Seventy-two toddlers participated in this study: twenty-nine with autism (mean age: 29 months, range: 12 to 46), thirteen with language delay (mean age: 19 months, range: 13 to 27), and thirty typically developing controls (mean age: 28 months, range: 13 to 46). All parents provided written informed consent and were paid for their participation. The UCSD human subject research protection program approved all experimental procedures. Toddlers were scanned late at night, during natural sleep, without the use of sedation.
Toddlers were diagnosed by a clinical psychologist with over 10 years of experience in autism using the three initial modules of the Autism Diagnostic Observation Schedule; toddler, 1, or 2 and the Mullen scale for early learning (Mullen, 1995
) (Figure S6
). Autism diagnosis was based on clinical judgment and ADOS scores, with those meeting the criteria having a composite ADOS score larger than 10. In all toddlers, behavioral exams were performed within 3 months of the fMRI scan (typically they were performed within the same week). The diagnosis of toddlers with autism who were younger than 24 months at the time of the scan was confirmed at later ages (Table S2
). Toddlers in the autism group did not include individuals with PDD-NOS or other less severe forms of autism. Toddlers were diagnosed with language delay if their expressive language score was below 40. On average, the expressive language scores were almost identical across autism and control groups, indicating a similar level of language difficulty/delay. However, only toddlers with autism exhibited the social and communication difficulties assessed by the ADOS test.
Data acquisition and preprocessing
Functional and anatomical data was acquired using a GE 1.5T Signa scanner located at the UCSD Radiology Imaging Laboratory in Sorrento Valley, California. Scanning was performed with a standard GE birdcage head coil used for RF transmit and receive. BOLD contrast was obtained using a T2* - sensitive echo planar imaging sequence (repetition time of 2000-2500 ms with 150-288 time-points in length depending on the precise protocol used, 31 slices, 3×3×3 mm voxels). Anatomical volumes were acquired with a T1-weighted SPGR pulse sequence (.94×.94×1.2 mm). Data was processed with the Brain Voyager software package (R. Goebel, Brain Innovation, Maastricht, The Netherlands). Preprocessing included 3D motion correction and temporal high-pass filtering with a cutoff frequency of 6 cycles per scan. In 18 cases (10 autism, 4 control, and 4 language delay), anecdotal head movements were found and the corresponding time-points were discarded. Functional images were aligned with the anatomical volume, and transformed to the Talairach coordinate system. Data was spatially smoothed using a Gaussian kernel with 8mm width at half height.
Four different types of stimulus protocols were included in this study. All included blocks of auditory stimulation containing words, pseudo words, sentences, tones, or environmental sounds (e.g. train, phone, plane, and dog bark), which were 20-35 seconds in length and were interleaved with rest blocks of equal length. Any possible evoked responses to the stimulus were regressed out of the data as described below.
Regressing out stimulus structure and global mean
To ensure that the analyzed data contained only spontaneous cortical activity and no auditory evoked responses, we regressed out the relevant stimulus structure from each fMRI scan (Jones et al., 2009
). This process included building a general linear model (GLM) of the expected hemodynamic responses to the auditory stimuli throughout the scan. We used linear regression to estimate the response amplitude (beta value) in every voxel to each stimulus condition and extracted the residual time course in each voxel. The analyses described throughout the manuscript were performed on these residuals. In a second step, we also regressed out the “global” (average) fMRI time-course across all gray matter voxels. We assumed that this average time-course reflected spontaneous “global” fluctuations due to arousal, heart rate, and respiration (Birn et al., 2006
). This step was performed in an identical way to that described above except that here the “global” time course was used in place of the GLM with the resulting residuals describing the variability in each voxel that was not explained by the “global” time course. This analysis was performed separately for each subject.
Region of interest (ROI) definition
We defined six anatomical ROIs individually for each subject, manually selecting voxels along the following anatomical landmarks separately in each hemisphere: 1. Lateral occipital area – voxels surrounding the lateral occipital sulcus, 2. Anterior intraparietal sulcus – voxels surrounding the junction of anterior intraparietal sulcus and post-central sulcus, 3. Motor and somatosensory cortex – voxels surrounding the central sulcus around the “hand knob” landmark, 4. Superior temporal gyrus – voxels in the posterior part of the superior temporal gyrus (commonly referred to as “Wernicke’s area”), 5. Inferior frontal gyrus – voxels in the posterior part of the inferior frontal gyrus (commonly referred to as “Broca’s area”), 6. Lateral prefrontal cortex – voxels in the anterior part of the middle frontal gyrus. An example of ROI selection is described in Figure S1
. Table S1
lists the average Talairach coordinates of each ROI in each group and Figure S1
shows a comparison of ROI sizes across the groups.
Seed correlation maps
Spontaneous fMRI activity was averaged across voxels of each left hemisphere ROI to compute six seed time-courses for each subject separately. The correlation between activity in each seed and the activity of every voxel in the cortex was then computed for each subject separately. Voxel-by-voxel correlation values were averaged across subjects of each group and displayed on the inflated brain of a representative subject (). The average correlation values were thresholded at 0.3 with voxels exceeding this threshold displayed in distinct colors corresponding to each of the six seeds. A similar analysis was performed with the 7 toddlers exhibiting weakest IFG inter-hemispheric correlations (Figure S3
Voxel-by-voxel interhemispheric correlation difference maps
To compare inter-hemispheric correlation strength across the groups, we first computed, separately for each subject, the correlation between the timecourses of each left-hemisphere voxel and its corresponding contralateral right-hemisphere voxel (determined by their Talairach X coordinate). This yielded a voxel-by-voxel measure of inter-hemispheric correlation for each subject, which was compared across groups using a random-effects analysis. Correlation values were normalized using the Fisher Transform and then two tailed t-tests were used to identify voxels with statistically significant between-group differences in correlation (). Only voxel clusters exceeding 50mm3 are displayed in the statistical map, which was overlaid on the inflated anatomy of an exemplar subject.
ROI Correlation analysis
Spontaneous activity was averaged across voxels to compute a single timecourse for each ROI in each hemisphere. The correlation between timecourses of right and left ROIs was computed for each subject separately and then averaged across subjects of each group. We used both standard t-tests and randomization tests to assess the significance of differences in correlation values across the three groups (). Randomization tests were carried out by generating a distribution of correlation differences for each pair of groups, according to the null hypothesis that there was no difference between groups, by randomly assigning individuals to either subject group (i.e., randomly shuffling subject identities). This randomization was repeated 10,000 times separately for each ROI to characterize ROI-specific randomized distributions. For the correlation difference between autism and either comparison group to be considered statistically significant, it had to fall above the 95th percentile of the relevant distribution (analogous to a one tailed t-test). Note that this statistical test does not assume that data are normally distributed and is, therefore, more conservative than a standard t-test. This was evident in that significance was always weaker when assessed with the former compared with the latter. The reported weaker inter-hemispheric correlations in autism () were significant using either statistical test.
The correlation between synchronization strength and behavioral measures (i.e. Mullen or ADOS scores, ) was computed for each ROI across individuals of each group separately. The statistical significance of these correlations was also determined using both randomization and t-test analyses. Here, the behavioral measures were shuffled across subjects to determine a distribution of correlation values expected by chance. For the real correlation to be considered significant, it had to exceed the 95th percentile of this random distribution. The reported significant relationships between synchronization strength and behavioral measures were significant when assessed with either statistical test.
Trigger average analysis
To determine whether there were any residual auditory evoked responses in the analyzed ROIs, we performed a “trigger average analysis”. Segments of data corresponding to the different blocks of stimulation were extracted, aligned to stimulus onset, and averaged. There were no visible BOLD increases at stimulus onset as would be expected from a stimulus evoked response in any of the ROIs or any of the groups (Figure S5