Neural synchrony examined with magnetoencephalography (MEG) during eye gaze processing in autism spectrum disorders: preliminary findings

Eighteen participants completed the study; 10 participants with ASD (M_Age = 16.6); M_IQ = 112) and eight neurotypical controls (M_Age = 17; M_IQ = 116). The groups did not differ significantly in age (U(16) = 1.79, P = 0.76), gender (χ² = 0.11), or Full Scale IQ, with both groups generally performing in the Average to Above Average range on the Wechsler Abbreviated Scale of Intelligence (WASI; U(16) = 40, P = 0.74) [35]. See Table 1 for a review of the demographic variables. There was equal age distribution between the genders of the two groups. To ensure that potential developmental differences between the groups did not account for our effect, significant relationships between our brain region coherence values with age were inspected. Age was related to coherence values in the beta frequency band only in the right inferior frontal to right superior temporal pathway in the NT controls (r = 0.52, P = 0.02). Five of the 10 ASD participants were on psychotropic medications. Two participants were prescribed a single medication, an antidepressant (Prozac) or antipsychotic (Risperdal). Two participants were prescribed an antidepressant and psychostimulant (Prozac or Zoloft and Concerta), and one participant was prescribed three medications, an antidepressant, anxioltyic, and mood stabilizer (Cymbalta, Buspar, and Depakote).

Individuals were recruited from and underwent MEG procedures at Henry Ford Hospital (HFH). Participants were diagnosed with Pervasive Developmental Disorder (PDD) (recently revised to ASD) based on the Diagnostic and Statistical Manual of Mental Disorders-Fourth Edition-Text Revision (DSM-IV-TR) [2] diagnostic criteria. Diagnoses were confirmed with the Autism Diagnostic Interview-Revised (ADI-R) [36]. Means and standard deviations for the ADI-R domain scores include: M_Social (SD) = 19.25 (4.59); M_{Communication} (SD) = 15.50 (4.34); M_Repetitive (SD) = 6.63 (1.69). Inclusion criteria included at least low average intelligence (≥80 Full Scale IQ scores on the WASI). Exclusionary criteria included any known history of head injury with unconsciousness, epilepsy, affective or anxiety disorders. No ASD participant had a history of a genetic disorder. NT participants had no history of developmental delay, learning disorder, or ASD in a first-degree relative. All APA Ethical Guidelines were followed and Institutional Review Board approval was obtained from all institutions participating in this study.

Cortical activity was recorded using a 148 channel whole head MEG system (4D Neuroimaging, Magnes WH2500) with magnetometer type sensors. During acquisition, the data were band-pass filtered 0.1 to 100 Hz and digitally sampled at 508.63 Hz. The timing of stimuli was recorded as pulse codes (representing the type of stimulus) on a trigger channel simultaneously collected with the MEG data. In postprocessing, noise artifacts due to heart and body movement were eliminated using an independent component analysis (ICA). Singular valued decomposition was used to remove any other artifacts in the data such as mouth movements if needed. Regarding movement artifact, runs are repeated if the coil on head positions exceeds 0.5 cm, although this did not occur during acquisition. As such, we combined the runs before source reconstruction. Regarding noise reduction, 4D Neuroimaging incorporates a set of reference sensors that are used to sample the environmental magnetic fields and create a file that has a set of weights. These weights are subtracted from the data during data collection. Data were band-pass filtered from 1 to 50 Hz. The locations of events on the trigger and response channels were used to select 2-s epochs of MEG data to examine average evoked responses for the stimuli requiring a conditional button press (that is, asterisk, face, or words). All trials within each condition were averaged to determine the evoked response. All epochs had a baseline of 500 ms before stimuli onset and 1,500 ms of data after stimuli onset.

The gaze cueing paradigm used was an adaptation of a paradigm used by Pelphrey and colleagues [37]. MEG field responses to gaze cues were collected for two 14-min trials in which five task conditions were administered: direct gaze, averted gaze, and gaze cueing to peripheral stimuli (asterisk, word, or face). In the direct and averted gaze conditions, participants passively viewed a central character. There were 30 trials in the direct gaze condition and 30 each (left and right) in the averted gaze condition. In each of the three additional gaze cueing conditions, the central character engaged in a random gaze shift toward the right or left for 1 s. A target (asterisk, word, or face) then appeared at either the right or the left of the subject for 3 s. See Figure 1 for an example of one of the target stimuli. The next trial began with the character returning to a direct gaze for 2 s with no stimuli in the periphery. The location of the target stimulus was either congruent or incongruent with the direction of the character’s gaze. Sixty targets were presented in each gaze-cueing condition, including 30 congruent and 30 incongruent trials. A conditional button press during gaze cues to the peripheral stimuli conditions ensured engagement during the passive conditions. For the purposes of this analysis, only passive viewing of direct and averted gaze was considered. Results from the evoked response data requiring active responses to the conditional stimuli (asterisk, faces, words) have been previously published [38].

Synchronization of neuronal activity was quantified by calculating coherence between cortical sites from MEG imaged brain activations [39, 40]. A model of the cortical brain surface was created from an age- and gender-appropriate standard MRI. The MRI was segmented and the brain surface was represented by a cortical model of approximately 4,000 dipoles each having an x, y, and z orientation at each site. Sites were distributed to represent the same volume of cortical gray matter. This model was then morphed to fit the digitized head shape collected during the MEG acquisition. To calculate coherence [40], the MEG data were first divided into 80 segments each containing 7.5-s segments of data and cortical activity in each segment was imaged on to the MRI using the MR-FOCUSS imaging technique [41]. Using the time sequence of imaged activity, coherence between active cortical model sites was calculated for each data segment and then averaged for the completed study. In addition, for each cortical model site, connectivity was quantified by a histogram of the number of sites to which the site had the same level of coherence. Statistical analysis of cortical coherence levels (0 to 1) were used to quantify differences in network connectivity between groups. Changes in coherence and connectivity between brain regions implicated as having deviant electrophysiological activity in the ASD brain were quantified and subjected to further statistical analysis.

Power spectra for activity at all active sites were also calculated and used to quantify differences in low frequency, beta, and gamma power. For the ‘lower’ frequency band, delta, theta, and alpha bands were collapsed for comparison with the alert working brain, specifically beta and gamma frequencies. For additional details of MEG coherence imaging, see our publication [40]. A region-of-interest (ROI) tool implemented in MEG Tools was used to identify 54 regions in the brain (27 in each hemisphere). See Figure 2 for a list of the brain regions. MEG Tools uses a non-linear volumetric transformation of the patient’s brain to transform MEG coordinates to standard Talairach [42] or MNI [43] coordinates [44]. This enables the ROI tool to access an atlas of Brodmann’s area identifiers and an atlas of cortical structures [43].

For each frequency band (low, beta, and gamma) within each condition (direct and averted gaze), a t-test was conducted to assess group difference in average coherence values for each pair of brain regions (N = 1,431). A P value was produced for each region pair. The false discovery rate (FDR) was used to adjust for multiple testing. Because of the large number of tests being performed simultaneously, using a significance level of alpha = 0.05 without adjusting for multiple testing would lead to a large number of false positive results. Bonferroni adjustments for multiple comparisons aim to control the Family Wise Error Rate. If a Bonferroni correction were applied to every test, there would be only a 5% chance of at least one false positive in all of the N = 1,431 tests. Bonferroni corrections require the P value to be less than 0.05/N, where N is the number of tests. With N = 1,431, this criterion becomes especially stringent, and many true differences may be missed (false negatives). In the era of large scale testing, a less conservative approach to adjusting for multiple testing has been developed. The FDR is the proportion of tests declared significant that are actually different only due to chance (or the proportion of significant tests that are false positives). The FDR is a widely accepted, less conservative approach to adjusting for multiple testing in large scale problems. The Benjamini-Hochberg algorithm [45] was used to control the FDR at 0.10. From each t-test, a z-score was computed according to the method of Efron [46] to summarize the difference in coherence values between ASD and NT. Positive z-scores indicate higher coherence in the ASD group. A series of chi-squares were computed to determine if the number of intra-hemispheric and inter-hemispheric cortical differences within the low (0 to 15 Hz), beta (15 to 30 Hz), and low gamma (30 to 45 Hz) frequency bands were statistically different between the groups.

To examine relationships between ASD clinical symptoms as reported on the ADI-R and neural oscillatory activity (coherence), a series of Kendall Tau correlation coefficients were computed.

There were no significant between group differences noted in error rates in responding to the conditional button press during the intervening task condition with respect to accuracy (t(16) = 0.70 P = 0.51), suggesting that both groups were equally engaged in the task. Reaction times were not statistically different (t(16) = -0.11 P = 0.92). See Table 2 for the accuracy (Total Correct Responses) and reaction time values for correct responses for the NT and ASD groups.

ADI-R scores were significantly correlated with coherence values in a predictable pattern among the ASD group under both the direct and averted gaze conditions (see Table 5). During direct gaze and collapsed over all frequencies, significantly higher coherence between left parietal regions (that is, angular and supramarginal gyri) was related to a greater number of ASD symptoms as reported on the ADI-R (≤0.005) while higher coherence between bilateral frontal cortical regions was related to fewer ASD symptoms. That is, between left inferior frontal/lateral orbitofrontal and right precentral gyri (≤0.01 and ≤0.005, respectively).

A similar pattern of relationships was noted during the averted gaze condition in the low frequency band. Significantly higher coherence between unilateral and bilateral parietal regions (P ≤0.005), between left parietal and left occipital cortices (P ≤0.005) as well as between left parietal and bilateral temporal regions (P ≤0.005), and between left temporal and bilateral occipital regions (R P ≤0.005; L P ≤0.01) was related to a higher scores on the ADI-R and more severe ASD symptoms. In contrast, higher coherence between unilateral right frontal (P ≤0.001) and bilateral frontal regions (P ≤0.005) as well as between right frontal and right parietal (P ≤0.005) and right frontal and right occipital (P ≤0.005) was related to fewer ASD symptoms.

Primarily negative correlations were obtained during the averted gaze condition in the beta frequency band. Higher beta frequency band coherence between bilateral frontal regions (P ≤0.005), between unilateral right frontal regions (P ≤0.001 - right inferior and superior frontal gyri; P ≤0.01 - right inferior frontal and middle frontal), between right frontal and right temporal cortices (P ≤0.01), and unexpectedly, between bilateral occipital regions (P ≤0.01) was related to lower scores and fewer ASD symptoms in the ADI-R. Higher symptom reporting was only related to increased coherence between the right precentral and right orbitofrontal regions.

In the gamma frequency band, higher coherence between bilateral frontal regions (particularly right precentral gyri) and all other cortical regions bilaterally was related to a lower number of ASD symptoms as reported on the ADI-R. Higher gamma frequency band coherence between the right precentral gyrus and left inferior frontal (P ≤0.001), left lateral orbitofrontal (P ≤0.005), left middle frontal (P ≤0.01), and the left inferior temporal cortex (P ≤0.005) was related to lower scores and fewer symptoms on the ADI-R. Higher gamma band coherence between right lateral orbitofrontal and left superior occipital (P ≤0.005) and the left angular gyrus (P ≤0.005) was also related to fewer ASD symptoms. Finally, higher gamma band activity between the right temporal gyrus and left superior frontal gyrus (P ≤0.01) was related to lower scores on the ADI-R.

An opposite and strongly lateralized right hemispheric finding was also obtained when examining gamma activity. Higher gamma frequency band coherence between the right parietal cortex (that is, angular gyrus) and right orbitofrontal (P ≤0.005), temporal (P ≤0.005 - fusiform; P ≤0.01 - inferior temporal), and occipital regions (P ≤ .001 - lingual and cuneus; P ≤0.005 - inferior occipital) as well as between bilateral occipital cortical regions and right temporal cortex (P ≤0.005 - bilateral fusiform; P ≤0.01 - inferior temporal) was related to a greater number of ASD symptoms as reported on the ADI-R.

Recent multi-modal imaging methods have started to illuminate networks involved in direct and averted gaze in NT using, for example, combined fMRI-DTI [61]. Dynamic gaze shifts have been found to increase activation in a well-established region of the social network including the right pSTS, anterior insula, and fusiform gyrus with direct connections noted between the right pSTS and anterior insula. These regions are thought to be supported by long-range network connections that project via the superior longitudinal fasciculus and are believed to be critical for extracting social meaning of eye gaze shifts [61]. Consistent with this purported social network, compared to those with ASD, our NT participants demonstrated significantly higher coherence between the right middle and inferior frontal gyri and the right superior temporal regions within the beta band frequency known to be critical for long-range connectivity.

In contrast, our ASD participants demonstrated higher gamma power in inter-hemispheric connections between the left and right parietal lobes and intra-hemispherically between the right parietal lobe (angular gyrus) and temporal regions. The angular gyrus has been implicated in a number of processes including reading and number comprehension, numerical processing, visual attention, and social cognition. It is a cross-modal region where sensory information from the visual, auditory, and tactile senses converge allowing for a combined and integrated percept. It is essential for the manipulation of mental representations and reorienting of attention [62]. We believe that heightened connectivity in these regions without appropriate regulatory or contextual feedback from frontal regions may result in exquisite sensory sensitivity, acceleration of letter, number, and word recognition with limited comprehension or applied skills, or an over-allocation of attention to information without a clear appreciation of its relevance; a neurocognitive pattern often noted in ASD. A strengthening of connectivity between the right angular gyrus and inferior temporal regions without frontal mediation, particularly from medial prefrontal regions, may further contribute to a heightened attention to the invariant features of the face or its components without an ability to extract essential social relevance. Our findings are consistent with models of ASD proposed by Brock and colleagues [25] as well as Belmonte et al.[27] who have suggested that abnormally high levels of high-frequency neural activity and over connectivity within localized brain regions causes impaired discrimination of brain signal from background brain noise. Interestingly and unexpectedly, our ASD participants demonstrated a greater number of low frequency cross-hemispheric connections, and particularly in posterior cortical regions. The results suggest that conceptualizing ASD as a disorder of either over- or under-connectivity in long- or short-range connections may not fully capture the pathophysiology.

Approximately 200 investigations have been published proposing neuroanatomical markers of ASD, although the results have often been in conflict or unreplicated [63]. The majority of this work has been in older children and adults using structural imaging. Recently, Ecker and colleagues [64] used linear support vector machines (SVM), a machine learning method that identifies patterns in data by identifying hyperplanes that maximally differentiate categories or groups, classified 20 adults with and without ASD. They used five structural classifiers of cortical gray matter to correctly classify 85% of those with ASD, with 90% sensitivity and 85% specificity. Left hemispheric classifiers were more accurate than right, and cortical thickness was the best classifier. Hemisphere laterality is an area that remains relatively unexplored in ASD, and with their reported method the authors were unable to determine if individuals with ASD displayed a higher (lower) degree of cortical asymmetry. Similarly, a recent fMRI investigation by Dinstein and colleagues [29] of sleeping toddlers with ASD revealed significantly weaker inter-hemispheric synchronization (that is, weak ‘functional connectivity’ across the two hemispheres) in inferior frontal (IFG) and superior temporal (STG) regions for which early ‘over-lateralization’ of language function was suggested. However, again, directionality of lateralization to the left or right hemisphere could not be determined from their data. Our method allows us to provide a direct numerical comparison between pathways, both inter- and intra-hemispherically, and to examine group differences between frequency bands known to underlie short- and long-range connectivity. Given the passive nature of the task, this methodology could be easily applied to preschoolers. Only 18% of children with ASD are identified by the age of 3 years; even later for children with milder forms (average, 6.3 years) [65]; well beyond when children can benefit maximally from early intervention. This method allowed coherence to be imaged in source rather than sensor space, achieving a significant degree of disentanglement of signals. Moreover, it provided better resolution and precision of the underlying networks generating the signals since coherence in sensor space is smeared due to current spread as described by Srinivasan et al.[66]. Neural synchrony between the frontal cortex and pre- and postcentral gyrus, between frontal and superior temporal cortex in the beta band, and between angular and fusiform gyri in the gamma band were particularly discrepant between the groups.

We recognize that there are a number of weaknesses with this pilot study. First, the sample size of this pilot phase was small which limits the external validity and conclusions that can be drawn. Although the findings were quite robust, a replication of this study with a larger number of participants is necessary. Moreover, the results suggest a strong need for funding of large scale and longitudinal studies with this technology in ASD.

Second, our ASD participants remained on their psychotropic medication regimens during the study given the perceived cost-benefit of motion artifact for participants taking ADHD medications as well as the practical challenges associated with withdrawing from medications with longer half-lives. However, we recognize that when one is evaluating a possible pathophysiological mechanism in any psychiatric or neurodevelopmental disorder, a vital issue is whether the observed neural abnormality is present at the outset of the disorder and predates medication exposure and the deleterious effects of long-term illness. As noted, five of our participants were on one or more of the following medications: psychostimulants, antidpressants, anxiolytics, antipsychotics, or mood stabilizers. Recent investigations exploring alterations in oscillatory activity in populations taking medications similar to our cohort (for example, ADHD, bipolar disorder, and schizophrenia) have reported variable results with respect to medication status on oscillatory activity and the findings to date have been primarily reported in adult populations [67‐69]. Wilson and colleagues [69] recently examined time estimation and oscillatory activity using MEG in medicated and unmedicated individuals with ADHD. Relative to controls, unmedicated participants were reported to display less accurate time estimation and weaker gamma activity in frontal cortex, specifically anterior cingulate, supplemental motor areas, and the left prefrontal cortex. Following medication administration, the patients demonstrated small but significant increases in gamma-band activity across the same neural regions, which was related to improved time estimation accuracy. Exploratory analysis also revealed stronger delta activity in frontal regions in the control participants relative to those with ADHD, regardless of medication status, possibly suggesting that such alterations in delta activity may not be responsive to stimulant medications. Although we recognize it is difficult to compare across populations, recent investigations have demonstrated high co-morbidity with ADHD and ASD, suggesting similar aberrant neural substrates. We further reviewed the literature to explore possible effects of antipsychotics on oscillatory power. Minzenberg and colleagues [67] recently reported reductions in gamma oscillatory power during an executive task that was independent of medication status in first-episode schizophrenia. Fifty-three first episode patients with schizophrenia (21 without antipsychotic treatment), aged 13 to 30 years, underwent EEG while performing a cognitive control task. Both the medicated and unmedicated patient subgroups were impaired on the behavioral task and displayed lower frontal induced gamma power. In contrast, there were no significant group differences in theta power between controls or the medicated or unmedicated patient subgroups. Given the above, we might have expected the effect of psychostimulants and antipsychotics to attenuate the frontal gamma frequency band group differences noted between our groups. Özerdem et al.[68] examined alpha and beta oscillatory activity during a visual odd-ball paradigm and EEG in 10 bipolar patients before and after valproate treatment. At baseline, drug-free individuals with bipolar disorder in the hypomanic or manic phase of illness demonstrated aberrant alpha and beta oscillatory activity to visual target stimuli compared to healthy controls. Individuals with bipolar disorder displayed significantly higher beta activity in occipital cortices compared to controls, and patients were devoid of an occipital-frontal alpha dominance that was noted in the control group. Following treatment with valproate, occipital beta responses dropped to control levels, and a further and significant decrease in occipital alpha and unchanged frontal alpha response was noted in patients with bipolar disorder. The authors suggest that the latter may suggest the inhibitory effect of valproate via GABAergic inhibition. That is, inhibition of occipital alpha activity while trying to manage cortical hyperactivity may be an unwanted effect of valproate use. Similarly, if the effect of valproate is an attenuation of occipital alpha activity, the impact of this medication would have also been to decrease our current effect of increased low frequency band activity in posterior regions. However, these assumptions should be interpreted cautiously as it is important to remember that investigations examining auditory and visual evoked oscillatory activity in humans have shown that the evoked responses are topography and stimulus modality dependent and that brain structures with different resonance properties may be dependent on the stimuli.

Finally, with respect to the frequency ranges that were examined, while we did not examine activity greater than 45 Hz, we recognize that the gamma band frequency extends to approximately 100 Hz. For this pilot project, we chose to remain below 60 Hz given that group differences in gamma power are the most robust in the 40 Hz range and in order to avoid additional artifact. Despite these limitations, this pilot study provided the first cogent understanding of whole-brain patterns of coherence that underlie direct and averted eye gaze in ASD.