order to compare how well various masks discriminated the social awkwardness clusters—in a manner that accounts for both sensitivity and specificity, and theoretically is not dependent on the arbitrary thresholds used to generate the masks—we here employ $d\prime$. We estimate hit rate from the proportion of the total number of voxels significantly correlated with social awkwardness falling within the candidate mask. We estimate false alarm rate from the proportion of the total number of voxels not significantly correlated with social awkwardness falling within the same candidate mask. The following formula is then used to estimate $d\prime$: norminv(hit rate)–norminv(false alarm rate), where norminv is the inverse cumulative distribution function of the standard normal. Using this metric, a binary mask of the Dufour et al. ROIs discriminated voxels significantly correlated with social awkwardness ($d\prime =.83$) better than many other binary masks generated from Neurosynth queries related to social perception and cognition (‘social’: $d\prime =.65$, ‘biological’: $d\prime =.59$, ‘tom’: $d\prime =.53$, ‘intentions’: $d\prime =.52$, ‘mentalizing’: $d\prime =.49$, ‘perspective’: $d\prime =.49$, ‘gaze’: $d\prime =.25$, ‘social cognition’: $d\prime =.23$, ‘eyes’: $d\prime =-.01$, ‘mirror neuron’: $d\prime =-.38$, ‘faces’: $d\prime =-.62$).

Whole brain analysis: group differences

We also examined whether any brain region across the whole brain showed group differences in response to social awkwardness (voxel-level threshold, P < 0.01; cluster-level threshold, P < 0.05). We found one contiguous cluster of voxels (2.8 cm³ in size) spanning across RTPJ and RSTS for which there was significantly greater correlation for the NT group compared with the ASD group (Figure 3). The MNI coordinates of voxels of peak group difference within this cluster were (52 −36 2; z = 3.62), (56 −44 18; z = 3.57) and (56 −42 20; z = 3.49); its center of mass was (57 −40 14).

Sixty-seven percent of this cluster’s voxels were found within a region we predefined as RTPJ, and 22% of the voxels were found within the posterior portion of predefined RSTS. However, many of the voxels of peak difference were located very near the border between these two regions—a border which is very approximate to begin with. We therefore cannot determine with confidence whether we observed group differences at two dissociable loci, or one locus at the border of these two areas.

As an exploratory analysis, we examined whether activity in this RTPJ/RSTS cluster was correlated with autistic symptom severity among the 14 subjects with ASD for whom we were able to obtain AQ scores. We did not observe a significant correlation ($r[12]=-.045,p=.88$); however, we acknowledge that this small sample underpowers this analysis for the reliable detection of any but the strongest effects.

Across the whole brain, we found no significant group differences in the opposite (ASD > NT) direction.

ROI analysis

Given the high degree of overlap between regions we expected a priori to be particularly active at socially awkward moments, the regions that were indeed correlated with this social awkwardness regressor, and a network of several ROIs identified independently by Dufour et al. (2013) in a recent fMRI study of autism, we focused a more detailed analysis on this particular network, and limit the following analysis to this subset of brain regions.

For each subject, we extracted the peak t-stat voxel within each of the seven respective ROIs defined by Dufour et al. (2013), each representing the voxel most reliably correlated with the subject’s observation of social awkward moments. For each ROI, we then tested whether there was a group difference (ASD vs NT) in the location of this peak voxel. We calculated the Euclidean distance between the two group mean peak locations (in 3D space), and performed a Monte Carlo approximation (10⁶ samples) of a non-parametric permutation test to derive the probability of observing a group mean difference of this size or larger given random rearrangement of group labels (Table 2). Only when examining the RTPJ ROI did we observe even a marginally significant group difference (P = 0.06).

We then performed full Bayesian analyses (Gallistel, 2009, using Matlab functions provided at http://cognitivegenetic.rutgers.edu/ptn/ and uniform priors across the predefined spatial limits of each particular ROI) to weigh the evidence for (or against) a group difference in peak voxel location in the x, y or z dimensions, respectively. For most ROIs, we found moderate evidence in favor of no group difference in all spatial dimensions (Table 2). The strongest evidence we found for a group difference in location was along the y dimension in RTPJ, where there was moderate to strong evidence that ASD peak voxels within this ROI were comparatively more posterior than NT peak voxels.

For each subject and for each ROI, we then defined a sphere of voxels within a 5 mm radius of this peak location. This sphere was allowed to extend outside of the predesignated ROI, but not outside of the predefined network (i.e. RTPJ $\cup$ LTPJ $\cup$ RSTS $\cup$ PC $\cup$ DMPFC $\cup$ MMPFC $\cup$ VMPFC). We derived Akaike Information Criterion (corrected for finite sample sizes, AICc)-adjusted likelihood ratios (Burnham and Anderson, 2002, 2004) to again weigh the evidence for (or against) a group difference (ASD vs NT), this time with respect to the mean t-value within the sphere defined around the individual’s peak voxel.¹ In Table 3, we also provide conventional (two-tailed) P-values to complement these less commonly used statistics.

For most of the predefined ROIs (LTPJ, PC, DMPFC, MMPFC and VMPFC), we found weak to moderate evidence in favor of the null hypothesis (i.e. no difference between group means). The exceptions were RTPJ and RSTS, for which we found some weak evidence of a group difference. These trends are consistent with our initial findings of a group difference in a (corrected) cluster straddling across these two predefined regions, and no significant group differences elsewhere. For this analysis, we initially assumed 5 mm radius spheres surrounding peak voxels in each ROI, but as illustrated in Figure 4, the choice of assumption (i.e. radius of sphere) does not affect the pattern of results.

Functional connectivity analysis

Our initial whole brain analysis of the fMRI data (the analysis described earlier) with respect to our social awkwardness regressor yielded evidence of a group difference in an area spanning RTPJ and RSTS (Figure 3). Having isolated this region, we searched for other regions with potentially high functional connectivity to the RTPJ/RSTS cluster using a conventional whole-brain temporal correlation analysis. To achieve this, we reran the full GLM including all nuisance regressors, but replaced the convolved social awkwardness regressor and its temporal derivative with the mean activity of voxels within the RTPJ/RSTS cluster shown in Figure 3. This regressor was generated by masking each subject’s time series with the supra-threshold cluster region from the group level contrast and taking the unweighted mean across these voxels for each TR.

Collapsed across groups, a whole brain analysis (voxel-level threshold P < 0.01; cluster-level threshold P < 0.05) revealed clusters of activity significantly correlated with this seed region in each of the ROIs that were earlier found to have been correlated with observation of socially awkward moments—RTPJ, RSTS, PC and MPFC. But additionally, in this case, we observed strongly correlated activity in LTPJ and LSTS, mirroring the activity of corresponding regions in the right hemisphere (Figure 5).

We did not observe group differences in functional connectivity within any of the regions for which we had strong a priori interest (i.e. the Dufour et al. ROIs). However, a 3.4 cm³ cluster of voxels, with local maxima in left orbitofrontal cortex (z = 3.86, MNI: −36 20 −14) and left temporal pole (z = 4.68, MNI: −26 14 −36), showed greater co-activity with RTPJ/RSTS in the NT group than the ASD group. In the opposite direction (ASD > NT), a 5.4 cm³ cluster of primarily white matter voxels in the left hemisphere showed greater co-activity with RTPJ/RSTS (peak: z = 4.09, MNI: −10 −12 36), likely reflecting a false positive. Neither of these clusters of group difference overlapped with any of our predefined ROIs (RTPJ, RSTS, LTPJ, PC, DMPFC, MMPFC or VMPFC). Though we report these results for completeness, we view these results as tentative and preliminary, especially since we had no a priori hypotheses about them.