Resting-state BOLD temporal variability in sensorimotor and salience networks underlies trait emotional intelligence and explains differences in emotion regulation strategies

A converging body of behavioural findings supports the hypothesis that the dispositional use of emotion regulation (ER) strategies depends on trait emotional intelligence (trait EI) levels. Unfortunately, neuroscientific investigations of such relationship are missing. To fill this gap, we analysed trait measures and resting state data from 79 healthy participants to investigate whether trait EI and ER processes are associated to similar neural circuits. An unsupervised machine learning approach (independent component analysis) was used to decompose resting-sate functional networks and to assess whether they predict trait EI and specific ER strategies. Individual differences results showed that high trait EI significantly predicts and negatively correlates with the frequency of use of typical dysfunctional ER strategies. Crucially, we observed that an increased BOLD temporal variability within sensorimotor and salience networks was associated with both high trait EI and the frequency of use of cognitive reappraisal. By contrast, a decreased variability in salience network was associated with the use of suppression. These findings support the tight connection between trait EI and individual tendency to use functional ER strategies, and provide the first evidence that modulations of BOLD temporal variability in specific brain networks may be pivotal in explaining this relationship.


Discussion
ER processes are daily employed by individuals to modify the emotional states they are experiencing. However, the individuals' variability to use one or other strategy is still poorly understood. How can such variability be explained? In the present study, we investigated the relationship between trait EI and ER abilities, on the basis of previous evidence that trait EI may represent a suitable measure for the individual predisposition to the choice of specific ER strategy (e.g. 18,25 ). Namely, we were interested in finding the neural bases associated with trait EI by analysing the functional connectivity of naturally grouping circuits decomposed by an unsupervised machine learning approach (ICA). One intriguing question was whether EI and ER share at least some neural bases. This would be an additional proof of their intimate relationship.
In terms of trait measures, we found that individuals with low levels of trait EI are associated with the use of non-adaptive ER processes (i.e., suppression, and self-blame). This result is complementary to past evidence 18,25,67 showing a relationship between high trait EI and adaptive coping strategies. The different, still complementary result can be explained by the fact that the previous studies (e.g. 25 ) focused mostly on coping abilities and not on typical emotion regulation strategies. Instead, by the wide range of strategies considered, we found that low scores in the trait EI subscale of self-control are associated with blaming others and rumination, whereas low scores in the well-being subscale are associated with the use of suppression, rumination, and catastrophizing. Finally, low scores in emotionality are associated to the use of suppression and rumination.
In literature, rumination and suppression strategies have been usually related to negative outcomes and psychopathologies 5 . Likewise, low level of traits EI are also associated with psychopathological disorders 68,69 , making reasonable the strong link we found between low trait EI and maladaptive strategies. The specific association between rumination and self-control, which is associated with difficulties in managing stressful situations and impulsive behaviours, is supported by recent evidence suggesting that impulsivity plays a critical role in rumination 70 . Low self-esteem-a relevant aspect in the trait EI subscale of well-being, already emerged as an important predictor of rumination 71 , and it has been indirectly linked to suppression since it is involved in shame emotion, which has been already coupled with this regulation strategy 56,72 . In addition, low emotionality and the related difficulty in emotions recognition and expression 62 , may lead individuals to use suppression that occurs late in the process of emotion generation 73 . In our study, that trait EI subscale of sociability did not predict any emotion regulation strategy, can be explained by evidence showing that low sociability is not associated with indexes of emotion dysregulation (e.g., high psychological reactivity, negative emotional intensity, dispositional negative affect, and personal distress). Low sociability is rather associated with low social support seeking 74 , Table 3. Regions identified in the independent component IC7, IC20, and IC16. Clustering indicates the spatial distribution of ROIs at the cortical and subcortical levels. Cluster threshold at p < 0.05 (pFDR corrected) and voxel threshold p < 0.001 (pFDR corrected, two sided). For each cluster we selected the brain regions with the highest covering proportion (%) of Harvard-Oxford Atlas ROI. Peak are reported in MNI coordinates. L left, R right, BA Brodmann area. www.nature.com/scientificreports/ strategy which may have escaped the taxonomy used in the ERQ and CERQ. In support of that, trait EI predicts social sharing when this latter is included in the set of regulation strategies 28 . Our result also showed that women are more likely to use suppression with respect to men. Although also previous studies reported gender difference in the use of suppression [75][76][77] ), this result is not consistent in literature (see 6,78 ), suggesting that gender difference possibly depend on other factors such as the situation itself 79 Further researches are needed to better understand the role of gender in the trait EI-ER relationship. Besides trait measures, functional connectivity results showed that modulations of BOLD temporal variability in sensorimotor (IC20 and IC16), visual (IC18), salience (IC7) and cerebellar (IC1) networks is associated with the total traits EI index and the four subscales. Most importantly for the present study, we found that increased BOLD variability especially in sensorimotor (when identified by the IC20 but not IC16) and salience networks also predicted the use of reappraisal strategy (in ERQ and CERQ), whereas decreased BOLD variability in salience network predicted the use of emotion suppression. These results suggest that BOLD temporal variability in sensorimotor and salience networks underlies trait EI and at the same time explains differences in ER strategies. Namely, this finding enriches our understanding of the relationship between trait EI and ER, by showing for the first time that different networks involved in the former, are also involved in the latter. The brain areas belonging to IC20 have been already related to the sensorimotor network 80 . Among these areas, the inferior frontal gyrus also emerged in previous resting-state study on trait EI 46 and it was interpreted www.nature.com/scientificreports/ as a part of a circuit related to social and emotional processing. Similarly, a positive correlation between r-state connectivity involving frontal regions and trait EI was reported by Takeuchi et al. 48 . On the other hand, the supplementary motor cortex (SMA) and the cerebellum positively correlated with trait EI 46 , playing a critical role in cognitive control mechanisms. In particular, the somatosensory cortex, and its portion of the supramarginal gyrus (SMG) are involved in the recognition of emotions, the understanding of the emotional states of others 81 , and are more generally part of the mirror neuron system 82,83 . In addition, other studies provided evidence that the somatosensory cortex is associated with emotion generation 84 and interoceptive awareness 85 . The insula, instead, is deemed to facilitate social interaction, and decision making by integrating sensory, affective, and bodily information, and it has been traditionally reported as neural correlate of trait EI. The implications of these areas in socio-emotional processing and cognitive control 46,86 make intuitive their involvement also in reappraisal strategy, which involve cognitive abilities (e.g., attention and memory) to control emotional responses 87 .
Consistently, there is evidence of a correlation between the frequency of use of reappraisal and the functional connectivity of the left insula, supplementary motor cortex (SMA) 49 , and inferior frontal gyrus 88 . These areas, more generally, represent a well-established network underlying cognitive control of emotions by reappraisal 34,38 .
Our result concerning the negative correlation between the BOLD variability in the IC16-related sensorimotor network and the positive reappraisal seems apparently incongruent, then. However, this incongruency may be better interpreted in line with our trait measures results, which show no association between the trait EIsociability scale and ER strategies. Similarly, indeed, that the IC16 (but not the IC20) is associated with the sociability scale may explain why this component yield a negative correlation with reappraisal. In addition, it is worth mentioning that the spatial match to template approach reports the IC20 (vs the IC16), as the best match for the sensorimotor network (see correlations in the Results section), and, yet, the two sensorimotor-related components involve overlapping, but still different brain regions, such as the middle temporal gyrus. Since this is the first attempt showing common neural substrates between trait EI and ER, further researchers are needed to better identify the specific involved brain areas. Along the sensorimotor network, our study showed the involvement of the salience network, the BOLD variability of which positively correlates and predicts adaptive (reappraisal) strategy, and negatively correlates with maladaptive (suppression) strategy. In line with our results, previous studies reported the link between a similar component and the salience network 89 , as well as the relationship between the key nodes of this network and trait EI and ER 90 . Especially the amygdala and the basal ganglia (i.e., putamen), allowing the organism to adaptively respond to the emotional context, are strictly related to the emotion processing 91 and as a such to the two constructs 48,49 . In addition, there is evidence that emotion suppression is associated with activation of anterior cingulate, ventrolateral prefrontal cortex 92 , inferior frontal gyrus, putamen, pre-supplementary motor area and supramarginal gyrus 93 .
The relation between BOLD variability, trait EI and adaptive vs maladaptive emotion regulation strategies, however, can be better understood coming back to what the BOLD variability means. Several studies point out that greater BOLD variability positively influences cognitive performance 94 , fluid intelligence 95 and most importantly for our results, interoceptive awareness 96 , adaptability, flexibility and efficiency of neural system in response to the multiplicity and uncertainty of environmental stimuli 54,56,97,98 . As a such, it is reasonable that cognitive and affective mechanisms related to the functionally connected regions are better implemented by individuals showing increased temporal variability in the network 96 . Accordingly, our findings suggest that the increased BOLD variability in the sensorimotor and salience networks play a critical role in predicting both high level of trait EI and adaptive emotion regulation strategy, in terms of a better social and emotional information integration, selfawareness along with a more efficient cognitive control. That temporal variability in these networks significantly predicts high traits EI and the frequency of use of Cognitive Reappraisal strategy is explained as an adaptive feature of the neural response, allowing the brain to easily access different "states", required to complete cognitive tasks 55,99 . By the same token, we could also infer that less variability in salience network could imply a difficulty of the subjects to process emotional information resulting in the maladaptive emotion suppression strategy. Then, greater temporal variability may represent a neural predisposition marker which facilitates individuals in the stages underlying the dynamic process of emotional regulation, identification, selection, and implementation 100 . This finding provides a context for and corroborates the hypothesis that regulation strategies and their outcome may depend on factors such as the individual differences. Neural flexibility and adaptability increase perception and control of the emotional event determining at the same time the success of an emotion regulation process 7 . Importantly, resting-state functional connectivity may be helpful to investigate task-independent constructs 101 such as those related to personal traits.
Besides these new findings, the study has some limitations to point out. While the data-driven approach allowed us to consider the activity of the whole brain and the role of naturally grouping circuits, theory driven analyses (i.e., dynamic causal model) that may facilitate inferences with respect to specific brain regions, or to identify causal relationships between them, may be a valuable and complementary alternative. Moreover, we acknowledge that a larger sample size (including thousands of participants) would have been ideal for increasing reproducibility and statistical power, as a recent paper suggested 102 . For what concerns the discussed networks as identified by the spatial match, they also included portions of executive networks. Future studies are needed to explore the contributions of such networks to both EI and ER. Finally, building on the existing literature, we focused on trait EI. However, it would be worthy not only to investigate other aspects of EI 18 , but also to extend the knowledge on the interaction between affective mechanisms and personality, considering the variability this latter may show across different situations 103 .
To conclude, the present findings reveal that the role of trait EI in predicting adaptive ER strategies relies on a shared and more efficient functional connectivity network involved in social and emotional information processing to understand self and others' affective states, and in higher cognitive mechanisms which contribute to the control of emotions. Consequently, our study not only further strengthened the association between low www.nature.com/scientificreports/ traits EI and maladaptive ER strategies but is also represents a first step to understand the neural mechanisms able to explain this relationship. Increased variability of the BOLD signal within a sensorimotor and salience network is a mainstay for the neural structure of high traits EI and at the same time predisposes to the use of adaptive emotional regulation strategy. By contrast, a decreased variability in salience network predisposed to the use of a maladaptive emotion regulation strategy.

Methods
Participants. The data analysed in this study were selected from the open-source dataset "Max Planck Institute Leipzig Mind-Brain-Body Dataset LEMON" 104 . All subjects were recruited by researchers at the University of Leipzig, in Germany, between 2013 and 2015, and data were collected in accordance with the Declaration of Helsinki for a study which protocol was approved by the ethics committee at the medical faculty of the University of Leipzig. For the present study, we extracted a subset of participants representing a young adult healthy population. Based on the information provided by the authors, indeed, we defined the sample by the following inclusion criteria: no substance use or abuse (negative at the Multi 8/2 Drogen-Tauchtest) 104 , and no past or present psychopathologies diagnosis, screened by the SCID-I, and by the Hamilton Depression Scale (HAM-D 105 ). The final subset was composed of 79 subjects (23 females; age range: 20-35 years; mean education: 12.39 years). For each participant, we extracted raw data from structural MRI scans (T1 Weighted-MP2RAGE) and functional MRI scans (rs-fMRI). With regards to trait measures, the scores of the following self-administered questionnaires were selected: TeiQUE-SF (Trait EI Questionnaire-Short Form), ERQ (ER Questionnaire) and CERQ (Cognitive ER Questionnaire).
The cognitive emotion regulation questionnaire (CERQ 108 ) was administered in German validated version 63 and consists of 36 items divided into 9 scales that measure nine strategies defined as adaptive: acceptance  Table 4 for a summary of sample's descriptive statistics for each scale and subscale).

Trait measures analyses.
To test whether trait EI predicts ER processes, we implemented two different analyses using SPSS Statistics for Windows, version 25.0 (SPSS Inc., Chicago, Ill., USA). In the first analysis we implemented a multivariate linear regression (MLR) with ERQ and CERQ questionnaires as dependent variables, while the total trait EI was included as a predictor. Moreover, to assess the effect of every subscale of trait EI, we next implemented a multivariate multiple linear regression (MMLR) with each subscale of the ERQ and CERQ questionnaires as dependent variables, and the four factors of the TeiQue-SF as predictors. Variance inflation factor (VIF) and Pearson's correlation among the TEIQue-SF subscales were used in order to examine multi-collinearity and relative association in the regression model. Gender was used as a categorical fixed factor to test its effect in the regression model. The type I error was controlled by applying false discovery rate (FDR) correction to p-values.
Neuroimaging analyses. Pre-processing and functional connectivity analysis were conducted using CONN MATLAB Toolbox (version 18b) 109 . Firstly, we implemented CONN's default pre-processing pipeline using SMP12 default parameters which includes the following steps: functional realignment and unwarping, translation and centering, functional outlier detection (conservative settings), functional direct segmentation and normalization (1 mm resolution), structural translation, and centering, structural segmentation and normalization (2.4 mm resolution), functional and structural smoothing (spatial convolution with Gaussian kernel 8 mm). Next, the denoising phase was implemented. The objective of this phase is the identification and elimination of confounding variables and artefacts from the estimated BOLD signal. Briefly, these factors are derived from three different sources (BOLD signal coming from white matter or cerebrospinal fluid masks, parameters and outliers defined in the pre-processing step, and an estimate of the pre-processing the subjects' motion parameters) 110 . Once identified, the factors are entered into a regression model (Ordinary Least Squares) as covariates. Finally, a 0.0008-0.09 Hz temporal band-pass filter standard for resting-state connectivity analyses was applied to the time series. Next, the functional connectivity analysis has been implemented. For this study, we chose to use a data-driven approach by implementing a group-independent component analysis (group-ICA). The group-ICA implemented by CONN includes the following steps: pre-conditioning variance normalization, concatenation of the BOLD signal along the temporal dimension, dimensionality reduction at the group level, fast-ICA for spatial component estimation, and the back-projection for spatial estimation on the individual subject 110 . The number of independent components to be identified was set to 20 as software CONN suggests as default, and in line with previous studies using low model order analysis [111][112][113] . In order to separate noise components from the underlying resting-state networks, every identified IC were visually inspected and compared with CONN's networks atlas using spatial match-to-template function. This feature measures the overlap between eight brain networks (Default Mode Network, Sensorimotor, Visual, Salience, Dorsal Attention, Frontoparietal, Language, Cerebellar; defined from CONN's ICA analyses of HCP dataset/497 subjects) and the IC's spatial map associated with each individual network component. One out of 20 ICs (IC17) did not allow for the delineation of specific areas due to its extent and it was discarded from the following analyses. We then extracted the temporal variability of each remaining IC's, calculated in CONN as SD of each BOLD time-series 110 . Type I error was controlled using cluster-size-based false discovery rate (FDR) correction (p < 0.05, voxel thresholded at p < 0.001 114 tab, within each analysis). Next, to assess the relationship between IC's temporal variability and both trait EI and ER, we implemented 2 different analysis by using SPSS Statistics for Windows, version 25.0 (SPSS Inc., Chicago, Ill., USA). Firstly, to address which of the 20 identified IC's predicted the trait EI, we tested the individual explanatory variables effect (IC's BOLD variability values) on the TEIQue-SF factors and total index by using a Multiple Linear Regression model (Ordinary Least Squares) with a stepwise method (forward) for each dependent variable, and gender as a categorical fixed factor in order to test its effect in the regression model. Since we do not expect that all the identified components were related to the investigated construct, we chose a method of fitting regression models in which the choice of predictor variables is made by an automatic procedure. This methodology consists of testing the incremental predictivity of the model: starting from a model with no predictor, each explanatory variable is added to the model and compared to the inclusion or exclusion threshold criterion (in our case predictor's p-value ≤ 0.05 for inclusion) until the model reaches its maximum predictivity. Finally, the BOLD temporal variability of IC's that resulted to be significant predictors of trait EI in the previous analysis were entered a Multivariate Multiple Regression (MMR) as independent variables to predict ER scores (ERQ and CERQ subscales) and gender as a categorical fixed factor in order to test its effect in the regression model. To avoid multiple comparisons issues, type I error was controlled applying false discovery rate correction (FDR) within each analysis.

Data availability
The complete LEMON Data can be accessed via Gesellschaft für wissenschaftliche Datenverarbeitung mbH Göttingen (GWDG) https:// www. gwdg. de/. Raw and preprocessed data at this location is accessible through web browser https:// ftp. gwdg. de/ pub/ misc/ MPI-Leipz ig_ Mind-Brain-Body-LEMON/ and a fast FTP connection