In a task demanding sustained, covert attention to lateral locations in the visual field, the autistic brain activated ventral occipital and striate regions instead of the normal network of superior parietal, middle temporal, dorsolateral prefrontal, premotor, and medial frontal cortices. In a further analysis of the effect of attention within individually mapped functional brain regions, the normal tendency to activate left ventral occipital cortex while attending to the right hemifield was reversed, and the normal tendency to activate left intraparietal sulcus while suppressing the right hemifield was augmented. In general, the autistic brain tended to be more active while attending to the left hemifield than while attending to the right. The data also suggested a possible relationship between neurophysiology and clinical behavioural features: the three subjects with clinical diagnoses of autism had particularly large reversed (i.e., left-attention) activations in left ventral occipital cortex, and the two males within this subgroup had the most augmented right-attention activations in right intraparietal sulcus.
As our experimental task has much in common with the classic oddball paradigm, results of our comparison between task and fixation may be expected to bear some similarity to the results of fMRI studies of oddball tasks. Indeed, our results show a pattern in which visual modality-specific oddball regions were significantly activated in the autism group, whilst modality-independent oddball regions were significantly activated in the normal group. In particular, the striate and extrastriate occipital activations in the autism group correspond to the modality-specific loci in fMRI studies of visual versus auditory oddball targets [91,77,52], and in the normal group the middle temporal, dorsolateral prefrontal, medial frontal, and premotor loci are a subset of the modality-independent loci common to visual oddball tasks [58,22,4] and oddball tasks in general [91,77,52]. Unlike most of these studies of oddball tasks, the current study did not find activations localised to insula, thalamus, or supramarginal gyrus, and the active regions that were identified were predominantly left-lateralised. This lack of complete overlap with previous results may stem from several methodological differences. First, our analysis compared task blocks to fixation blocks rather than target events to non-target events, and as a result was more sensitive to brain regions involved in the task in general rather than those involved specifically in evaluating targets. Second, our small number of subjects no doubt compounded the effects of spatial blurring and whole-brain spatial normalisation to decrease the sensitivity of our analysis. Third, the demands of our task differed from those of the oddball tasks used in these previous fMRI studies: in our paradigm, target stimuli differed from non-targets in colour rather than form, and in addition, a continuous stream of distractors was present at an unattended location.
In the attention comparison, our particular regions of interest were chosen in order to examine the effects of these two spatially separated streams of attended and suppressed stimuli, and the region-of-interest approach in general was applied in order to avoid the deleterious effects of blurring and normalisation - effects that are particularly problematic in some of the brain areas of interest. In ventral occipital cortex, the effect of selective attention is relatively easy to discern statistically since the sensory signals driving it are strong and the anatomical variability across subjects is low. The same cannot be said, however, for the effect in intraparietal sulcus. Within individuals, this attention effect is subtle, reaching significance in a few of the autism subjects but in none of the normal controls. This subtlety of the underlying signal within subjects is exacerbated by anatomical variation between subjects: in a full 30% of the normal population, the intraparietal sulcus follows an unpredictable, zigzag course through the parietal lobe, sending off varying numbers of small rami as it descends toward the transverse occipital sulcus [64]. Spatial averaging in coordinate systems based on global landmarks would therefore tend to eliminate localised, weak activations. In light of this high degree of normal variation in parietal anatomy, several authors [66,24,90,11] have noted the necessity of examining functional anatomy within individuals rather than in a spatial average of individual brains.
In patient populations, and particularly in autism where cerebral anatomic abnormalities have been identified, the weakening of statistical power inherent in whole-brain spatial averaging can only be magnified [75,70,63,16,1]. Given the weakness of the underlying BOLD signal and the variability of the relevant anatomical features both within and possibly between groups, our strategy of within-subjects functional mapping and individual permutation testing followed by groupwise parametric testing is very important in revealing the effect that we have demonstrated. Permutation testing is particularly powerful in situations such as this, where the number of subjects available is small and where anatomical variation demands within-subjects statistical analysis [65].
In the comparison of task to fixation, the finding of heightened ventral occipital activation and lowered prefrontal, parietal, and temporal activation in autism as compared to controls is consistent with the theme of abnormally heightened early sensory activation that has emerged from several fMRI studies of autism. In the further comparison of attention conditions within the task, the findings of generally increased activation during attention to the left hemifield and an absent effect of spatial attention in ventral visual areas are consistent with earlier quantitative electroencephalographic findings [9]. The heightened activity in intraparietal cortex is a new finding, and helps to sketch the outlines of abnormal information flow in autistic perception. We propose that this pattern of information flow is characterised by three elements: (1) hyper-arousal, that is, primary sensory processing that is abnormally intense [72,7,32,75,70] and abnormally generalised across anatomical regions and functional systems [49,50,9], and (2) impaired early selection of relevant stimuli [79,9], leading to (3) overloading of higher-order processes, such as the suppressive process identified in the current study.
What physiological causes and effects might explain and relate these three elements? A recent neuropathologic study [17,18] identified in postmortem autistic brains a reduction in the size of cortical minicolumns and an increase in cell dispersion within minicolumns. The authors speculated that a resulting increase in the total number of minicolumns would lead to overconnected and insufficiently inhibited neural networks, with consequent reductions in information content, signal-to-noise, and discriminability of competing inputs - in other words, hyper-arousal and impaired selection. This hyperconnectivity may be akin to that induced by a failure of synaptic pruning in Fragile X [42], a syndrome that shares many cognitive characteristics with autism [33]. Indeed, quantitative MRI anatomical studies have revealed a pattern of early brain overgrowth in autism [15,27,16,76,5], a finding that may be the gross manifestation of such a failure.
The flood of input generated by over-aroused, under-selective primary processing would from the earliest months of infancy overload nascent higher-order cognitive processes - processes whose development may be independently sabotaged by the same neuropathology that affects primary regions. Faced with this bottleneck in higher-order cognition, the developing and plastic brain would likely evolve a cognitive style that emphasises low-level features and eschews reliance on global patterns [43] - in other words, weak central coherence. Weak coherence in turn may impair the use of contextual information in complex perceptual and executive tasks [38], including theory-of-mind tasks [74], impede the development of joint attention and shared affect [51,73], and perturb or prevent the activity-dependent development [26,1] of specialised modules for such tasks as theory-of-mind, face processing, and language. This failure to use context and to implement a theory of mind may result in a style of unsupervised learning founded on statistical associations rather than learning that is directed by the intentionality of others [37], and to a preference for ritualised, scripted, and repeatable interactions.
We have considered autistic cognition in terms of a bottom-up developmental influence of low-level attentional processing on higher-order cognitive processing. However, top-down influences are also possible, and indeed these may seem a more parsimonious explanation of autistic deficits in late stages of information processing affecting memory, language, and reasoning [59]. In this view, hyperactivation of primary perceptual systems may arise as a consequence of a more fundamental deficit in complex information processing. Functional imaging of the adult autistic brain, though more practical than imaging of autistic children, cannot distinguish between a bottom-up developmental chain of dysfunction, a top-down chain, or multiple simultaneous loci of dysfunction.
We have interpreted our intraparietal result in terms of the active suppression of unattended stimuli. An alternative interpretation is that intraparietal activation reflects oculomotor activity in response to stimuli in the unattended hemifield. Although in the present study gross eye movements were identified and rejected, subjects still could have made subtle deviations in eye position below the resolution limit of our eye tracker. More significantly, oculomotor-related activation in intraparietal sulcus could have arisen even in the absence of actual eye movements, due to planning of unexecuted reflexive saccades [39] and/or inhibition of reflexive saccades [23]. Although recent results in non-spatial tasks suggest that the role of intraparietal sulcus is more general than simple saccade preparation [90], in paradigms that include a spatial component it is difficult to factor out the role of oculomotor planning and inhibition. This question of distractor suppression versus saccadic activity can be addressed more specifically in future studies using tasks of non-spatial selective attention.
Anxiety is always a potential confound in experimental studies of autism, since people with autism are so prone to anxiety in novel circumstances and in situations that demand rapid or accurate performance. Attempts were made to reduce levels of anxiety by offering subjects ample opportunity to practise the task before entering the scanner, and by allowing a variable rest period rather than a strictly timed one. Although this variable rest period between trials may itself have posed a confound, in our estimation the anxiety which these breaks helped to alleviate would have posed an even greater obstacle to comparison of the two groups.
Although our method of individual functional mapping solves the problem of individual variation in functional anatomy, it introduces an assumption that the regions identified in each subject group are functionally comparable. While it seems likely that regions bearing similar relationships to local landmarks and similar functional correlations with the task subserve similar processes, purely correlational methods such as functional imaging cannot guarantee this. When the locations of our regions of interest were transformed to Talairach space (Table 4), the coordinates in autistic brains seemed posteriorly shifted by 5mm to 8mm relative to normal. Our 8mm interval between the centres of neighbouring coronal slices makes it impossible to determine whether this apparent shift is a genuine phenomenon, especially given only six subjects with autism. If this posterior shift is real, it may be a byproduct of abnormality in distant structures that has perturbed the positions of global landmarks. Such a perturbation might, for instance, occur as an indirect effect of developmental overgrowth of the anterior cerebrum.
Sensitive statistical methods can, of course, be a double-edged sword; an increase in statistical power can come hand in hand with increased susceptibility to Type 1 error. Our strategy of computing z-scores for each individual subject and then computing a second, group statistic over these individual statistics implements a mixed-effects (often referred to as `random-effects') statistical model [47]. Although mixed-effects analysis does prevent perturbation of the statistical result by one or a few individuals who are atypical of the rest of the sample, it cannot guard against the possibility that the sample as a whole is atypical of the population. The circumstances of the current study admit at least two possible types of sample bias: a chance bias in random sampling, and a systematic bias to due selection of only high-functioning cases. In the former case, the small size of the the experimental sample increases the risk that this sample is not representative of the population from which it has been drawn. In the latter case, findings may be peculiar to high-functioning people with autism. It may especially be the case that compensatory cognitive strategies and other secondary phenomena differ in high-functioning and `low-functioning' subpopulations. Further work that includes larger samples and more directly addresses autism's developmental aspect will be of value in examining these possible biases.
We have presented physiological evidence for impaired early selection of relevant stimuli and for compensatory suppression of irrelevant stimuli in adults with autism spectrum disorders. We have also outlined a sequence of physiological and cognitive dysfunctions by which hyper-arousal and impaired selection may produce weak central coherence and other higher-order cognitive patterns typical of autism. Our findings, though statistically robust within our sample, are based on a small number of adult subjects with heterogeneous diagnoses within the autism spectrum. Future work would benefit from younger, larger, and separate samples of autism, Asperger syndrome, and also `unaffected' sibs who have features of the broader autism phenotype. Contrasting across all these sub-populations the relationships between elementary abnormalities of attention and higher-order abnormalities of cognition could highlight the junctures at which cognitive development goes awry. Discovering the maturational windows at which a mild proto-deficit becomes magnified into the full syndrome of autism would be a key step towards the development of targeted interventions.