Abnormal Variability and Distribution of Functional Maps in Autism: An fMRI Study of Visuomotor Learning

The study subjects were eight male autistic patients (ages 15–41 years) and eight gender-matched healthy comparison subjects (ages 21–43 years). The patients met the DSM-IV criteria for autism and the criteria for a diagnosis of autism according to the Childhood Autism Rating Scale (42) and the Autism Diagnostic Interview—Revised (43), except for patient 6, whose scores fell below the cutoff on the Childhood Autism Rating Scale but were otherwise consistent with a diagnosis of autism (Table 1). None of the autistic patients had fragile X syndrome (as determined by DNA or chromosomal analyses), had used psychotropic medication, or had a history of seizures or any additional psychiatric or neurologic diagnoses. All autistic patients were nonretarded (i.e., full-scale IQ >70), as assessed by either the Wechsler Intelligence Scale for Children—Revised (44) (patients 1 and 4) or the WAIS–R (all other patients). Neuroradiological examination showed no abnormalities in any of the patients besides borderline findings that are often associated with a diagnosis of autism (such as reduced cerebellar or parietal volume [10, 45]).

Comparison subjects were free of any history of developmental, psychiatric, or neurologic disorders. The autistic and healthy comparison groups were matched for age, sex, and handedness. Patient 3, who was ambidextrous, was matched with a left-handed comparison subject. We analyzed data for complete groups, including left-handed subjects, who performed finger movements with the left hand. Motor control conditions, which were also performed with the dominant hand, were expected to remove lateralizing effects related to side of execution. This expectation was empirically confirmed (see the Results section).

We refrained from trying to match the autism and comparison groups for IQ. Intelligence scales tap into multiple cognitive domains that are known to be impaired in autism. Since these impairments are probably an integral component of the disorder (46, 47), they should not be factored out. There appear to be only two alternatives to this approach, both of which are less desirable: 1) the inclusion of comparison groups of (borderline) retarded subjects, which is problematic given that retardation may involve diverse forms of neural pathology, and, thus, retarded subjects do not constitute “normal comparison subjects” for psychiatric populations; and 2) the exclusive study of very high-functioning autistic patients with average IQ, which may not yield results that are representative of more typical autistic populations.

The study was approved by the institutional review boards of the Children’s Hospital and the University of California, San Diego. Written informed consent was obtained from each subject, according to the procedures outlined in the Declaration of Helsinki. The 15-year-old autistic patient gave additional written informed assent.

Images were acquired on a General Electric Signa 1.5-T scanner (GE Medical Systems, Milwaukee) by using a custom-made head gradient coil. Sagittal and axial localizer scans were used to select sagittal slices for echoplanar image acquisition, ensuring complete coverage of the head. After manual shimming for reduction of magnetic field inhomogeneities, echoplanar images were acquired with a single-shot gradient-recalled pulse sequence (interleaved slice acquisition; TR=2.5 seconds; TE=40 msec; flip angle=90°). The field of view was 24 cm, and 19 sagittal slices were acquired with a thickness of 7 mm (1 mm gap) and a 64×64 matrix. For each subject, the time series contained 98 echoplanar images. Phase maps were acquired for correction of echoplanar image distortions. A high-resolution structural volume was acquired in the same session by using a three-dimensional magnetization prepared rapid gradient echo pulse sequence (TR=30 msec; TE=5 msec; flip angle=45°; matrix=256×256×128; field of view=24 cm; slice thickness=1.2 mm).

Subjects viewed a screen through an individually adjusted mirror inside the head coil. Visual stimuli were displayed on a screen approximately 12 feet from the subject’s head. Control and experimental (or task) conditions were alternated every 40 seconds in a block design (CECECE) for a total of six 40-second blocks. Each block started with the visual instruction “Press,” signaling a change in condition. Subjects viewed a diagram of a hand (roughly corresponding to the position of their own dominant hand on a four-button press device) with a blue dot appearing every 550 msec on one of the four fingers (excluding the thumb). Subjects were instructed to press the corresponding button as soon as the dot was detected. All subjects performed the movements with their preferred hand (right hand in ambidextrous patient 3). The tasks were briefly practiced immediately before the fMRI session, with digit sequences different from those used during scanning.

An unwarping algorithm (48) was applied to correct for echoplanar image distortions due to magnetic field inhomogeneities by using phase maps acquired for each session. The first two time points of each series, which are typically affected by signal instability, were discarded. All image volumes were intensity normalized, motion corrected (49, 50), spatially normalized (49), and smoothed with a Gaussian kernel of 6×6×6 mm (full-width half-maximum).

Head movement prior to correction was measured by using a center-of-intensity procedure. In all subjects and on all axes, detected motion was less than 1 mm. Groupwise means for head movement were between 0.23 and 0.49 mm for all axes and both experiments. There were no significant group differences for any axis or for mean motion across all three axes (all p≥0.14, two-tailed t test). The surprisingly limited motion in autistic patients is consistent with our experience from previous studies (25, 29) that high-functioning autistic patients are often extremely cooperative after instructions have been carefully explained.

Hemodynamic changes associated with the alternation of task and control periods (“activations”) were statistically analyzed by means of regression against a hemodynamic ideal (51). We used a smoothed boxcar shifted by 5 seconds (to correct for hemodynamic latency) as the hemodynamic ideal. The main effects of task (i.e., fit of observed signal intensity changes with the hemodynamic model) were examined within each group. Using a factorial design, we also tested for task-by-group interactions in each experiment (cf. reference 52, p. 618). All z maps were thresholded at p≤0.05, with Bonferroni correction for multiple comparisons (based on the number of resolution elements).

Two types of analyses assessed individual variability of activation foci within each group for the most consistent regions of activation (premotor cortex [Brodmann’s area 6] and superior parietal cortex [area 7] in both hemispheres). First, we calculated the distance in stereotactic space between the strongest groupwise activation peak in a given region and the closest peak found in each individual of the respective group. For example, in experiment 1 the strongest peak in the right premotor cortex (lateral area 6) in the autism group occurred at coordinates x=42, y=–3, z=48 in stereotactic space. In patient 1, the closest significant peak occurred at x=30, y=3, z=51, resulting in a distance of 13.75 mm, calculated as follows:

In fMRI, z scores tend to vary markedly across individuals, and in a few subjects the closest peak was slightly below the Bonferroni-corrected significance threshold (p≤0.05). In these cases, we applied a slightly more liberal threshold of p≤0.0005 (uncorrected). This procedure appeared adequate because it was applied to regions of expected activation. There were no significant group differences in maximum z scores for individual subjects. This finding of generally normal blood-oxygen-level-dependent (BOLD) effects is consistent with previous studies of autistic patients (24, 25, 29).

The distance from groupwise activation peaks may underestimate the scatter of activations in autistic patients, which is expected on the basis of previous studies (25, 28, 29). Even if peaks were randomly scattered across a given region, a few would occur close to a groupwise peak by chance alone. We therefore examined spatial variability in a second way, in terms of the overall group variance of distances from the groupwise peak on each axis.

All subjects were task compliant throughout the study, but response times and accuracy varied widely, especially in the autism group. The mean number of errors per block (button presses that did not correspond to the target sequence) was 12.2 in the autism group (SD=8.4) and 3.4 in the comparison group (SD=4.3). The mean response time was 574.1 msec in the autism group (SD=179.9) and 513.9 msec in the comparison group (SD=81.0). Group differences were significant for the number of errors (t=2.63, df=14, p<0.03) but not for response time (t=0.85, df=13, p=0.41). In the control condition, there was no significant difference in the mean number of index finger presses per block (autism group: mean=53.4, SD=24.1; comparison group: mean=59.1, SD=3.0; t=0.66, df=7, p=0.53). The mean number of errors decreased significantly from trial 1 to trial 10 within task blocks in both groups (autism group: p=0.004; comparison group: p=0.01 [paired one-tailed t tests]). Significant response time decreases were also seen in both groups (autism group: p<0.05; comparison group: p=0.003 [paired one-tailed t tests]).

This experiment included a more complex control condition that was more tightly matched to the experimental condition in terms of complexity of motor execution. Nonetheless, the number of errors was significantly lower in both groups for the control condition (regular sequence) compared to the task condition (sequence learning). There were also significant performance differences between groups, both for the control and the task conditions, with greater numbers of errors in the autism group. In the learning condition, the mean number of errors per block (button presses not corresponding to the target sequence) was 11.1 (SD=9.6) in the autism group and 2.4 (SD=3.3) in the comparison group. The mean response time was 492.9 msec (SD=78.7) in the autism group and 501.1 msec (SD=68.3) in the comparison group. Group differences were significant for the number of errors (t=2.42, df=9, p<0.04) but not for response time (t=0.21, df=12, p=0.84). As in experiment 1, decreases in response time from trial 1 to trial 10 within each task block occurred in both groups (p<0.001, paired one-tailed t tests). Error means also decreased, but these differences did not reach significance (autism group: p=0.18; comparison group: p=0.08 [paired one-tailed t tests]). Mean numbers of errors and mean response times were slightly lower in both groups for experiment 2, compared to experiment 1, but these differences were not significant (all p≥0.33, paired two-tailed t tests).

We performed additional analyses using data from right-handed subjects (N=5 per group). As Figure 2 shows, task effects were highly similar for the full groups and right-handed subgroups. Handedness is therefore unlikely to account for the task and group effects described earlier.

Group differences in performance constituted a further possible confound. We therefore performed all analyses using data from subjects who were approximately matched on performance, removing the two highest-performing comparison subjects and the two lowest-performing autistic patients (autistic patients 1 and 2 in experiment 1 and autistic patients 1 and 3 in experiment 2). These subgroups did not significantly differ in the number of digit sequence errors (p>0.21 in both experiments, two-tailed t tests) and in response time (p>0.60, two-tailed t tests). Figure 3 shows that the main findings for the frontal and parietal lobes for task-by-group interactions in the full groups and in the performance-matched subgroups were comparable.

As expected, spatial variability (distance between a groupwise peak in a region of interest and the closest significant peak in individual subjects) was consistently higher in the autism group than in the comparison group (Figure 4). Although the in-plane resolution in our study was high (3.75 mm), the effective slice thickness of 8 mm may have limited the detection of spatial distances. However, mean distances per region in the autism group were between 12 and 21 mm and were thus well above the spatial resolution in this study. Across both experiments and the four regions of major activation (left and right premotor and superior parietal cortex), the difference between the comparison group and the autistic group was highly significant (t=4.08, df=107, p<0.001). The difference between groups was also significant when experiment 1 (t=3.94, df=40, p<0.001) and experiment 2 (t=2.09, df=56, p<0.05) were considered separately (for single-region comparisons, see Figure 4). Greater variability of activation loci was also reflected in the greater variance of stereotactic coordinates in the autism group for each of the four regions (Figure 5).

The finding of activation abnormalities in the frontal and parietal lobes in autism is not unexpected. Parietal involvement in autism is suggested by cholinergic abnormalities (63) and by association of parietal hypoplasia (45) with visuospatial attention deficits (40, 64, 65). Our task paradigm required visual attention to a spatially displayed digit sequence. Reduced activity in superior parietal loci that were activated in normal adults could be related to impairments of visuospatial attention in autism. However, a deficit interpretation alone would not account for enhanced activation in more inferior and posterior loci in the autism group (see later discussion).

Frontal involvement in autism has been hypothesized on the basis of neuropsychological evidence of “executive” deficits (47, 66–68), cytoarchitectonic abnormalities (69), neuroimaging studies showing reduced blood flow (26, 70, 71) and serotonin metabolism (15), and abnormal event-related potentials at frontal sites for attentional tasks (41, 72). Again, however, such evidence of frontal impairment in autism does not readily account for our finding of greater-than-normal prefrontal involvement in motor learning.

In each of the few published functional mapping studies of autism and Asperger’s disorder (20–23, 25, 29, 65, 73), significant differences from normal hemodynamic responses to a variety of tasks have been observed. Such abnormalities support the expectation that autistic pathogenesis is associated with neurofunctional disturbances. However, ascertaining neurological abnormality in a disorder that by consensus is neurally based is not truly enlightening. Some studies have demonstrated reduced or absent activation in regions that are functionally well characterized in normal brains, such as the fusiform gyrus for face perception (24, 25) or the amygdala for emotional processing (21). Even though such findings may be linked to respective deficits in autism, the view that recruitment of differently localized cortical tissue would necessarily reflect functional impairment remains an assumption. Given our insufficient understanding of neurofunctional organization in autism and the known multipotential qualities of developing cortical tissue (see later discussion), this assumption is speculative. In addition, the vast array of sensorimotor and cognitive disturbances observed in autism makes it unlikely that this disorder can be neurofunctionally described in terms of a localized or “modular” deficit.

In view of the pervasive developmental nature of the disorder, the question of abnormal neurofunctional maps observed in autistic adults must be approached developmentally: How do functional maps differentiate in normal development, and how do developmental disturbances in autism interfere with this normal differentiation? Since fMRI studies of autistic patients at the time of initial pathogenesis (prenatal or early postnatal) are impractical, developmental considerations can apply only post hoc. Autism is a behaviorally defined psychiatric disorder, and diagnostic criteria may not correspond to a unique developmental pathology. On the contrary, the diversity of findings in structural (5, 74) and functional imaging (19) and in genetic linkage studies (75, 76) suggests that multiple etiological pathways result in phenotypes captured under the diagnostic umbrella term of autism. In this context, it becomes crucial for neuroimaging studies to examine data on an individual level. To our knowledge, only three studies have attempted this approach.

A study of face perception (25) showed that absence of normal activation (in the fusiform gyrus) on groupwise analyses does not imply absence of activation in individual autistic patients. Instead, the majority of patients in this study showed activation in individually varying loci around the fusiform gyrus. Our previous study of index finger movement (29) demonstrated abnormal individual variability of activations in autism for simple motor control, with individual maps showing scattered activation beyond the perirolandic cortex. Very similar effects were also observed in the cerebellum in a separate study of simple thumb movement (28). Among the three previous studies examining fMRI findings on a patient-by-patient basis, there was thus complete convergence indicating individually variable scatter of activation beyond the foci seen in normal subjects. In the present study, we again found overall significantly greater spatial variability of activation foci in the autism group, compared to the healthy subjects (Figure 5 and Figure 6), which supports our hypothesis that such variability may be a general characteristic of neurofunctional organization in autism and may directly reflect developmental disturbances.

Despite important genetic influences on brain morphology (77) and cognitive abilities (78), normal neocortical differentiation into functionally specialized regions is not strictly predetermined genetically (79, 80). Crossmodal plasticity of developing cortical tissue has been demonstrated in animal models (81, 82) and in human imaging studies (83, 84). Animal studies have further established that thalamocortical afferents play an important role in such differentiation (81), as well as in lesion-induced cortical plasticity (85). Intact thalamocortical afferents are thus a prerequisite for the normal development of cerebral cortical functional maps. There is some, as yet limited, evidence for thalamic abnormalities in autism (15, 26, 86–89). It is known that the thalamus participates in extensive cerebello-thalamo-cortical pathways reaching almost all cerebral cortical regions, including association cortex (90). Among the many sites of neurological abnormality proposed in autism research, the cerebellum has been identified through convergent evidence, including findings of postmortem (17), volumetric (10, 91–93), neurochemical (15, 16, 94), behavioral (95), and activation studies (26, 27, 65). Postmortem histology suggests that the most prominent cerebellar abnormality is a reduction in Purkinje cell numbers (17, 96). Fatemi and colleagues (94) observed reduced levels of reelin and Bcl-2 protein in autistic postmortem cerebella. These proteins play important roles in migration and cortical lamination (reelin) and apoptosis (Bcl-2). The postmortem findings from adult brains do not establish whether similar regulatory protein reduction could explain developmental abnormalities in the autistic cerebellum. However, an absence of “empty baskets” suggests that Purkinje cell reduction in adults reflects maldevelopment rather than post hoc loss (17, 97). Growth interaction between basket cells and Purkinje cells occurs before the differentiation of Purkinje cells is completed (around postnatal day 8 in the rat [98], roughly corresponding to the end of the second trimester in human gestation [99]). Inhibitory synapses of Purkinje cells in deep cerebellar nuclei probably form around the beginning of the third trimester in humans (99, 100). In the thalamus, afferents from deep cerebellar nuclei are present at birth in the rat, i.e., preceding the events described earlier (101). Early Purkinje cell loss in autism is thus likely to affect developing cerebello-thalamo-cortical connectivity and cerebral cortical functional differentiation, which may explain abnormally scattered fMRI activation patterns, as observed here.

Representational scatter in autism (as depicted in Figure 6) may imply phenomena analogous to those of “crowding” suggested for lesion-induced interhemispheric reorganization (102, 103). Currently limited fMRI evidence suggests that ontogenetically earlier functions (e.g., simple motor control, visual perception) invade brain areas normally involved in more complex multimodal processing. The opposite is not observed. For instance, we did not see greater-than-normal activity in the primary motor cortex associated with visuomotor coordination. Instead, scatter was directed toward the prefrontal cortices, i.e., the relatively late-maturing cortices that normally assume higher “executive” functions. Our hypothesis implies that due to impaired cerebral cortical differentiation (caused at least partly by disturbances in thalamocortical afferents), early-developing sensorimotor functions require a larger cortical processing territory. Such territory will subsequently be less available to later-developing complex multimodal processing (e.g., language, executive functions). This hypothesis of hierarchical crowding thus implies that more elementary functions impinge on more complex ones in the developing autistic brain.

This scenario is not a fully comprehensive model of autistic neuropathology. Relatively elementary functions (such as motor dexterity, auditory perception, spatial attention) are often impaired in autism, whereas some high-functioning patients, including those with Asperger’s disorder, perform well in higher cognitive domains. The hypothesis of hierarchical crowding does not address potential effects of compensatory neuroplasticity, known from the literature on patients with early brain damage (104, 105). Furthermore, given the plethora of pathological findings in autism, additional etiological mechanisms besides early Purkinje cell reduction and hypothesized cerebello-thalamo-cortical disturbances—such as diffuse growth disturbances (91, 106)—are likely to play important roles.

It should be noted that the proposed model has been prompted by imaging findings in rather small groups of high-functioning autistic patients, and replication will be required. Nonetheless, our model generates a number of hypotheses that may inform future fMRI work. To examine these hypotheses, it will be important to go beyond the customary procedure of groupwise analyses and examine the autistic brain on an individual level. First, our model implies predictions for fMRI studies on elementary and complex cognitive functions in various domains and modalities. Second, the model would predict functional impairment of the thalami beyond the currently limited evidence. Third, it is predicted that axonal pathways connecting cerebellar deep nuclei, the thalamus, and the cerebral cortex are affected.