Prediction and prediction error in autism: A meta-analysis of functional magnetic resonance imaging results

In autism spectrum condition (ASC), altered priors/predictions or prediction errors have been hypothesized to increase the influence of bottom-up sensory input, relative to top-down prior knowledge. Such alterations could account for several autistic features, but their empirical basis is unclear. In neurotypical individuals (NT), multiple neuroimaging meta-analyses have aimed to outline domain-general ‘prediction networks’ of the brain. However, there has not been a similar meta-analysis in autism. We searched the literature for functional magnetic resonance imaging and magnetoencephalography studies with autistic participants. The contrasts that explicitly or implicitly involved the processing of priors/predictions and prediction errors were selected. Contrasts were divided into those that reported stronger activation in ASC compared to NT (ASC > NT; 8 contrasts; 139 ASC and 150 NT participants), and vice versa (NT > ASC; 13 contrasts; 261 ASC and 289 NT participants). The convergence of differences between the groups was then tested using activation likelihood estimation meta-analysis. We additionally identified 38 contrasts without significant group differences. ASC > NT did not result in significant convergence. NT > ASC converged in a cluster in the medial frontal gyrus/cingulate gyrus. We found converging NT > ASC activation differences in an area associated with error monitoring and uncertainty estimation. Our results are generally consistent with notions of altered predictions or prediction errors in autism, pointing to differences at high levels of the information-processing hierarchy. However, we recommend a cautious interpretation, given the limited number of available contrasts and the high proportion of null results. Several features of autism have been hypothesized to be related to alterations in the way the brain predicts upcoming events and inputs. In this meta-analysis of functional magnetic resonance imaging (fMRI) studies, we show that when autistic and neurotypical groups engage in tasks and activities that directly or indirectly involve predictions, there are converging brain activation differences in the anterior cingulate cortex, a region involved in uncertainty estimation and error detection. While this finding could be consistent with hypotheses of prediction-related differences in autism, we also find that over half of fMRI experiments do not detect reliable group differences.