The Neural Substrate of Reward Anticipation in Health: A Meta-Analysis of fMRI Findings in the Monetary Incentive Delay Task

On 15/12/14 we searched the NICE Healthcare Database including EMBASE (Ovid), MEDLINE (Ovid), PsycINFO (Ovid) and CINAHL (EBSCO) using the terms ((“monetary” AND “incentive” AND “delay” AND “fmri”).ti,ab OR (“monetary” AND (“reward” OR “incentiv*” OR “anticipat*”) AND “fmri”).ti,ab) AND “article” [Limit to: Publication Year 2000–2014]. We complemented this with a cross-reference search of Pubmed on 16/12/14 with the general search term “monetary reward incentive anticipation fMRI”. Study inclusion criteria were (i) inclusion of healthy adults, (ii) used fMRI, (iii) used Monetary Incentive Delay Task, (iv) article available in English, (v) published in a peer-reviewed journal, (vi) conditions and contrasts of interest included, (vii) whole-brain analysis reported. We also included the placebo condition of intervention studies in healthy participants meeting criteria. Full articles were read and excluded if the above inclusion criteria were not met. The search was repeated using the same search terms and databases on 19/1/17, ranging from December 2014 onwards and excluding any repeated articles.

For all the articles that satisfied study inclusion criteria, all authors or corresponding authors were contacted by published email address in March 2015 requesting whole brain maps. Research groups who sent maps were included in the Monetary incentive delay Task Analysis Consortium (MTAC) (Supplementary References 1). If no maps were available for use, we manually extracted the whole-brain only coordinates from the published article for the conditions of interest for the healthy adult sample. For all articles (both maps-received and coordinates only), information on the MRI scanner, scanning sequence and timings, parameters of the monetary incentive delay task and performance data where available were extracted manually by two researchers (Colizzi and Wilson) and cross-checked for errors. This process of contacting authors was repeated for a follow-up literature search in March 2017. If no maps were available, we did not extract published article because of positive effect-size bias detected in the first round (see Results).

Meta-analysis was carried out using the seed-based d mapping software employing previously described methods (Radua et al. 2014; Radua and Mataix-Cols 2009; Radua et al. 2012, 2010). In brief, the aim of seed-based d mapping is to create voxel-level maps of effect-size measured as Hedge’s g allowing modelling of both positive and negative activations on the same map. For each reported peak coordinate and value (t, z or p), seed-based d mapping ensures surrounding voxels have a similar but smaller estimated effect size by multiplication with an un-normalised Gaussian kernel. If a voxel should have a value assigned from more than one coordinate, the values are averaged weighting by the square of the distance to each close peak. The data from each study are then weighted by the inverse of the sum of variance plus between study variance and combined using the random effects model DerSimonian-Laird estimator (DerSimonian and Laird 1986). This approach allows for studies with a larger sample size or lower variability to contribute more and creates a map of heterogeneity.

In neuroimaging analyses, as well as in meta-analyses, the need for appropriate statistical thresholding to minimize the extent of false positive results must be balanced against the need for avoiding false negatives (Lieberman and Cunningham 2009). Typically, neuroimaging meta-analyses control for the false discovery rate to minimize false positive results, which we have also reported. However, it is worth noting that the choice of threshold is best guided by the specific research context (Müller et al. 2018). In the present study, we did not employ the conventional threshold, but instead combined an intensity threshold (p < 0.005) with a cluster extent threshold (cluster >10 voxels) that has been shown to result in acceptable Type II error rates (Lieberman and Cunningham 2009).

Thus seed-based d mapping generated mean maps were thresholded at the validated default settings p < 0.005, seed-based d mapping-z > 1.0, cluster>10 voxels. In seed-based d mapping, using a cluster size of 10 voxels and an uncorrected p = 0.005 has been shown empirically to be equivalent to corrected p = 0.05, optimally balancing sensitivity and specificity, and seed-based d mapping-z > 1 reduces the false positive rate (Radua et al. 2012). Given 100% parametric maps, this results in 100% sensitivity and a 3.5% false positive rate to produce the final whole brain maps of statistical significance, accompanied by an HTML document of main peaks, p value, seed-based d mapping-z-score, MNI coordinates, number of voxels per cluster and significant sub-peaks within each cluster (also with p value and seed-based d mapping-z-score). All results at the default seed-based d mapping-z > 1 threshold are reported in the supplementary material for the interested reader, however in the main text we have reported only those results that survived a more conservative threshold (z-score ≥ 5.0). The ‘5-sigma’ threshold for significance (Horton 2015) is a consensus agreement in other scientific disciplines, reflecting five standard deviations from the mean, or p < 0.0000003). We have adopted this threshold in order to further minimize the probability of detecting an effect by chance and to identify the most critical regions involved in motivational salience, reflecting high confidence that these are true positives when contrasting anticipation win or loss with neutral.

Jacknife sensitivity analysis of replicability, analysis of heterogeneity and publication bias are automated within the seed-based d mapping program. Jacknife analysis involves re-analysing mean maps multiple times by leaving out a single study each time. In order to interpret the many Jacknife mean maps, we thresholded for significance, binarized the data and combined into a single overlapping density map of significant data. This allowed visual inspection of areas of low density, signifying lower replicability across studies. Inter-study heterogeneity was calculated in which Q_H statistics are converted to standard z values to create a map. This map was overlaid on the final mean map for visual inspection of areas of overlapping significant heterogeneity with areas of thresholded activation or deactivation. Publication bias was estimated using standard funnel plots and Egger’s test for each reported peak. The funnel plots consisted of effect-size on the x-axis and standard error on the y-axis with bias tested using Egger’s test for asymmetry of the funnel plot.

The anatomical location of the peaks were identified using the Atlas of the Human Brain and associated BrainNavigator program (Mai et al. 2007). As the atlas did not cover the cerebellum, so the Talairach Daemon (Lancaster et al. 1997; Lancaster et al. 2000) was used, following MNI to Talairach conversion.

Following mean map analysis, the two contrasts AWAN and ALAN were compared using the automated linear model analysis which calculates the between-group difference based on statistical significance after Monte Carlo randomisation. In order to study potential confounders, automated linear model meta-regression was calculated, whereby the difference between the minimum and maximum values of regressors are returned with statistical significance based on Monte Carlo randomisation.

This contrast was examined in a total sample size of 274 participants (Table 1).

This contrast was examined in a combined sample of 246 participants (Table 1).

On between-group linear model comparison between AWAN and ALAN, we found significantly greater deactivation in AWAN compared to ALAN in the left inferior frontal gyrus opercular part (Fig. 4, Supplementary Table 6).

Funnel plots revealed no clear evidence of marked asymmetry on visual inspection for the higher thresholded main peaks for either contrast (Supplementary Fig. 1), also surviving Egger’s test for bias (Tables 2 & 3). Maps of significant heterogeneity (Q_H) were overlaid on the mean maps for both contrasts to visually inspect for overlapping areas (Supplementary Fig. 2). For AWAN significant heterogeneity overlapped with activation in the bilateral ventral striatum and bilateral dorsal thalamic nuclei. For ALAN, overlap was seen in bilateral nucleus accumbens, bilateral medial superior frontal gyrus and the left precentral gyrus. Visual inspection of Jacknife analysis overlaid on the heterogeneity maps showed good replicability of all areas reported in both contrasts (Supplementary Fig. 3).

Outside of the task conditions, two independent regressors were available, namely, placebo intervention and scanner strength. Seventy of the 274 (26%) subjects in AWAN and 59 of the 246 (24%) in ALAN were given placebo (wash-out period 1 to 2 weeks). Overlaying heterogeneity maps for each contrast with meta-regression for placebo showed overlap with activation in AWAN in the bilateral ventral striatum, but no visual overlap in the ALAN contrast (Supplementary Fig. 4, Supplementary Table 8).

Regarding field strength, for AWAN 7 of 14 mean maps were generated using a 3 T scanner, for ALAN 5 of 11 maps were 3 T and the remainder for both 1.5 T. No areas of overlap in heterogeneity corresponded to areas of activation or deactivation in the mean maps (Supplementary Fig. 5, Supplementary Table 9).

The aim of the present study was to investigate which regions of the brain are robustly engaged by reward processing in healthy adult humans, specifically looking at motivational salience, that is, presentation of reward incentive and preparation of approach behaviour. Using all available fMRI group-map data for the anticipation of reward and loss conditions of the monetary incentive delay task in healthy adult humans, our meta-analysis shows that there are large areas of both activation and deactivation across the whole brain in motivational salience. For the first time, we report the pattern of deactivation in anticipation of reward in the monetary incentive delay task in healthy adult humans. We also report evidence of positive effect-size bias in the anterior cingulate (AWAN) and striatum (AWAN and ALAN) in the literature.

The results presented here are consistent with those reported in a recent activation likelihood estimation based meta-analysis of reward anticipation using the monetary incentive delay task in healthy adults (Oldham et al. 2018), but we report far more extensive areas of activation involving frontal, temporal, parietal and cerebellar regions. Furthermore, we have reported on areas of relative deactivation as well as publication bias, heterogeneity, replicability, and the potential effects of placebo and scanner strength.

We have confirmed a number of regions of activation reported in the original monetary incentive delay task studies of healthy adults (Knutson et al. 2001; Knutson et al. 2000) for AWAN, including bilateral insula (left insula originally reported as motor cortex), right nucleus accumbens, nucleus accumbens, bilateral caudate, left putamen, thalamus, right amygdala, right anterior cingulate gyrus (both reported as ‘right mesial prefrontal cortex’), right superior frontal gyrus medial part (reported as ‘right SMA’) and right cerebellum anterior lobe culmen (reported as cerebellar vermis). However, we did not find any activation in the right amygdala or left nucleus accumbens, an area of significant heterogeneity. We have also greatly extended the pattern of activity such that all previous unilateral peaks, except right amygdala, were found to be bilateral including putamen, superior frontal gyrus (medial and lateral), anterior cingulate gyrus, cerebellum anterior lobe culmen, and thalamic nuclei. We have provided greater resolution within the thalamus itself, finding activation in bilateral medial dorsal thalamic nuclei and two left ventral anterior thalamic nuclei. We report nine additional bilateral areas of activation including middle frontal gyri and inferior temporal gyri, precentral gyri, frontal operculae, paracingulate cortex, postcentral gyri, superior parietal lobules, precunei and insular gyri. We report twenty additional unilateral areas of activation (5 main peaks) including areas known to be implicated in both salience processing, such as the hippocampus (Crottaz-Herbette et al. 2005), and reward processing, such as the right parahippocampal gyrus and right inferior frontal gyrus (Brooks et al. 2013).

As predicted, the striatum is strongly engaged in both anticipation of reward and loss, though not clearly differentiated when comparing the two contrasts. We found significant heterogeneity in this area in the anticipation of reward across all studies which may be explained in part by placebo effects.

Both reward and loss anticipation robustly engage key nodes of the salience network including anterior insular and anterior cingulate cortex. However, a mixed picture emerged for the anterior cingulate with anticipation of reward strongly activating bilateral anterior cingulate with a single main peak of deactivation in the left anterior cingulate. In contrast, anticipation of loss strongly activated the left anterior cingulate only with no detected deactivation. This pattern of activation may be reflected in the AWAL between-group comparison (Supplementary Table 6) showing a subthreshold (cluster size 8 voxels) difference in activation in the left anterior cingulate. The anterior insula was activated bilaterally in both contrasts. The insula cortex is considered a major cortical target of ascending interoceptive and visceromotor signals passing through thalamic nuclei (Uddin 2015) found to be functionally connected to amygdala, dorsomedial thalamus, hypothalamus peri-aqueductal grey matter (Seeley et al. 2007). We confirmed activity in these regions, though seemingly a different pattern for each contrast (AWAN - bilateral dorsomedial thalamic nuclei, right periaqueductal grey matter; ALAN - right posterior hypothalamic area, left basomedial nucleus of the amygdala).

The right anterior insular cortex, in particular, is also thought to be involved in switching between two other networks, namely, the central-executive and the default-mode networks (Sridharan et al. 2008). The central-executive network is considered the neural substrate underlying cognitive processes such as inhibition, interference control, working memory and cognitive flexibility (Diamond 2013) and is thought to be anchored in the dorsolateral prefrontal and lateral parietal cortex (Sridharan et al. 2008). We found strong bilateral activation of dorsolateral prefrontal cortex in both contrasts in the bilateral middle frontal gyri and bilateral superior frontal gyri, lateral part. Regarding the lateral parietal cortex we found significant but differing patterns of activation for each contrast. In both conditions, there was bilateral activation of the superior parietal lobule and unilateral activation of the right angular gyrus. However in AWAN there was unilateral activation of the left supramarginal gyrus and deactivation in the left angular gyrus, and in ALAN bilateral activation of the supramarginal gyrus and unilateral deactivation of the right angular gyrus.

The default-mode network is considered a state of cortical activity independent of external stimuli (originally considered to be ‘waking rest’) convergent with areas active in resting state fMRI (Andrews-Hanna et al. 2014) and task-induced deactivation (Andrews-Hanna, 2012). This network has been further subdivided into a ‘core’ anchored in the posterior cingulate cortex and anteromedial prefrontal cortex, as well as the ‘medial temporal’ and’ dorsal medial subsystems’ (Andrews-Hanna 2012). The core is associated with self-referential processes, the medial temporal subsystem corresponds to past and future autobiographical thought, episodic memory and contextual retrieval, and the dorsal medial subsystem corresponds to social cognition, story comprehension and semantic processing (Yeo et al. 2011). We found bilateral activation of the ‘core’ posterior cingulate in both anticipation contrasts. Regarding the medial temporal subsystem, we observed a differing pattern of activation in the ventromedial prefrontal cortex with unilateral activation of the right posterior orbital gyrus in AWAN and unilateral activation of the left intermediate orbital gyrus in ALAN. In AWAN in the medial temporal lobe, there was bilateral deactivation in the parahippocampal gyrus in conjunction with unilateral activation in the left parahippocampal gyrus. Yet in ALAN we reported unilateral activation of the left hippocampus CA1 region. With respect to the dorsomedial subsystem, on the one hand we found robust bilateral activation of the superior frontal gyrus, lateral part in both contrasts. On the other hand, we saw a mixed picture in the lateral temporal cortex. The inferior temporal gyrus was activated in both contrasts. In AWAN, there was activation in the right middle temporal gyrus, deactivation in the right middle gyrus and deactivation in the right superior temporal gyrus. In ALAN there was activation in the left middle temporal gyrus, deactivation in bilateral middle temporal gyri and deactivation in the left superior temporal gyrus.

The monetary incentive delay task is also a motor processing task and, as such, we observed broad bilateral activation of primary motor cortex (precentral gyrus), somatosensory cortex (post central gyrus), the supplementary motor area and multiple thalamic nuclei in both contrasts. The striatal region of the basal ganglia was activated, but no peaks were seen in the globus pallidus or subthalamic nuclei. Significant activity was also found in the cerebellum. In AWAN there was strong activation in bilateral anterior lobe of the cerebellum and a peak of deactivation in the right posterior lobe, but in ALAN a mixed picture of unilateral activation and deactivation in the right anterior lobe. It is known that there is an important function for the cerebellum in motivational salience, in keeping with a recent animal study (Cutando et al. 2013) suggesting a role in encoding expectation of reward, the growing understanding of the reciprocal connections between cerebellum and basal ganglia (Niendam et al. 2012) and cerebellar involvement in addiction (Moulton et al. 2014).

These observed patterns of activation are controlled by the neutral condition in which the only theoretical difference in task activation is the absence of motivational salience. Could it be that common areas of activation and deactivation are part of a motivational salience network, robustly engaging the striatum, salience network, parts of the central executive and default networks and associated motor regions? The picture emerging from the central executive and default network appears more complicated with robust activation in the posterior cingulate and dorsolateral prefrontal cortex, but a differing pattern emerging between anticipation of reward and loss in the lateral parietal cortex, ventromedial cortex, medial (including hippocampus) and lateral temporal cortex.

The first meta-analysis of healthy controls discussed earlier compared uncontrolled anticipation of reward directly with loss (Knutson and Greer 2008). We used a different approach (between-group linear model) comparing mean group maps for two controlled conditions, AWAN and ALAN. We report only one significant peak above threshold in the left inferior frontal gyrus operculum. This region was deactivated in both conditions, but significantly more deactivated in anticipating loss than reward. Additionally, the left anterior cingulate gyrus was active in both contrasts, significantly more so anticipating reward than while anticipating loss, but just under threshold (Supplementary Table 6). The left inferior frontal gyrus has previously been implicated in response inhibition (Swick et al. 2008; Bhattacharyya et al. 2015) and the attribution of aspects of stimulus salience and attentional allocation of resources (Seeley et al. 2007; Bhattacharyya et al. 2012; Downar et al. 2002). Might this left inferior frontal activation reflect differential allocation of cognitive resources according to motivational salience or could this directly reflect a representation of valence?

Our findings are limited by the reported sample demographics, constraints of the monetary incentive delay task, the data acquisition and analysis, the methods of meta-analysis and the brain atlas used. Across all the studies, we were only able to report sample size, age, gender and handedness. Other variables such as substance use, education, IQ, socioeconomic status and ethnicity were too inconsistently reported for any meaningful interpretation.

In terms of the monetary incentive delay task, there was significant variation across designs. The maps we received were from many different international locations including Europe, the USA, Japan and South Korea. Therefore the financial incentive could have different meanings to different samples, although superficially they appear to be similarly small amounts of money. There was inconsistent reporting of, and variation in, the duration of all the phases of each trial, including monetary incentive stimulus presentation, anticipation, target, feedback and inter-stimulus interval. The data are gathered over multiple trials varying from 44 to 180 in each study, and the target hit rate was either unreported or varied from 50% to 75% which could have influenced any neural substrate for the temporally bound reward prediction error signal. The monetary incentive delay task is not designed to capture temporal learning or any reward prediction error signal. Variants of the monetary incentive delay task have been developed that introduce contingency, varying the predictability of reward receipt following action, and we have included a single group map in the AWAN condition introducing this (Li et al. 2014). Lastly, overall performance data was not consistently reported, preventing meaningful behavioural interpretation.

Regarding data acquisition and analysis, a variety of scanners and sequences were used in each study. There was variation in acquisition and echo time, number and thickness of slices, resolution and analysis software. All of these are documented in Supplementary Table 3 for reference. Only two potential confounders were available for statistical analysis, scanner strength and placebo condition, though meta-regression and comparison with heterogeneity and mean maps did not suggest any significant effects on our results.

The most important limitation to our method was the exclusion of over half of the studies. We felt this to be an essential step after identification of the positive effect-size bias in published coordinate-based data. The main peaks of the omnibus analysis (text coordinates and group maps) are reported in Supplementary Table 4 for reference. Secondly, because we found such a large number of significant peaks, we could not report all the data generated by this meta-analysis and applied a very conservative threshold. Subsequently, on the one hand, we feel our results are extremely robust, but on the other, many regions which may be of interest are not discussed. However, all these data are recorded in the Supplementary Table 5.

In terms of the between-group linear model comparison of AWAN and ALAN, there was a difference in gender proportion between AWAN (78%) and ALAN (71%), though this effect just fails to achieve nominal significance (χ2 (1) = 3.34, p = 0.07). However, 10 of the 15 studies (n = 194) for which we received group maps, sent both AWAN and ALAN contrasts, such that the majority of the data was repeated-measures within-group, strengthening the results.

Finally, the brain atlas used for anatomical location, while strong on detail, is based on the dissection of a single 24 year old male brain and does not cover the cerebellum necessitating additional use of the Tailarach Daemon.

Notwithstanding these limitations, this image-based meta-analysis uses data from a large sample of healthy individuals and reveals robust engagement of brain areas that have previously been linked to the anticipation of reward as well as novel brain areas, providing a definitive map of the reward anticipation network in the healthy human brain.