Disrupting The Resting State: Meta-Analytic Evidence That Mindfulness Training Alters Default Mode Network Connectivity


 This meta-analysis sought to expand upon neurobiological models of mindfulness through investigation of inherent brain network connectivity outcomes, indexed via resting state functional connectivity (rsFC). We conducted a systematic review and meta-analysis of rsFC as an outcome of mindfulness training (MT) relative to structurally-equivalent programs, with the hypothesis that that MT would increase cross-network connectivity between nodes of the Default Mode Network (DMN), Salience Network (SN), and Frontoparietal Control Network (FPCN) as a mechanism of internally-oriented attentional control. Texts were identified from the databases: MEDLINE/PubMed, ERIC, PSYCINFO, ProQuest, Scopus, and Web of Sciences; and were screened for inclusion based on experimental/quasi-experimental trial design and use of standardized mindfulness-based interventions. RsFC effects were extracted from twelve studies (mindfulness n = 226; control n = 204). Voxel-based meta-analysis revealed significantly greater rsFC (MT > control) between the left middle cingulate (Hedge’s g = .234, p = 0288, I2 = 15.87), located within the SN, and the posterior cingulate cortex, a focal hub of the DMN. Egger’s test for publication bias was nonsignificant, bias = 2.17, p = .162. In support of our hypothesis, results suggest that MT targets internetwork (SN-DMN) connectivity implicated in the flexible control of internally-oriented attention.


Introduction
Human waking life contains many moments in which the mind is not engaged by external goals or tasks, but is instead absorbed in a state of stimulus-independent thought (SIT) (Christoff et al., 2004;Mason et al., 2007), commonly referred to as "mind wandering" and "resting state activity". Experience sampling studies suggest that SIT is often self-re ective and focuses on the evaluation of past experiences and the simulation of future events (Diaz et  . Such studies support the involvement of candidate interacting brain networks-including the default mode network (DMN), salience network (SN), and the frontoparietal control network (FPCN)--through which mindfulness may facilitate the exible allocation of attentional resources between introspective and perceptual processes (Dixon et al., 2018). However, there remains little consensus about how mindfulness alters functional connectivity within and between these networks.
The current meta-analysis sought to update brain-based models of mindfulness by comprehensively examining mindfulness-driven resting state functional connectivity (rsFC) outcomes.

The Neurocognitive Features of Stimulus Independent Thought (SIT)
When measured as functional connectivity, changes in neural circuitry re ect strengthened or weakened coordination between regions and, at a larger scale, between networks of interest (Van Dijk et al., 2010;Buckner et al., 2013). Analogous to the spontaneous ow of thought characteristic of resting states, the brain at rest spontaneously emits blood oxygen level dependent (BOLD) uctuations that self-organize into cohesive neural networks (Raichle et al., 2001;Fox & Raichle, 2007) that are commonly termed 'resting-state' or 'intrinsic' functional networks (Snyder & Raichle, 2012). Although there is little consistency in the taxonomy of intrinsic functional networks (for review see Uddin et al., 2019), it is generally accepted that a minimum of 10 networks are observable during the resting state (Damoiseaux et al., 2006). Among recognized large-scale networks, the DMN, SN, and FPCN have been frequently implicated in explanatory frameworks of SIT. The DMN has received the most attention historically for its role in internally-directed mentation (Andrews-Hanna et al., 2014;Raichle, 2015 According to the dynamic framework of mind-wandering (Christoff et al., 2016), internal experiences--maintained by the DMN--may be deliberately or automatically constrained via coordination with the FPCN and SN, respectively. The FPCN, characterized by dorsolateral PFC and anterior inferior parietal structures (Uddin et al., 2019), is well-known for its integral role in cognitive control processes (Zanto & Gazzaley, 2013). Studies combining experience sampling with neuroimaging suggest that the FPCN exibly couples with the DMN to deliberately regulate internal mentation (Cole et al., 2013), and that such coordination is used to inhibit internal thought when external attention is required (Spreng et al., 2010;Gao & Lin, 2012).
In contrast, the SN has been theorized to support automatic constraints on internal experiences through cognitively e cient cross-network signaling (Christoff et al., 2016). Key nodes of the SN include the midcingulate cortex (MCC) [also known as the dorsal anterior cingulate cortex (dACC) (

Mechanisms of Mindfulness: Theory and Empirical Support
Mindfulness, commonly de ned as the act of attending to present-moment thoughts, emotions, and sensations without judgment or appraisal (Brown & Ryan, 2003), is relatively unique as a treatment of maladaptive SIT. Unlike popular cognitive behavioral therapies meditation trains the practitioner to focus on and maintain attention to a neutral sensory object (e.g., the breath), and direct attention back to that object when the mind begins to wander (Lutz et al., 2008). This recursive process of shifting and sustaining attention has previously been linked to enhanced functional connectivity within the FPCN of experienced meditators , suggesting that FA improves top-down cognitive control needed to disengage from distracting thoughts and emotions.
Alternative models of mindfulness posit that mindfulness may indirectly regulate SIT by promoting awareness of internal experiences (Vago & Silbersweig, 2012). According to this framework, the sustained concentration conferred by mindfulness facilitates awareness (or mindful meta-awareness), de ned as the capacity to observe one's mental patterns with a sense of equanimity and psychological distance (Lutz et al., 2008;Dunne et al., 2019). This internal awareness thereby supports recognition of thoughts and feelings as discrete mental states, and in turn, improves exible, adaptive responding (Lutz et al., 2008). The cultivation of mindful awareness has been theoretically attributed to enhanced functional cohesion of networks linked to self-awareness (e.g., default mode network) and attention monitoring Although insightful, such studies typically suffer from low statistical power, a limitation endemic to research relying on high-cost neuroimaging modalities such as fMRI. Addressing such concerns, meta-analytic approaches may be used to pool information from well-controlled studies while modeling convergence of effects across pooled samples. Thus, we conducted a systematic review and meta-analysis of rsFC outcomes of mindfulness training relative to structurally-equivalent programs (i.e., active controls). To test the neuroplastic changes associated with mindfulness skills--namely, executive functioning and meta-awareness--we hypothesized that 1) mindfulness training would enhance rsFC between the FPCN and DMN as an indicator of enhanced cognitive control; and 2) mindfulness training would enhance rsFC between the DMN and SN, re ective of meta-awareness.

Results
Study characteristics and participant demographics are displayed in Table 1. Overall, studies used standardized mindfulness-based interventions and structurally equivalent control interventions ranging from 3 days to 8 weeks in length. Study samples varied in terms of age (M = 45.80; SD = 13.15) and clinical characteristics with the majority of studies recruiting healthy adults (see Table 1). The majority of individuals from the pooled sample identi ed as female (n = 286, 62.72%) followed by male (n = 170, 37.28%). No studies reported data from trans or non-binary participants. Of the 12 included studies, only 5 reported racial/ethnic demographic information. From this subsample of 152 participants, 100 (65.79%) identi ed as white, 31(20.39%) as Black/African American, 12 (7.89%) as Mixed Race/Other, 5 as Hispanic/Latino (3.29%), 3 (1.97%) as Asian, and 1 as Southeast Asian (0.66%).
Seed regions were pooled according to standardized resting state network location. This process demonstrated that eligible studies used seed regions from four networks, the default mode network (DMN; n = 11), the midcingulo-insular network (M-CIN; n = 7), the dorsal attention network (DAN; n =1), and the frontoparietal control network (FPCN; n =2) (see Supplementary Table S1).
SDM meta-analysis was used to test for signi cant training condition effects. Meta-analysis results identi ed one signi cant cluster of 57 voxels located in the paracingulate gyri (Table 2, Figure 2), Hedge's g = .234, p = . 0288, I 2 = 15.87, suggesting that mindfulness training, relative to control training programs, increased connectivity to bilateral median cingulate regions and the left anterior cingulate (nonsigni cant cluster outcomes reported in Supplementary Table S2). Yeo's cortical parcellation atlas (2011) places the peak coordinates of this cluster (0,20,34) in the ventral attention network (VAN), more recently taxonomized as the mid-cingulo insular network (within the anatomical domain) or salience network (within the cognitive domain) (Uddin et al., 2019). Egger's test for publication bias was nonsigni cant; bias = 2.17, p = .162, and funnel plots did not suggest the in uence of small study effects (Figure 3).
Examining the original articles revealed increased functional connectivity to the median cingulate via the posterior cingulate cortex (PCC), a region canonically situated within the default mode network (DMN). Notably, studies within the sample exclusively reported increased connectivity between median cingulate effect regions and PCC seed regions. Therefore, the ndings suggest that mindfulness training increased resting state functional connectivity between the default mode network and salience network.

Discussion
This meta-analysis is the rst to systematically examine the effects of mindfulness-based training on resting state functional connectivity (rsFC), a neural marker of cognitive and emotion regulation. A systematic review of the literature revealed that rsFC has been sparsely investigated as a target of mindfulness training, with 12 studies meeting the eligibility criterion in the current review. Meta-analysis results partially supported our hypotheses, indicating that relative to training in one or another structurally equivalent (active control) program, mindfulness training increased functional connectivity to the midcingulate cortex (MCC), and that such connectivity was seeded to the posterior cingulate cortex (PCC). Results did not support our rst hypothesis, which predicted enhanced cross-network connectivity between the FPCN and DMN as a mechanism of cognitive control. In support of our second hypothesis, these results suggest that mindfulness training strengthened cross-network connectivity between regions associated with the default mode network (DMN) and salience network (SN). Implications for null and supported ndings are explored in the following sections.
Along with the frontoparietal control network (FPCN), the DMN and SN operate synergistically to support self-regulation, and aberrant coordination within and between these networks has been associated with emotional dysfunction, de cits in attentional control, and maladaptive stimulus-independent thinking (SIT) (e.g., rumination, worry, intrusive thought) ( The duration and therapeutic focus of mindfulness interventions likewise warrant consideration. Although the mindfulness interventions were standardized and well-matched against active control programs, these studies report a wide range of intervention duations (3 days -16 weeks) with different degrees of intensity (i.e., retreat verses remote-delivered trainings) and therapeutic focus (e.g., PTSD, parentfocused stress reduction). There is currently little research examining the dose-response relation for mindfulness-based treatments; however, a recent meta-analysis of this topic suggests that brief trainings may be equally as effective as longer interventions for the treatment of stress, depression, and anxiety (Strohmaier, 2020). Nevertheless, the relation between mindfulness training dose and neuroplasticity remains poorly understood, a matter that is further complicated when applied to clinical and aging populations with Interpretation is likewise limited by a priori seed selection, which was requisite for all included studies (see protocol from Kaiser et al., 2015). This study failed to detect signi cant FPCN cross-network connectivity, and while this null effect may stem from erroneous theoretical assumptions, it may instead be the consequence of preferencing seeds within the DMN. Behavioral neuroimaging studies have previously shown strong functional coupling between the FPCN and DMN across multiple tasks (e.g., Farb et al, 2007, Spreng et al., 2010Fleming & Dolan, 2012), and one study recently demonstrated how robust co-activation of FPCN and DMN regions may obscure detection of group effects (Dixon et al., 2021). Further research is needed to de nitely determine if and how mindfulness training interacts with FPCN neurocircuitry during task-related and resting brain states.
Finally, it warrants noting that there is currently no standard classi cation system for large-scale functional networks. The lack of universal nomenclature presents a signi cant barrier to interpreting neural outcomes, especially given the multitude of naming schemes (Uddin et al., 2019) and inconsistent application of common labels (see meta-analysis of executive control network topography; Witt et al., 2021). We use the functional network terms "frontoparietal control network", "default mode network", and "salience network" primarily due to their ubiquity in cognitive neuroscience (Uddin et al., 2019) and the mindfulness literature. Nevertheless, such naming conventions based on functional properties are problematic for the reasons described above.We recommend that future research move towards more transparent network taxonomies incorporating anatomical properties (see Uddin et al., 2019).
This meta-analysis highlights the potential value of rsFC as a window into understanding the mechanisms of mindfulness. However, research on this topic is in its early stages, and considerably more research is necessary to qualify speci c rsFC effects as mechanistic targets of mindfulness training. Without additional research, the effects of mindfulness on resting state cognition remains speculative. Researchers may consider novel sampling methods including lab-based phenomenological reporting (e.g., ) and ecological momentary assessment, a method used to capture day-to-day lived experiences (Berkman & Falk, 2013; Gillan & Rutledge, 2021). Investigators should additionally consider examining multiple representations of connectivity--including effective connectivity, white matter connectivity, and dynamic connectivity--which probe different features of brain network organization (e.g., directional effects; temporal dynamics, etc.). The comparison of such representations has the potential to reduce ambiguity and improve interpretation of connectivity-based effects (Bijsterbosch et al., 2020).
Further research is also needed to determine the relevance of rsFC plasticity for different psychiatric conditions. The successful use of network neuroscience for clinical diagnosis and treatment is a lofty goal, which will require overcoming considerable methodological challenges (see Woo  Studies were excluded if they met one or more of the following criteria: (1) experimental interventions predominantly featured training elements other than mindfulness meditation (e.g., yoga, transcendental meditation, loving-kindness or compassion meditation, tai chi; integrative body-mind training); (2) the study design lacked an active control condition or used a cross-sectional design; (3) resting state functional connectivity (rsFC) group contrasts were not reported; (4) seed-based rsFC methods were not used or seed-based rsFC indices were not reported.
The systematic literature review identi ed 7,092 records, from which 7,071 were excluded upon abstract review (see PRISMA ow diagram in Figure 1

Data Extraction and Coding
The meta-analysis was coordinate-based (for example see Kaiser et al., 2015), in which extracted coordinates re ected locations of signi cant group differences in resting state functional connectivity across time. Given that all studies used seed-based rsFC analyses, coordinates were categorized as belonging to either seed anatomy or effect anatomy. Using this coding scheme, 11 sets of seed coordinates were extracted, with coordinates re ecting each seed anatomy's reported center of mass. In the instance that seed anatomy coordinates were not reported because the seed region was de ned from a standardized atlas or subject-speci c spatial map, the center of mass was estimated using meta-analytic maxima reported from the open source platform, neurosynth.org (Yarkoni et al., 2011). All studies reported peak coordinates of signi cant between-group effects, yielding a total of 65 effect coordinates. After extracting seed and effect coordinates, all coordinates were categorized into rsFC networks as de ned from a standardized network cortical parcellation atlas (Yeo et al., 2011). Effects were likewise characterized by direction of effect with the term hyperactivation re ecting relatively greater rsFC effects in the mindfulness group (MT > CT), and hypoactivation indicating relatively greater rsFC in the control group (CT > MT). In addition to main effects, demographic data and intervention characteristics were extracted for review. Effects were extracted by four independent reviewers who collected data from independent reports.

Interrater Reliability
We conducted a two-way random-effects ICC modeled with absolute agreement. Speci cally, we tested for signi cant effects of rater ID on each quantitative measure. ICC estimates were within acceptable range (ICC > .6; p < .005), indicating a high degree of rater agreement. Next, unweighted Cohen's kappa was calculated to examine reliability of extracted categorical variables. Results indicated substantial agreement (Altman, 1999;Landis & Koch, 1977) between the raters' judgements, k = .741 (95% CI .182 to .884). Finally, z statistics were converted to two-sample t-test statistics, assuming equal variances in both conditions. If records did not report z statistics or two-sample ttest statistics, p values were used to estimate t-test statistics used in the meta-analysis.

Voxel-based meta-analysis
Voxel-based meta-analysis was conducted using seed-based d mapping (SDM-PSI version 6.21) with permutation of subject images (Albajes-Eizagire et al., 2019b). Unlike traditional meta-analytic approaches which test for the presence or absence of peaks of statistical signi cance (i.e., null hypothesis testing), seed-based d mapping (SDM) uses multiple imputation to model effect sizes for each study before conducting a random-effects meta-analysis. Notably, SDM has the advantage of controlling false positive and false negative rates through subject-based permutation tests (Winkler et al., 2014) and threshold-free cluster enhancement (Smith & Nichols, 2009).
SDM preprocessing was rst used to convert t-values of peak coordinates into Hedge's g effect sizes. Study-level images of upper and lower bounds of probable effects were then constructed for all voxels using multiple imputation (Rubin, 2004; Albajes-Eizagirre et al., 2019a) with anisotropic Gaussian kernels. Using SDM we then calculated most likely effect size and standard error via multiple imputations of Maximum Likelihood Estimation (MLE) with a jackknife procedure. Finally, FWE corrections were performed via subjectbased permutation testing and TFCE-corrected effect sizes were calculated to estimate group differences. Main analytic ndings were scrutinized for small study effects and excess signi cance (i.e., publication bias) through Egger's tests and examination of funnel plots.

Declarations Author Contributions
The authors con rm contribution to the paper as follows: study conception and design: HR, KWB; data extraction and curation: MP, AA, NL; analysis and interpretation of results: HR; KWB, DRV; draft manuscript preparation: HR, KWB, DRV; study conception and design: HR, KWB  Figure 1 The top gure portrays the location of the PCC ROI seed. The bottom gures show a 3D map of voxelwise z-scores, with signi cant cluster effects localized to the bilateral paracingulate gyri. Figure 2