Introduction

Goal-directed behavior often involves regulating the interference between competing response tendencies via inhibiting task-irrelevant but predominant responses and maintaining task-relevant processes. The cognitive control of interference has been extensively examined using validated task paradigms, particularly the classical Stroop (MacLeod 1991; Stroop 1935), Simon (Simon et al. 1971), and Flanker tasks (Eriksen and Eriksen 1974). Across these tasks, the responses to a target stimulus are typically slower and less accurate when the stimulus and response are incongruent (e.g., RED printed in green) than when they are congruent (e.g., RED printed in red), i.e., an interference effect. At the neural level, interference control is supported by a neural circuit encompassing the lateral prefrontal cortex (PFC), anterior insula (AI), dorsal anterior cingulate cortex (dACC), and parietal cortex (Cieslik et al. 2015; Derrfuss et al. 2005; Laird et al. 2005b; Nee et al. 2007; Cole and Schneider 2007).

While classical paradigms mainly target non-emotional interference processes, conflicts in daily life often relate to emotional events (e.g., smiling in an angry situation). The inability to control or inhibit emotional distracters furthermore characterizes a variety of psychiatric disorders, ranging from anxiety to depressive, trauma-related, and addictive disorders (Bishop et al. 2004; Etkin et al. 2010; Wang et al. 2008; Froeliger et al. 2012). Accordingly, the past decade has witnessed a surge of interest in examining the neurobiological foundation underlying the control of emotional distracters (Etkin et al. 2006; Egner et al. 2008; Ochsner et al. 2009; Kanske and Kotz 2010; Dolan et al. 2001; Beall and Herbert 2008; Xue et al. 2016; Klasen et al. 2011; Chiew and Braver 2011; Chen et al. 2014, 2016).

These studies usually employ paradigms that allow for a measure of emotional interference comparable to the cognitive interference assessed in the classical paradigms. For instance, in an emotional variant of the Stroop task, participants are presented with facial expressions (e.g., fearful or happy) overlaid with congruent or incongruent emotional labels (e.g., “fearful” or “happy”) and are required to indicate the facial expressions while ignoring the distracter of emotional word labels across the face (Houwer and Hermans 1994; Stenberg et al. 1998; Etkin et al. 2006). In such cases, the emotional interference is based on incompatible response tendencies between emotional expressions and word labels (Krug and Carter 2012; Etkin et al. 2006; Bang et al. 2016; Stenberg et al. 1998; Haas et al. 2006). Similarly, another paradigm requires participants to indicate the valence of emotional prosody in the presence of congruent or incongruent semantic cues (Mitchell 2006a, b, 2013; Rota et al. 2008; Schirmer et al. 2004; Wittfoth et al. 2010). Furthermore, in the emotion-word Flanker task, participants are required to indicate the valence (positive or negative) of a central target word while ignoring flanking emotional words mapped onto either the congruent or incongruent valence of the target (Ochsner et al. 2009; Samanez-Larkin et al. 2009). In these paradigms, incongruent trials in comparison with congruent trials typically induce an emotional interference effect similar to that probed in the classical paradigms (Vanderhasselt et al. 2013; Stenberg et al. 1998). Using these paradigms in combination with neuroimaging methods, numerous studies have investigated the neural signatures underlying emotional interference and attempted to determine the common and specific neural substrates of emotional and cognitive interference processing.

However, previous neuroimaging studies have revealed inconsistent results regarding the specific neural systems that mediate emotional interference. On one hand, many studies have reported that emotional interference involves a domain-general cognitive control network comprising the dACC and lateral PFC, which are also strongly engaged in non-emotional interference processing (Chechko et al. 2009, 2012, 2013; Krug and Carter 2010, 2012; Ochsner et al. 2009; Torres-Quesada et al. 2014; Frühholz et al. 2009; Godinez et al. 2016; Rey et al. 2014; Kühn et al. 2010; Müller et al. 2011). Accordingly, recent meta-analyses on emotional variants of the Stroop task demonstrated a robust engagement of cognitive control regions including the dACC and lateral PFC (Song et al. 2017; Xu et al. 2016). In line with the involvement of this domain-general network, several studies have linked inter-individual variations in emotional interference to neural activity in the dACC and lateral PFC (Chechko et al. 2012; Ovaysikia et al. 2011). Moreover, similar regions have been consistently identified in tasks involving the regulation of emotional experiences or emotion–cognition interactions (Ochsner et al. 2002; Buhle et al. 2014; Cromheeke and Mueller 2014). These findings together prompt a domain-general account holding that a domain-general cognitive control network may be responsible for controlling interference induced by different types of stimuli (Buhle et al. 2014; Ochsner et al. 2009; Ochsner and Gross 2005; Collignon et al. 2008; De Gelder and Vroomen 2000).

On the other hand, it has been demonstrated that emotional interference involves brain regions that are particularly sensitive to emotional stimuli, particularly the amygdala and rostral ACC (rACC) (Etkin et al. 2006, 2010, 2011; Park et al. 2008; Mitchell et al. 2003; Mitchell 2006b; Watson et al. 2013; Eugène et al. 2010). For instance, the amygdala and rACC have been specifically identified during emotional interference processing and have been directly related to the behavioral interference effects in emotional context but not in non-emotional context (Ochsner et al. 2009; Etkin et al. 2006). Likewise, rACC lesions have been specifically associated with impairments in emotional interference resolution, while cognitive interference resolution remained intact (Maier and Di Pellegrino 2012). Moreover, individuals with disorders characterized by emotional dysregulation, such as anxiety or depression, exhibit rACC perturbations in response to emotional interference (Etkin and Schatzberg 2011), which have been linked to the severity of symptoms (Offringa et al. 2013; Etkin et al. 2010; Fonzo et al. 2017). Finally, functional connectivity findings revealed that control of cognitive interference recruits a cognitive control circuit that engages the dACC, lateral PFC, and sensory cortex, whereas control of emotional interference relies on an affective control circuit that encompasses the dACC, rACC, and amygdala (Egner et al. 2008). These findings resonate with a domain-specific account suggesting that the rACC and amygdala specifically mediate emotional interference (Etkin et al. 2011, 2015). Importantly, however, the domain-general and domain-specific accounts are not necessarily mutually exclusive given that common and distinct neural bases of the emotional and cognitive interference are highly conceivable (Zaki et al. 2010; Fleury et al. 2014; Comte et al. 2014). Indeed, a recent meta-analytic study revealed both common and distinct networks for the control of emotion (i.e., emotion regulation) and action (Langner et al. 2018).

Against this background, the present study aimed to achieve a more fine-grained characterization of the neural networks implicated in emotional interference based on meta-analytic and data-driven analyses. First, we utilized a coordinate-based meta-analysis to quantitatively synthesize previous neuroimaging findings on emotional interference with the goal of determining the underlying neural architecture of the phenomenon. We next assessed the correspondence between emotional and cognitive interference to determine the common and distinct neural circuits underlying different types of interference. Second, we examined both task-dependent and task-independent functional connectivity profiles of the identified regions with meta-analytic connectivity modeling (MACM) and resting-state functional connectivity (RSFC), respectively. Third, large-scale network correlates of the identified regions and corresponding functional connectivity profiles were determined by mapping them to seven canonical functional networks covering the whole brain (Yeo et al. 2011; Choi et al. 2012). Finally, the functions of the identified domain-general network were decoded using the BrainMap database, allowing a data-driven quantitative inference on mental processes associated with the network (Laird et al. 2009b).

Together, the complementary analytical strategies employed in the current study not only allowed us to localize brain regions associated with emotional interference, as in previous work (Song et al. 2017; Xu et al. 2016), but also allowed us to first characterize network-level mechanisms underlying emotional and cognitive interference as well as to decode the functional significance of the underlying network (Zhang et al. 2017). In the context of previous work suggesting that the regulation of emotional and cognitive interference might depend on extensive interactions between distributed brain networks (Pessoa 2008; Mesulam 1990), the present large-scale network analysis further supports a system-level understanding of these behavioral domains.

Materials and methods

Meta-analysis

Literature search and selection

A systematic online database search was performed in April 2017 using PubMed, ISI Web of Science and Google Scholar by entering various combinations of relevant search items [e.g., (‘affective conflict’ OR ‘emotional conflict’ OR ‘affective interference’ OR ‘emotional interference’ OR ‘affective compatibility’ OR ‘emotional compatibility’ OR ‘affective stroop’ OR ‘emotional stroop’) AND (‘fMRI’ OR ‘magnetic resonance imaging’ OR ‘neuroimaging’)]. In addition, we explored several other sources, including (1) the BrainMap database (http://brainmap.org), (2) the bibliography and citation indices of the pre-selected articles, and (3) direct searches on the names of frequently occurring authors. The obtained studies were further assessed according to the following criteria: (i) subjects were free from psychiatric or neurological diagnoses; (ii) subjects performed tasks in which they were exposed to interference induced by incompatible emotional events. This included but was not limited to the emotional Stroop (e.g., Chechko et al. 2014) and emotional Flanker tasks (e.g., Ochsner et al. 2009) commonly employed in the literature. For instance, in the approach–avoidance task, participants were asked to respond to stimuli associated with both positive and negative (and thus conflicting) outcomes (Aupperle et al. 2015). Notably, the following frequently used paradigms were excluded in the current meta-analysis: (a) traditional emotion-word Stroop tasks were not included because the stimuli used in this task did not induce incompatible response tendencies (see also Feng et al. 2018; Etkin et al. 2006); and (b) cognitive interference tasks (e.g., a classical Stroop task) performed in the context of emotional stimuli (e.g., Kanske and Kotz 2010; Hart et al. 2010) were not included, since interference in these tasks is induced by conflicts in the cognitive rather than emotional domain. (iii) fMRI was used as the imaging modality; (iv) whole-brain general-linear-model-based analyses [rather than region of interest (ROI) analyses] were applied; (v) fMRI results were derived from a general-linear model based on either a binary contrast or parametric analyses; and (vi) activations were presented in a standardized stereotaxic space (Talairach or Montreal Neurological Institute, MNI). Note that for studies reporting Talairach coordinates, a conversion to the MNI coordinates was employed, as implemented in the GingerALE software with Brett’s mni2tal algorithm (http://imaging.mrc-cbu.cam.ac.uk/imaging/MniTalairach).

Filtering search results according to the inclusion/exclusion criteria yielded a total of 48 “incongruent > congruent” or “incongruent > neutral/baseline” contrasts (503 foci, 1125 subjects) (Table 1). In line with previous meta-analytic studies (Fan et al. 2011; Bartra et al. 2013; Luo et al. 2018; Vartanian and Skov 2014; Fouragnan et al. 2018), activation peaks from binary contrast and continuous parametric analyses were included because both analysis strategies aim to determine brain regions engaged during interference processing. Second, we allowed for variations in contrasts rather than restricting our analyses to a single contrast, as long as the different contrasts (e.g., incongruent > congruent, or incongruent > baseline) carried information about interference processing. This is an advantage provided by neuroimaging meta-analytic methods (see also Wager et al. 2007), such that brain regions consistently reported in different contrasts are detected while those that vary across contrasts (due to such factors as participants, stimuli, and specific contrasts) are likely canceled out. Similar ideas allowed us to compare emotional interference with cognitive interference, the findings of which were derived from different tasks and stimuli.

Table 1 Summary of the studies included in the meta-analysis focusing on emotional interference
Full size table

Another meta-analysis on neuroimaging studies of cognitive interference processing was conducted based on one of our recent publications (Cieslik et al. 2015). In particular, we searched for neuroimaging studies that employed the Stroop, Simon and Flanker tasks to identify brain regions consistently involved in cognitive interference processing. The search and selection procedures resulted in 96 “incongruent > congruent” or “incongruent > neutral/baseline” contrasts (1022 foci, 1689 subjects) (Table S1). Importantly, tasks involving an emotional component were excluded from the meta-analysis of cognitive interference (see also Cieslik et al. 2015). For studies that employed both cognitive and emotional versions of the interference task, coordinates were separately coded for the meta-analysis of emotional and cognitive interference (e.g., Ochsner et al. 2009). Furthermore, the selection criteria used for cognitive studies were similar to those for emotional studies, such that only results from healthy subjects and whole-brain analyses were included.

Main activation likelihood estimation (ALE) analysis

A coordinate-based meta-analysis of reported fMRI studies was conducted, employing the ALE algorithm (in-house MATLAB scripts) (Eickhoff et al. 2009). ALE determines the convergence of foci reported from different functional [e.g., blood-oxygen-level dependent (BOLD) contrast imaging] or structural (e.g., voxel-based morphometry) neuroimaging studies with published foci
in Talairach or MNI space (Laird et al. 2005a; Turkeltaub et al. 2002). ALE interprets reported foci as spatial probability distributions, whose widths are based on empirical estimates of the spatial uncertainty due to the between-subject and between-template variability of the neuroimaging data (Eickhoff et al. 2009). The ALE algorithm weights the between-subject variability based on the number of subjects analyzed in the studies, modeling larger sample sizes with smaller Gaussian distributions and thus presupposing more reliable approximations of the ‘true’ activation for larger sample sizes (Eickhoff et al. 2009).

The union of the individual modulated activation maps first created from the maximum probability associated with any one focus (always the closest one) for each voxel (Turkeltaub et al. 2012) is then calculated to obtain an ALE map across studies. Then, using a non-linear histogram integration algorithm, this ALE map is assessed against a null distribution of random spatial association between studies (Eickhoff et al. 2012; Turkeltaub et al. 2012). In addition, the average non-linear contribution of each experiment for each cluster was calculated from the fraction of the ALE values at the cluster with and without the experiment in question (Eickhoff et al. 2017). Based on the calculated contribution, we employed additional criteria to select significant clusters: (1) the contributions for one cluster were from at least two experiments, so that the finding was not driven by a single experiment; and (2) the average contribution of the most dominant experiment (MDE) did not exceed 50%, and the average contribution of the two most dominant experiments (2MDEs) did not exceed 80% (Eickhoff et al. 2017).

Main conjunction/contrast analysis

After obtaining the consistent maxima separately for emotional and cognitive interference, a conjunction analysis between them was conducted to assess the correspondence. In addition, differences between emotional and cognitive interference were tested by first performing separate ALE analyses for each condition and then computing the voxel-wise difference between the ensuing ALE maps (Eickhoff et al. 2011).

Validation analysis

We implemented additional analyses to validate findings derived from conventional ALE meta-analysis and conjunction analyses. First, we implemented a leave-one-experiment-out (LOEO) analysis for the ALE meta-analyses on the emotional interference. On each fold, one contrast was excluded, and the ALE meta-analysis was conducted on the remaining N − 1 contrasts (N − 1 = 47). Afterwards, we conducted a conjunction analysis on the ALE results of all folds to identify the brain regions that were robustly engaged in emotional interference. These analyses were employed to make sure that our main ALE meta-analysis findings on emotional interference were not driven by a single contrast/experiment.

Second, the unbalanced number of contrasts between emotional and cognitive interference (48 vs. 96) could lead to differences in statistical power for the contrasts to be compared. To address this issue, we employed a re-subsampling approach for the meta-analysis of cognitive interference. The re-subsampling was repeated 1000 rounds. On each round, 47 contrasts of cognitive interference were randomly selected to match the number of contrasts in emotional interference. The ALE meta-analysis was then conducted for the selected contrasts of cognitive interference. Afterwards, the ALE results of all rounds were averaged to represent the re-subsampling findings. These analyses were employed to validate our main ALE meta-analysis findings on cognitive interference.

Third, we aimed to validate our main conjunction findings with the following analyses: we computed the overlap between regions that were most robustly engaged in emotional interference (i.e., the voxel-wise probabilities of being declared active in all folds of the ALE-LOEO analysis) and those in re-subsampling analyses for cognitive interference. Importantly, this conjunction analysis was based on equal numbers of contrasts associated with emotional and cognitive interference.

Finally, the ALE approach treats different contrasts within a study as distinct experiments (Laird et al. 2009a). This approach (i) ensures that all relevant information from each study is accounted for in the meta-analysis and (ii) avoids the selection of a single contrast based on subjective selection criteria. However, a disadvantage of this approach is that multiple contrasts from a single study could be related. Although this potential issue could be partly addressed by the LOEO and resampling methods mentioned above, we implemented an additional supplementary analysis by combining different contrasts from the same publication into a single experiment. In this way, the approach prevented multiple non-independent contrasts from a single publication from influencing the results of the meta-analyses (Turkeltaub et al. 2012). Combining within-study contrasts resulted in 34 contrasts for emotional interference and 84 contrasts for cognitive interference.

All maps were thresholded using a cluster-level family-wise error (cFWE) correction (P < 0.05) with a cluster-forming threshold of P < 0.001 using 10,000 permutations for correcting multiple comparisons. Notably, the current analyses were implemented using key functions of the most recent version of GingerALE (version 2.3.6), which uses valid multiple-comparison corrections (Eickhoff et al. 2017).

Task-based connectivity: MACM analyses

To examine the co-activation patterns of the bilateral dACC, AI, left inferior frontal gyrus (IFG) and superior parietal lobe (SPL) commonly recruited by emotional and cognitive interference processing, we conducted MACM analyses—with these regions as ROIs—using the BrainMap Database (Laird et al. 2009a). MACM delineates patterns of co-activation across thousands of studies using neuroimaging databases and produces data-driven functional connectivity maps based on pre-defined ROIs (Langner et al. 2014). The BrainMap database (http://www.brainmap.org/) was used, which at the time of assessment contained coordinates of reported activation foci and associated meta-data of more than 8400 neuroimaging experiments. For our analysis, only whole-brain neuroimaging studies that report activation in standard stereotaxic space and recruit a healthy population were included, while other studies investigating differences in age, gender, handedness, and training effects or clinical populations were excluded. For the dACC, 333 experimental contrasts and 4910 foci from 4555 participants were identified; for the left AI, 302 experimental contrasts and 5068 foci from 4845 participants; for the right AI, 264 experimental contrasts and 4381 foci from 4035 participants; for the left IFG, 219 experimental contrasts and 3250 foci from 3367 participants; and for the left SPL, 168 experimental contrasts and 2893 foci from 2728 participants. First, whole-brain peak coordinates of all those studies from BrainMap were downloaded if the study reported at least one focus of activation within each ROI. Next, coordinates were analyzed with the ALE algorithm (as described above) to detect areas of convergence of coactivations with each seed. Finally, the ALE maps were family-wise error (FWE) corrected at a threshold of P < 0.05 (corresponding to Z = 5.03) at the voxel-level and converted into Z scores for display.

Task-free connectivity: RSFC analyses

Neuroimaging data

To complement task-based connectivity derived from MACM analyses, whole-brain RSFC of the bilateral dACC, AI, left IFG and SPL as ROIs was assessed. Specifically, resting-state fMRI images of 192 healthy volunteers were obtained from the Enhanced Nathan Kline Institute-Rockland Sample (http://fcon_1000.projects.nitrc.org/indi/enhanced/) (Nooner et al. 2012). During resting-state acquisition, subjects were instructed to look at a fixation cross, not to think about anything in particular and not to fall asleep (which was confirmed by post-scan debriefing). For each subject, 260 resting-state echo planar imaging (EPI) data were acquired on a Siemens TimTrio 3T scanner using the following parameters: repetition time (TR) = 1.4 s, echo time (TE) = 30 ms, flip angle = 65°, voxel size = 2.0 mm × 2.0 mm × 2.0 mm, 64 slices.

Image preprocessing

Physiological and movement artifacts were removed from the resting-sate data using FIX [FMRIB’s independent component analysis (ICA)-based Xnoiseifier, version 1.061 as implemented in FSL 5.0.9] (Salimi-Khorshidi et al. 2014; Griffanti et al. 2014). Unique variance related to the identified artifactual independent components was then regressed from the data together with 24 movement parameters (including derivatives and second-order effects as previously described and evaluated) (Satterthwaite et al. 2013). Data were further preprocessed using SPM8 (Wellcome Trust Centre for Neuroimaging, London) and in-house Matlab scripts. The first four scans were excluded prior to further analyses. The remaining EPI images were corrected for head movement using a two-pass (alignment to the initial volume followed by alignment to the mean after the first pass) affine registration. The mean EPI image for each subject was then spatially normalized to the ICBM-152 reference space using the “unified segmentation” approach (Ashburner and Friston 2005). The resulting deformation was applied to the individual EPI volumes, which were subsequently smoothed with a Gaussian kernel with a full width at half maximum (FWHM) of 5 mm to improve the signal-to-noise ratio and to compensate for residual anatomic variations.

Seed-to-voxel connectivity

Implementing a seed-based analysis, the functional connectivity (bivariate correction) between the average BOLD signals from given seed regions (the bilateral dACC, AI, left IFG and SPL) and all other voxels in the brain was computed. The voxel-wise correlation coefficients were then transformed into Fisher’s Z scores and tested for consistency across subjects by a second-level analysis of variance (ANOVA, including appropriate nonsphericity correlation). Consistent with our MACM analyses, results were FEW-corrected at a threshold of P < 0.05 (corresponding to T = 4.92) at the voxel-level.

Large-scale network analysis

To assess the underlying large-scale network correlates, clusters that were revealed by meta-analysis, MACM, and RSFC analyses were overlaid onto seven canonical functional networks covering the cerebral cortex and striatum (Choi et al. 2012; Yeo et al. 2011). Canonical networks include the fronto-parietal network (FPN), dorsal attention network (DAN), ventral attention network (VAN), somatomotor network (SMN), visual network (VN), affective network (AFN), and default mode network (DMN). The relative distribution was computed by the proportion of activated voxels of a given network vs. all activated voxels, while the absolute distribution was calculated by the proportion of activated voxels of a given network vs. voxels of that template network (Zhang et al. 2017).

Functional decoding

After identifying a domain-general network in the ALE meta-analysis, and the MACM and RSFC analyses, we decoded the potential functions of the network based on the BrainMap database (http://www.brainmap.org/). In particular, we performed the functional characterization of the network based on the behavioral domain meta-data categories available for each neuroimaging experiment included in the BrainMap database. Behavioral domains included the main categories cognition, action, perception, emotion, and interoception, as well as related sub-categories (Turner and Laird 2012) (see http://brainmap.org/scribe/).

We determined the individual functional profile corresponding to the domain-general network using forward and reverse inference approaches. Forward inference refers to the probability of identifying activity in a brain network given knowledge of the psychological process, whereas reverse inference refers to the probability of a psychological process being present given knowledge of a particular brain network. In the forward inference approach, a network’s functional profile was determined by identifying taxonomic labels, for which the probability of finding activation in the respective network was significantly higher than the overall chance (across the entire database) of finding activation in that particular network. In particular, we assessed whether the conditional probability of activation given a particular label [P (Activation|Task)] was higher than the baseline probability of activation in the network in question per se [P (Activation)]. In the reverse inference approach, a network’s functional profile was determined by identifying the most likely behavioral domains given activation in a particular network. This likelihood [P (Task|Activation)] can be derived from [P (Activation|Task)] as well as [P (Task)] and [P (Activation)] using Bayes rule. Significance (at P < 0.05, corrected for multiple comparisons using the FDR method) was then assessed by means of a Chi-square test.

Results

Main ALE findings

For emotional interference, the ALE meta-analysis revealed significant convergence of activity in the bilateral dACC/SMA, AI, left IFG, SPL and inferior temporal gyrus (Fig. 1a; Table 2). Twenty-five out of 48 contrasts contributed to the cluster in the dACC/SMA (MDE = 10.84%; 2MDE = 20.96%, Table 3). Eleven out of 48 contrasts contributed to the cluster in the left AI (MDE = 22.07%; 2MDE = 36.63%, Table 3). Twenty-five out of 48 contrasts contributed to the cluster in the right AI (MDE = 10.08%; 2MDE = 16.42%, Table 3). Sixteen out of 48 contrasts contributed to the cluster in the left IFG (MDE = 18.50%; 2MDE = 31.89%, Table 3). Twelve out of 48 contrasts contributed to the cluster in the left SPL (MDE = 15.79%; 2MDE = 31.48%, Table 3). Seven out of 48 contrasts contributed to the cluster in the left inferior temporal gyrus (MDE = 26.04%; 2MDE = 50.97%, Table 3).

Fig. 1
figure 1

Significant clusters from the main coordinate-based activation likelihood estimation (ALE) meta-analysis (cluster-level family-wise error correction (P < 0.05) with a cluster-forming threshold of P < 0.001 using 10,000 permutations) for emotional interference, cognitive interference, and the common clusters for both. a Consistent maximum for emotional interference was found in the bilateral dACC/SMA, AI, left IFG, SPL and inferior temporal gyrus. b Consistent maximum for cognitive interference was found in the bilateral dACC/SMA, AI, left IFG, SPL/IPL, middle frontal gyrus and middle occipital gyrus. c Consistent maximum for the conjunction of emotional interference and cognitive interference was found in the bilateral AI, dACC/SMA, left IFG and SPL. L left, R right

Full size image
Table 2 ALE meta-analysis results for emotional interference and cognitive interference
Full size table
Table 3 Average contribution of each experimental contrast for significant clusters identified for the meta-analysis of emotional interference
Full size table

By overlapping these clusters with the seven functional networks, it was revealed that the identified areas were primarily distributed in the VAN (relative: 22.67%; absolute: 4.24%), DAN (relative: 21.57%; absolute: 3.59%), FPN (relative: 18.94%; absolute: 2.49%), and VN (relative: 17.15%; absolute: 1.97%) (Fig. 2a, b).

Fig. 2
figure 2

Network distribution of significant clusters for emotional interference, cognitive conflict, and the common clusters for both. a Relative network distribution of clusters from primary meta-analyses. b Absolute network distribution of clusters from primary meta-analyses. c Relative network distribution of clusters from validation meta-analyses. d Absolute network distribution of clusters from validation meta-analyses. VN visual network, SMN somatomotor network, DAN dorsal attention network, VAN ventral attention network, AFN affective network, FPN fronto-parietal network, DMN default-mode network

Full size image

The ALE meta-analysis on the cognitive interference revealed significant convergence of activity in the bilateral dACC/SMA, AI, left IFG, SPL/IPL, middle frontal gyrus and middle occipital gyrus (Fig. 1b; Table 2). Fifty-four out of 96 contrasts contributed to the cluster in the right dACC/SMA (MDE = 5.00%; 2MDE = 9.68%, Table S2). Forty-seven out of 96 contrasts contributed to the cluster in the left AI (MDE = 5.92%; 2MDE = 11.22%, Table S2). Fifty-six out of 96 contrasts contributed to the cluster in the right AI (MDE = 6.42%; 2MDE = 11.94%, Table S2). Forty-five out of 96 contrasts contributed to the cluster in the left IFG (MDE = 7.05%; 2MDE = 12.44%, Table S2). Sixty-two out of 96 contrasts contributed to the first cluster in the left SPL/IPL (MDE = 5.25%; 2MDE = 9.97%, Table S2). Thirty-three out of 96 contrasts contributed to the second cluster in left SPL/IPL (MDE = 7.30%; 2MDE = 12.51%, Table S2). Twenty-two out of 96 contrasts contributed to the cluster in the left middle frontal gyrus (MDE = 11.67%; 2MDE = 22.24%, Table S2). Twenty-four out of 96 contrasts contributed to the cluster in the left middle occipital gyrus (MDE = 13.62%; 2MDE = 25.07%, Table S2).

By overlapping these clusters with the seven functional networks, it was revealed that the identified areas were primarily distributed in the FPN (relative: 26.92%; absolute: 9.50%), DAN (relative: 26.00%; absolute: 11.61%), VAN (relative: 21.29%; absolute: 10.68%), and VN (relative: 16.07%; absolute: 4.94%) (Fig. 2a, b).

Main conjunction/contrast findings

A conjunction analysis revealed a common activation maximum in the bilateral AI, dACC/SMA, left IFG and SPL for emotional and cognitive interference (Fig. 1c; Table 2).

By overlapping these clusters with the seven functional networks, it was revealed that the identified areas were primarily distributed in the VAN (relative: 24.90%; absolute: 2.53%), FPN (relative: 21.87%; absolute: 1.56%), DAN (relative: 19.10%; absolute: 1.73%), and VN (relative: 18.05%; absolute: 1.13%) (Fig. 2a, b).

Finally, contrast results (Table 2) showed that the right lateral orbital frontal cortex (lOFC) and right IFG were more activated in the emotional interference than cognitive interference. In contrast, the left AI, IFG, inferior parietal lobule, right middle frontal gyrus and bilateral dACC/SMA were more activated in the cognitive interference than emotional interference. Notably, regions revealed by the contrast analyses were adjacent to those revealed by the conjunction analysis (Fig. S1).

Validation of main findings

LOEO findings

Consistent activation maxima for the emotional interference were found in the bilateral dACC, AI/IFG, left SPL and inferior temporal gyrus (Fig. 3a; Table 4). Therefore, the results of the LOEO approach corroborated the findings of the standard ALE meta-analysis.

Fig. 3
figure 3

Significant clusters from the leave-one-experiment-out (LOEO) analysis for the emotional interference, the re-subsampling results for the cognitive interference, and the common clusters for both (cluster-level family-wise error correction (P < 0.05) with a cluster-forming threshold of P < 0.001 using 10,000 permutations). a Consistent maximum for emotional interference was found in the bilateral dACC, AI/IFG, left SPL and inferior temporal gyrus. b Consistent maximum for cognitive interference was found in the bilateral dACC, AI/IFG, SPL, thalamus, left dlPFC, middle occipital gyrus and right caudate. c Consistent maximum was found in the bilateral dACC, AI, IFG, left SPL and middle occipital gyrus when overlapping average results of LOEO for emotional interference and re-subsampling ALE results for the cognitive interference. L left, R right

Full size image
Table 4 Validation of main meta-analyses and conjunction analyses derived from leave-one-experiment-out and re-subsampling approaches
Full size table

By overlapping these clusters with the seven functional networks, it was revealed that the identified areas were primarily distributed in the VAN (relative: 36.93%; absolute: 7.65%), FPN (relative: 34.70%; absolute: 5.05%), and DAN (relative: 15.01%; absolute: 2.77%) (Fig. 2c, d).

Re-subsampling findings

The re-subsampling ALE meta-analysis on the cognitive interference revealed consistent maxima in the bilateral dACC, AI/IFG, SPL, thalamus, left dlPFC, middle occipital gyrus and right caudate (Fig. 3b; Table 4).

By overlapping these clusters with the seven functional networks, it was revealed that the identified areas were primarily distributed in the FPN (relative: 35.04%; absolute: 19.01%), DAN (relative: 29.34%; absolute: 20.15%), and VAN (relative: 24.90%; absolute: 19.20%) (Fig. 2c, d).

Conjunctions between LOEO and re-subsampling findings

The minimum conjunction between the probability map for emotional interference and the cognitive interference processing was identified in the bilateral dACC, AI, IFG, left SPL and middle occipital gyrus (Fig. 3c; Table 4).

By overlapping these clusters with the seven functional networks, it was revealed that the identified areas were primarily distributed in the VAN (relative: 43.29%; absolute: 6.46%), FPN (relative: 35.70%; absolute: 3.75%), and DAN (relative: 16.02%; absolute: 2.13%) (Fig. 2c, d).

Within-study contrasts combined

This analysis revealed essentially the same findings as those from the main ALE meta-analyses and conjunction analysis (Figure S2 & Table S3), further validating our findings.

MACM and RSFC results

We conducted task-based (MACM) and task-free (RSFC) connectivity analyses to further investigate the functional characterization and connectivity of brain regions commonly recruited by emotional interference and cognitive interference processing, including the bilateral dACC, AI, left IFG and SPL. MACM and RSFC analyses showed remarkably similar results for the functional connectivity of these seed regions (Fig. 4).

Fig. 4
figure 4

Results for the conjunction of task-based connectivity analysis (MACM) and task-free connectivity analysis (RSFC) for the regions commonly involved in the emotional and cognitive interference (voxel-wise P(FWE) < 0.05). a Conjunction of MACM and RSFC for each seed region. b Relative network distribution of clusters from conjunction of MACM and RSFC for each seed region. c Absolute network distribution of clusters from conjunction of MACM and RSFC for each seed region. L left, R right, dACC dorsal anterior cingulate cortex, AI anterior insula, IFG inferior frontal gyrus, SPL superior parietal lobule, MCAM meta-analytic connectivity modeling, RSFC resting-state functional connectivity

Full size image

For the dACC (Fig. S3 & S4 & Table S4), both MACM and RSFC analyses revealed functional connectivity with the medial frontal gyrus, AI/IFG, SPL, putamen, thalamus, middle frontal gyrus, precentral gyrus and cerebellum, which were primarily distributed in the FPN (relative: 32.32%; absolute: 10.57%), VAN (relative: 29.82%; absolute: 13.86%), and DAN (relative: 16.74%; absolute: 6.93%).

For the left AI (Figs. S3 & S4 & Table S5), both MACM and RSFC analyses revealed functional connectivity with the dACC/SMA, AI/IFG, SPL, thalamus and middle frontal gyrus, which were primarily distributed in the FPN (relative: 39.77%; absolute: 13.69%), VAN (relative: 31.58%; absolute: 15.45%), and DAN (relative: 16.90%; absolute: 7.36%).

For the right AI (Figs. S3 & S4 & Table S6), both MACM and RSFC analyses revealed functional connectivity with the dACC/SMA, AI/IFG, SPL, caudate, thalamus, precentral gyrus and middle frontal gyrus. These regions were primarily distributed in the FPN (relative: 42.72%; absolute: 14.10%), VAN (relative: 30.90%; absolute: 14.49%), and DAN (relative: 17.47%; absolute: 7.30%).

For the left IFG (Figs. S3 & S4 & Table S7), both MACM and RSFC analyses revealed functional connectivity with the dACC/SMA, AI/IFG, SPL, middle temporal gyrus, precentral gyrus and cerebellum, which were primarily distributed in the FPN (relative: 43.25%; absolute: 13.80%), DAN (relative: 24.12%; absolute: 9.74%), DMN (relative: 15.56%; absolute: 3.20%) and VAN (relative: 14.79%; absolute: 6.70%).

For the left SPL (Fig. S3 & S4 & Table S8), both MACM and RSFC analyses revealed functional connectivity with the dACC/SMA, AI/IFG, SPL, middle frontal gyrus, inferior occipital gyrus and cerebellum, which were primarily distributed in the DAN (relative: 51.13%; absolute: 15.11%) and FPN (relative: 38.61%; absolute: 9.01%).

Functional decoding of the domain-general network

A conjunction analysis across meta-analysis, MACM, and RSFC revealed a consistent network comprising the bilateral AI, dACC/SMA, left IFG and SPL (Fig. 5). Afterwards, functional characterization according to the BrainMap meta-data was performed for the network. Forward and reverse inference alike indicated a significant association of the network with language-related cognition (especially phonology, semantics, and orthography) and working memory (Fig. 5). Reverse inference additionally revealed an association with attention, pain, and activation inhibition (Fig. 5).

Fig. 5
figure 5

The domain-general network and functional decoding. a Conjunction across meta-analyses, MACM and RSFC analyses. b Quantitative forward inference on the domain-general network. c Quantitative reverse inference on the domain-general network. L left, R right

Full size image

Discussion

The past decade has witnessed an explosion of interest in examining neural signatures of emotional interference (Huang et al. 2013; Comte et al. 2014; Chen et al. 2014, 2016; Chiew and Braver 2011). Neuroscientific studies on this topic often utilize paradigms that induce emotional interference comparable to cognitive interference and assess the neuropsychological underpinnings of the interference effect (e.g., Etkin et al. 2006; Ochsner et al. 2009). The objective of the current work was to quantitatively synthesize the results of previous fMRI studies on emotional interference and to characterize large-scale network correlates of brain regions consistently involved in emotional interference. Our results identified a convergence of reported activation foci in neural circuits important in cognitive control (MacDonald et al. 2000; Ridderinkhof et al. 2004). Specifically, consistent involvement of the bilateral dACC, AI, left IFG, and SPL was identified for emotional interference, and conjunction analyses demonstrated correspondence in these regions between emotional and cognitive interference. These meta-analytic findings were further validated using LOEO and re-subsampling analyses. Furthermore, characterization of connectivity profiles revealed that these regions were extensively connected to each other in both task and resting states, in addition to connectivity with other regions. Finally, large-scale network analyses indicate that these regions as well as their functional connectivity profiles were primarily mapped onto the fronto-parietal network, the ventral attention network, and the dorsal attention network. These findings together suggest that there is a domain-general neural system recruited by different types of interference and that control of emotional and cognitive interference depends on interactions between distributed brain networks.

In line with the domain-general account, the current meta-analysis demonstrated overlap in cognitive control systems consisting of the dACC, AI, IFG and SPL between emotional and cognitive interference. Specifically, the dACC and AI have been implicated in the detection of interference induced by both emotional and non-emotional stimuli (Chechko et al. 2012; Egner et al. 2008; Aupperle et al. 2015; Wittfoth et al. 2010). Furthermore, these regions have been consistently activated by the expectancy violations (Garrison et al. 2013; Wu et al. 2016; Feng et al. 2015) and by events salient to current goals (Seeley et al. 2007; Menon and Uddin 2010). Accordingly, the dACC and AI have been thought to contribute to generic monitoring processes that signal the need for adjustment in cognitive control (Menon and Uddin 2010; Menon 2011). The IFG is frequently engaged by response inhibition and interference suppression, as revealed in different lines of previous neuroimaging studies (Bunge et al. 2002; Levy and Wagner 2011; Aron et al. 2004). The activity of the IFG has thus been attributed to inhibitory control of prepotent responses and information from semantic memory (Swick et al. 2008; Aron et al. 2004; Whitney et al. 2010; Thompson-Schill et al. 1998). Finally, the SPL plays critical roles in selective attention and working memory (Behrmann et al. 2004; Koenigs et al. 2009; Yantis and Serences 2003). For instance, the SPL is usually engaged in the condition where attention needs to be shifted to goal-related features (Yantis et al. 2002; Liu et al. 2003). Considering the consistent co-activation of these regions across emotional and cognitive interference, it is possible that they work together with each other as a connected network to support different aspects of interference control (Spielberg et al. 2015). Specifically, different types of interference might be registered in the dACC and AI as generic control signals driving subsequent increases in cognitive control implemented in the IFG and SPL, among other regions. Those conflicts are in turn resolved by inhibiting task-irrelevant responses (Ovaysikia et al. 2011; Egner et al. 2008; Etkin et al. 2006) and/or by directing attention to task-relevant features (Lee et al. 2007; Chechko et al. 2014).

In accordance with the current findings, many studies have indicated that control of emotion consistently recruits cognitive control systems. For instance, a domain-general control network consisting of prefrontal and parietal regions is often activated in emotion regulation tasks in which participants are explicitly asked to regulate their emotions (Buhle et al. 2014; Kohn et al. 2014; Kalisch 2009; Ochsner et al. 2002, 2012). The regulatory role of these regions is further supported by growing evidence demonstrating that neural activity in the dACC and lateral PFC is directly related to emotion regulation success, presumably via their regulatory control over limbic emotion processing regions (Wager et al. 2008; Kober et al. 2010; Lee et al. 2012; Silvers et al. 2017). Moreover, reliable increases in the dACC and lateral PFC activity are accompanied by successful treatment of patients with mood disorder (Delaveau et al. 2011). Finally, the presence of emotional distractors in cognitive tasks consistently enhances the engagement of cognitive control systems, especially when emotional information impairs ongoing task performance (Cromheeke and Mueller 2014). Considering these data in conjunction with the current observations, implicates a potential neuropsychological mechanism in the control of emotional processing with respect to the pivotal role of the domain-general cognitive control system.

Our connectivity and network analyses further revealed that these regions were connected to each other and mainly engaged the fronto-parietal network, the ventral attention network, and the dorsal attention network. These findings are in line with the domain-general account, considering that the engaged networks have demonstrated functional generality across diverse cognitive challenges, ranging from the control of attention to memory and response (Duncan and Owen 2000; Duncan 2010, 2013; Lückmann et al. 2014; Vincent et al. 2008; Dosenbach et al. 2006). For instance, the fronto-parietal network is thought to contribute to task-set maintaining, long-term planning, and response suppression and selection, among other high-order control processes (Menon 2011; Seeley et al. 2007; Cole et al. 2014; Miller and Cohen 2001). The dorsal attention network presumably implements top-down attention orientation to task-relevant features, whereas the ventral attention network is recruited by the detection of salient stimuli (Corbetta et al. 2000; Kim 2014; Cieslik et al. 2010). Other studies have distinguished the functional roles of these networks as initiating/adjusting control or providing stable maintenance of task sets (Dosenbach et al. 2007, 2008). Engagement of those distinct large-scale networks by emotional and cognitive interference is in line with hypothesis that human emotion and cognition emerge from dynamic interactions between spatially distributed and functionally dissociable brain networks (Barrett and Satpute 2013; Pessoa 2008; Mesulam 1990; Sha et al. 2017). Complementary to the present meta-analytic findings, recent studies employing network analyses have demonstrated that the dACC, AI, lateral PFC, and SPL form coherent large-scale networks, which change their topological properties and dynamic interactions as a function of the level of interference (Spielberg et al. 2015).

Finally, we revealed additional possible functions of the identified domain-general network using the functional decoding analysis. The network was associated with higher-level cognitive domains such as working memory, attention, language, and action inhibition. Therefore, the identified network is associated with functions or processes representing facets of cognitive control, which aligns with our meta-analytic results indicating an involvement of these regions in top-down control across emotional and cognitive interference. Notably, the network is also associated with emotion (pain) processing. These findings are consistent with the idea that interference might be in and of itself affective (Botvinick 2007; Dreisbach and Fischer 2012, 2015). This conjecture has been supported by evidence from both behavioral and brain imaging studies (Schouppe et al. 2015; Dreisbach and Fischer 2015; Braem et al. 2017; Pan et al. 2016).

The current study is the first to employ meta-analytic connectivity modeling, resting-state functional connectivity, and functional decoding to characterize brain networks associated with emotional and cognitive interference. Accordingly, the current study extends beyond localizing brain regions of interference processing (for recent meta-analyses, see also Song et al. 2017; Xu et al. 2016) and unveiled the underlying neural networks and their functional profiles, which could contribute to a better understanding of interference processing. Moreover, the current work is the first to apply functional decoding to characterize the network associated with emotional and cognitive interference, pointing to possible novel functions of brain systems not necessarily addressed in individual studies. Furthermore, this study included the largest number of datasets (48 experiments) and utilized advanced methods (LOEO and resubsampling) that allowed a more precise estimation of the effects than previous ALE meta-analyses of 14 (Xu et al. 2016) or 5 experiments (Song et al. 2017). Relatedly, the current meta-analysis was not restricted to the word-face Stroop paradigm, which has attracted the most attention in the field and was, therefore, the focus of the two previous meta-analyses. Instead, our study attempted to bring together a variety of tasks developed for examining emotional interference with the aim of providing a larger scope and determining inference-related networks across different paradigms.

Although the current findings provide converging evidence for a domain-general neural network involved in emotional and cognitive interference, several important issues remain open. First, contrast analyses revealed that brain regions showing differences in emotional and cognitive interference are adjacent to those regions consistently engaged across both domains. On the one hand, these findings suggest that emotional and cognitive interference processing engage similar regions but to different extents. On the other hand, it is possible that these adjacent regions are engaged in different functions. Indeed, adjacent or overlapping brain activations do not necessarily imply identical functions for all neurons within these regions (Haxby et al. 2001). Further studies are needed to examine the functions of the adjacent/overlapping regions with more fine-grained techniques or specifically designed task paradigms. For instance, advanced imaging techniques (e.g., multi-voxel pattern analysis) provide promising approaches to compare the neural patterns of adjacent/overlapping regions at the sub-voxel level (Norman et al. 2006; Peelen and Downing 2007).

Second, an alternative strategy to test the functional significance of the identified network is to examine the specificity of adaptation of the network across emotional and cognitive interference. That is, it would be interesting to examine whether the presence of emotional interference in the current trial reduces the neural responses (i.e., adaptation effects) of the network in the subsequent trial. Similar paradigms have frequently been used to test the specificity of interference adaptation across different cognitive tasks in the behavioral domain (Braem et al. 2014; Egner 2008, 2014). Further studies are needed to test whether interference adaptation transfers across emotional and cognitive domains at both the behavioral and neural levels.

Third, the current work focused on the contrast of incongruent vs. congruent, which is most frequently reported in the literature. Therefore, our study did not allow for the distinctions between different processes involved in interference processing, such as conflict monitoring and conflict resolution. This distinction requires experimental design with conflict adaptation manipulations, but these have been employed in only few previous studies (e.g., Egner et al. 2008; Etkin et al. 2006, 2010; Etkin and Schatzberg 2011; Chechko et al. 2009, 2014; Jarcho et al. 2013). Employing conflict adaptation manipulations, previous studies have identified domain-specific regions in emotional interference processing, e.g., the amygdala is involved in conflict monitoring (Etkin et al. 2006), whereas the rACC is involved in conflict resolution by dampening activity in the amygdala or dACC (Etkin et al. 2006; Egner and Hirsch 2005). It is thus possible that these emotion-related regions are involved in emotional interference processing but require more sensitive experimental designs to detect them (but see also Chechko et al. 2009, 2014; Krug and Carter 2010). Alternatively, it is possible that the rACC engagement in the emotional interference is further modulated by personality attributes, such as anxiety or depression (Krug and Carter 2010; Rey et al. 2014; Haas et al. 2007).

In either case, previous studies employing adaptation paradigms indicate that emotional and cognitive interference might also depend on domain-specific neural circuits. Indeed, the domain-general network identified in the current meta-analysis does not preclude the possibility that they depend on additional domain-specific processes. This conjecture fits with recent findings that control of emotion and action recruits both common and distinct brain systems (Langner et al. 2018). In line with the current findings, Langner et al. (2018) identified a core network—consisting of the dACC, AI, and SPL—commonly involved in emotion regulation and action control. However, the authors further revealed specific networks for each domain (Langner et al. 2018). Discrepant findings between the present meta-analysis and Langner et al.’s study may be related to the disparity in the topics covered. In the current study, we compared the interference in emotional and cognitive domains, which are similar to each other in terms of topic, tasks, and contrasts (although variations exist). In contrast, the two fields compared by Langner et al. (2018) often employed very different tasks (e.g., “enhancing your emotional experience while viewing the picture” vs. “responding to the color of the word while ignoring meaning of the words”) and stimuli (e.g., arousing affective pictures vs. neutral words), and they concerned related but distinct contrasts (e.g., “enhancing emotion > passive viewing” vs. “incongruent > congruent”). For these reasons, it is not surprising that Langner et al. (2018) revealed more distinct processes between two domains.

Fourth, the limited number of studies for each type of paradigm (e.g., emotional Flanker, emotional Stroop) and stimulus modality (e.g., visual, auditory) did not allow for the examination of neural substrates within each task and modality (e.g., Watanabe et al. 2013), but an accumulating number of neuroimaging studies on emotional interference will enable such an investigation in future meta-analyses.

In summary, our findings indicate that interference induced by emotional stimuli modulates neural mechanisms that strongly resemble those for cognitive interference at both the regional and network levels. These findings coincide with the domain-general hypothesis holding that different types of interference share similar neural mechanisms implemented in domain-general cognitive control systems (Buhle et al. 2014; Ochsner et al. 2009; Ochsner and Gross 2005). Our findings might have significant clinical implications with respect to the critical roles of domain-general cognitive systems in psychiatric distress. That is, our results suggest that cognitive control deficits could account for the etiology and symptoms (e.g., emotion regulation deficits) of a variety of mental illness. In line with this conjecture, a recent meta-analysis of the gray matter volume identified common structural perturbations in the dACC and AI across psychopathologies (Goodkind et al. 2015). In a transdiagnostic meta-analysis of neuroimaging studies on cognitive control tasks (e.g., classical Stroop, Go/No-Go task, working memory), McTeague et al. (2017) identified a common neural circuit disruption across psychiatric diagnoses that parallels the currently observed neural network, including the dACC, lateral PFC, and parietal cortex. The current results complement previous observations by suggesting that disruptions in cognitive control systems could impair not only cognitive functions, but also appropriate control of emotional distress.