We conducted a systematic literature search to identify neuroimaging studies of prediction error using PubMed¹, Web of Science², and Neurosynth³ databases. We searched for articles in the English language using the keywords “prediction error” AND “fMRI” and did not specify a time span for date of publication. The search revealed 8’610 results as of July 2016. To narrow the results, a second search was performed using the keywords “prediction error” AND “fMRI” AND “behavior change” as well as “prediction error” AND “fMRI” AND “observational learning”. Again, no time span was specified. These searches revealed 111 results and four results, respectively. The abstracts of these articles were examined to select potential matches for our inclusion criteria. We also scanned the reference lists of the results to search for additional articles, which met our inclusion criteria. We retrieved the full text of 72 articles for further examination. We predefined study selection criteria to minimize ambiguousness in the study selection. The criteria can be requested of the corresponding author. Studies were included when they met the following criteria: (1) experimental prediction error paradigm and (2) report of voxel-wise-brain analysis for a prediction error main effect, which yielded a total of 59 articles. We excluded studies which did not report prediction error for healthy adults or used medication in their experiment (n = 6 studies excluded). We did this to include only prediction errors which arise from an unexpected change in contingencies in alignment with the clinical model. Of these studies, we precluded those articles failing to experimentally induce a prediction error by changing the contingencies between stimuli and outcome (n = 10 studies excluded). A flowchart of the selection process is shown in Figure 1. The studies included in the meta-analysis are listed in Table 1.

We included all analyses which contrasted prediction error brain activity with brain activity during expectation confirming trails or paradigm specific variations of these contrasts. Of the 43 studied that met all inclusion criteria, we included 60 contrasts in the analysis. If the coordinates were reported in Talairach space they were transformed to Montreal Neurologic Institute (MNI) space using the GingerAle software (Eickhoff et al., 2009, 2012), which utilizes the icbm2tal transform algorithm (Lancaster et al., 2007). In total, we included 446 foci into the analysis.

We performed an activation likelihood estimation (ALE) analysis using the Software GingerAle (Eickhoff et al., 2009, 2012). The algorithm assesses above-chance clustering between experiments, using a probability distribution centered at each of the foci used in the analysis. Since the spatial relationship is assumed to be fixed in each experiment, the ALE analysis infers random effects (Eickhoff et al., 2009). We used the algorithm described in Turkeltaub et al. (2012), which organizes the foci by subject group (as opposed to study affiliation). This prevents an influence of multiple foci from one experiment on the Meta-Analysis results (Turkeltaub et al., 2012). We performed three Meta-Analyses: (1) studies which encourage a behavioral strategic adaptation following a prediction error, (2) studies, which employed a passive observational paradigm, and (3) an analysis of all studies, which was necessary to perform the contrast analysis. In line with previous studies (Garrison et al., 2013), we defined a false discovery rate (FDR) method with p < 0.05 and a minimal cluster volume of 50 mm³. We then performed a contrast analysis of the “active” behavioral subset and the “passive” observational study subset. This analysis allows the subtraction of two datasets to compare differences in brain activity between these two. To this end, a pooled dataset is created, which then serves as basis for two randomly created datasets with the same number of foci as the original datasets. A permutation of subtractions of simulated datasets are compared to the results of the original datasets. We used an uncorrected p-value p < 0.05 since the single analyses were already corrected with FDR (Eickhoff et al., 2011). We chose a minimal cluster volume of 50 mm³ for the contrast analysis. Papaya⁴ was used to superimpose the ALE cluster results on a T1 brain template (Colin27_T1_seg_MNI.nii⁵).

Twenty-one significant clusters were identified by the ALE meta-analysis of all 43 studies. The results show activation in the right basal ganglia and the right insula (see details in Figure 2 and Table 2). There was no clear indication of laterality in the main analysis.

When analyzing all prediction error studies, which employed a behavioral reaction following a prediction error, the ALE meta-analysis revealed 17 significant clusters. We found activation in the striatum, the insula and the claustrum (see details in Figure 3 and Table 3).

The ALE meta-analysis of all prediction error studies, which employed a passive paradigm revealed four significant clusters. We found activation in the putamen, the lateral globus pallidus, declive and the lingual gyrus (see details in Figure 4 and Table 3).

In both analyses, no clear indication of laterality was found.

When performing a meta-analysis containing all prediction error studies our results are in line with previous research (Garrison et al., 2013). We found activation in the striatum, the insula, thalamus as well as fronto-medial structures. The Putamen and the Caudate body are part of the striatum whose association with memory processes is consistent with previous literature (Grahn et al., 2009; Provost et al., 2015). The insula has been associated primarily with fear conditioning (Kircher et al., 2013) but also with reinforcement learning for reward (Lawrence et al., 2014) as well for avoidance learning (Palminteri et al., 2012). Consistent with our hypothesis we also found activation in the areas associated with the fronto-striatal circuits (Leh et al., 2007). In addition to the striatum, we found activation in the globus pallidus, the thalamus and frontal structures, i.e., the left superior, media and middle frontal gyrus.

When contrasting the differences in neuronal activity of prediction errors computed in active behavioral adaptation and passive observational paradigms we found higher activation in the striatum, the insula and the fusiform gyrus.

The medial globus pallidus is part of the four corticostriatal loops, which are responsible for executive function, visual processing, motor function and motivational evaluation (Seger, 2006). It serves as an output nucleus of the basal ganglia and projects to the thalamus, the centromedian nucleus, and the pedunculopontine nucleus (Nauta and Mehler, 1966). These structures are associated with goal-directed motor actions as well as reward learning and evaluation (Hong and Hikosaka, 2008; Haber and Knutson, 2009; Sescousse et al., 2013).

The putamen is associated with novel motoric executions as well as in ambiguous action tendencies, i.e., if the best motoric strategy is unclear (Grahn et al., 2009). Moreover, due to findings of strong connectivity of the putamen with prefrontal regions, it is suggested that the putamen has a cognitive rather than solely motoric function (Provost et al., 2015).

The caudate body has been shown to be involved in cognitive tasks such as categorization and reward information assessment in monkeys (Yanike and Ferrera, 2014) as well as in humans (Packard and Knowlton, 2002). Further, it has been suggested, that the caudate nucleus is involved in evaluating outcomes post-decision (Badre, 2012; Kepecs and Mainen, 2012).

Most studies do not specifically differentiate between distinct parts of the striatum, but investigate the striatum in its entirety. The striatum has been associated with strategizing in avoidance learning (Palminteri et al., 2012), failure or success to learn associations in instrumental conditioning (Schönberg et al., 2007; Horga et al., 2015), decision making and motor initiation (Nagano-Saito et al., 2014).

The insula is associated with the perception and processing of interoception of emotional states (Zaki et al., 2012; Simmons et al., 2013).

The fusiform gyrus is associated with facial and body recognition (Peelen and Downing, 2005) as well as a sensitivity to visual words (McCandliss et al., 2003). The area of the fusiform gyrus showing peak activation is also associated with object recognition (Bar et al., 2001).

The functions of these areas can be incorporated into the processing of prediction errors computed in a behavioral paradigm. The higher activation of the putamen in prediction errors with behavioral changes might be due to the determination of a novel motoric behavior and its initiation. Due to its’ evaluative properties, the caudate nucleus could function as a constant evaluation unit, comparing expected and actual outcomes. The involvement of the insula cannot be explained by the emotional valence of the stimuli used in the studies, since not all the studies comprising the insula cluster contained emotional content, such as negative feedback. They share, however, a high level of uncertainty in their paradigms, e.g., temporal uncertainty or ambiguous stimuli or categories. The processing of uncertainty has also been shown to be associated with insula activity (Simmons et al., 2008; Sarinopoulos et al., 2010), which could be interpreted as an aversive and thus emotional state.

Rief et al.’s (2015) model proposes that following an expectation violation, various information processing mechanisms decide whether an expectation is changed and integrated or maintained and reinforced. In order to shed light on the cognitive processes involved in an expectation change following an expectation violation, we investigated the brain areas more active in paradigms encouraging a behavioral strategic change following a prediction error.

The striatum might be involved in learning the specific contingencies between stimulus and outcome. This might eventually form an expectation about the action strategies resulting in a rewarding outcome. However, when facing an expectation violation, the caudate body might signal a non-rewarding outcome, even though the same behavioral strategy has been employed. On the other hand, if the environment encourages a passive behavior, i.e., no action has to be taken following an expectation violation, the individual is not required to determine a behavioral alternative. The difference in expectation and outcome could be solved by mechanisms such as immunization, leading to an expectation persistence (Kube et al., 2016). In contrast, if the environment encourages or even enforces the use of action alternatives, e.g., an active prediction error paradigm or a therapeutic setting, a behavioral reaction to the situation would be necessary. In such a case, the putamen could be involved in determining a novel behavior and initiate the action alternative by projecting to the medial globus pallidus. This structure could then initiate the motoric aspect of the action alternative. The thalamus, the centromedian nucleus, and the pedunculopontine nucleus could be involved in assessing the reward when employing the new behavior. If the action alternative leads to a satisfying result, i.e., a rewarding outcome, the behavior is integrated and leads to an expectation change.

The involvement of the insula especially in prediction error paradigms encouraging behavioral adaptation suggests an emotional component to be important. Due to its association with avoidance learning, the insula might be involved in assessing aversive outcomes following an action alternative. This contrasts with the reward assessment of thalamus, centromedian nucleus and pedunculopontine nucleus. It might be possible, that the avoidance of an unwanted outcome, e.g., a negative emotional state, is as important as the gain of a rewarding outcome. The aversion of a negative emotional state might be a rewarding outcome in itself which has to be considered when assessing the reward of an action alternative.

A few methodological limitations have to be considered. First, the studies we included used various paradigms, showing a wide range of stimuli, tasks and underlying mathematical models. However, there is evidence of different brain activity involved in various types of learning and in particular in model-based (i.e., goal-directed actions) vs. model-free (i.e., habit-based actions) approaches (Maia, 2009; Wunderlich et al., 2012). Moreover, the ALE meta-analysis itself has a few limitations. Coordinate-based analyses accumulate power across studies (Costafreda, 2009) and cannot reproduce the same quality in results as image-based meta-analyses (Salimi-Khorshidi et al., 2009). A third limitation is that the results of the subtraction analysis do not survive a FDR corrected threshold. Considering this restriction, the results of the contrast analysis have to be interpreted with caution. Eickhoff et al. (2016) recommend a minimal sample size of 17 studies for the ALE meta-analysis. We could only include 19 studies using a passive observational paradigm. This suggests that the statistical power may be rather small for the subtraction analysis, explaining why our results did not survive the FDR corrected threshold. For future research, it is necessary to repeat the analyses with a larger study sample to increase the statistical power. This will allow a more decisive analysis of the differences in neurological activity between active behavioral and passive observational prediction error paradigms.