Introduction
Rumination is a maladaptive pattern of repetitive and passive thinking in which individuals persistently focus on their negative emotions, symptoms of distress, and the possible causes and consequences of these feelings, rather than engaging in active problem solving or taking action (Bartoskova et al., Reference Bartoskova, Sevcikova, Durisko, Maslej, Barbic, Preiss and Andrews2018; Lyubomirsky & Nolen-Hoeksema, Reference Lyubomirsky and Nolen-Hoeksema1995). Rumination is a risk factor for the development and maintenance of major depressive and bipolar disorder (Ehring, Reference Ehring2021). It is not only a feature of acute depressive episodes but also persists into remitted depression (RD), where elevated rumination prospectively predicts subsequent depressive episodes (Kovacs et al., Reference Kovacs, Takacs, Toth, Simon, Schmelowszky and Kokonyei2020; Michalak, Holz, & Teismann, Reference Michalak, Holz and Teismann2011).
The Response Styles Theory conceptualizes rumination as a habitual response that can be automatically triggered by negative contextual cues and is often difficult to voluntarily control (Hjartarson et al., Reference Hjartarson, Snorrason, Bringmann, Ogmundsson and Olafsson2021; Nolen-Hoeksema, Reference Nolen-Hoeksema1991; Wood & Neal, Reference Wood and Neal2007). Such responding can occur with limited awareness or intent, even when individuals recognize that it is unhelpful (Watkins & Nolen-Hoeksema, Reference Watkins and Nolen-Hoeksema2014). Accordingly, rumination may persist despite symptom reduction and represent a vulnerability process during remission, highlighting the need to understand how individuals disengage from rumination when it is triggered (Hijne, Penninx, van Hemert, & Spinhoven, Reference Hijne, Penninx, van Hemert and Spinhoven2020; Moulds & McEvoy, Reference Moulds and McEvoy2025).
Mindfulness training, which involves cultivating purposeful, curious, nonjudgmental, and non-reactive momentary awareness, is an effective tool for reducing rumination and preventing relapse in patients with RD (Chiesa & Serretti, Reference Chiesa and Serretti2009; Nandarathana, Ranjan, & Sinha, Reference Nandarathana, Ranjan and Sinha2025). From a habit perspective, mindfulness is thought to interrupt the reinforcing dynamics between contextual cues to ruminative responding and to reduce the perceived uncontrollability of rumination (Bolzenkotter, Neubauer, & Koval, Reference Bolzenkotter, Neubauer and Koval2025; Brewer, Reference Brewer2019; Raes & Williams, Reference Raes and Williams2010). To reduce relapse risk, mindfulness-based cognitive therapy (MBCT) trains remitted individuals to decenter from ruminative processing (Kuyken et al., Reference Kuyken, Hayes, Barrett, Byng, Dalgleish, Kessler, Lewis, Watkins, Brejcha, Cardy, Causley, Cowderoy, Evans, Gradinger, Kaur, Lanham, Morant, Richards, Shah and Byford2015; Teasdale et al., Reference Teasdale, Segal, Williams, Ridgeway, Soulsby and Lau2000; Wei et al., Reference Wei, Qin, Yu, Cao and Li2025). However, although it is clinically crucial to understand how the brain supports the application of mindfulness skills in the aftermath of rumination, previous neuroimaging studies have typically examined mindfulness or meditation states in isolation, without embedding them within an explicit rumination-to-mindfulness sequence (Calderone et al., Reference Calderone, Latella, Impellizzeri, de Pasquale, Famà, Quartarone and Calabrò2024; Rahrig et al., Reference Rahrig, Vago, Passarelli, Auten, Lynn and Brown2022). Elucidating this mechanism will provide critical insights into the neural dynamics that enable the transition from ruminative states to mindful awareness, thereby offering a neurobiological foundation for optimizing mindfulness-based interventions and improving relapse prevention in recurrent depression.
In recent years, three functional networks, the default mode network (DMN), frontoparietal network (FPN), and salience network (SN), have emerged as the most consistently implicated systems in RD during ruminative states (Karim, Reference Karim2024; Lydon-Staley et al., Reference Lydon-Staley, Kuehner, Zamoscik, Huffziger, Kirsch and Bassett2019). In RD, rumination has been consistently linked to increase within-DMN connectivity and reduced engagement of frontoparietal control circuitry, together with altered coupling among the DMN, FPN, and SN, consistent with greater cognitive control demands during ruminative thought (Lydon-Staley et al., Reference Lydon-Staley, Kuehner, Zamoscik, Huffziger, Kirsch and Bassett2019; Park, Kuplicki, Paulus, & Guinjoan, Reference Park, Kuplicki, Paulus and Guinjoan2024; Shu et al., Reference Shu, Ou, Ma, He, Zhang, Huang, Chen, Deng, Li, Xi, Meng, Liang, Gao, Peng, Liao, Zhang, Wang, Liu, Liu and Zhang2025; Song et al., Reference Song, Long, Wang, Zhang and Lee2022; Zhou et al., Reference Zhou, Chen, Shen, Li, Chen, Zhu, Castellanos and Yan2020). In contrast, mindfulness practices have been linked to reductions in DMN–FPN and DMN–SN connectivity, and in within-DMN connectivity, accompanied by increased connectivity within the SN, indicating that mindfulness may support the maintenance of task-oriented attention across a range of regulatory demands and diminishing spontaneous, self-generated thoughts about past events (Fox et al., Reference Fox, Dixon, Nijeboer, Girn, Floman, Lifshitz, Ellamil, Sedlmeier and Christoff2016; Prakash et al., Reference Prakash, Shankar, Tripathi, Yang, Fisher, Bauer, Betzel and Sacchet2025; Sezer, Pizzagalli, & Sacchet, Reference Sezer, Pizzagalli and Sacchet2022). However, it remains unclear how these networks reorganize when individuals with RD attempt to shift from a ruminative mode into mindful awareness within the same session. Thus, rather than examining rumination or mindfulness as isolated conditions, the present study focuses on whole-brain network reconfiguration during the transition from deliberate rumination to a mindfulness attempt, a clinically relevant process that may reveal residual vulnerability mechanisms in RD.
In the present study, we used a six-state fMRI paradigm to characterize how large-scale functional networks reorganize as individuals with RD move from baseline rest, through negative autobiographical recall and rumination, into a mindful mode of processing, distraction, and a subsequent resting state. We conceptualized these states as successive phases of a single affective and regulatory sequence, with an a priori focus on the mindfulness state immediately following rumination as the key transition point. To localize sequence-dependent group differences, for each of the six states, we computed whole-brain functional connectivity and compared RD and healthy controls (HCs) using network based statistics (NBS) to identify any connected components showing group differences. We then applied graph theoretical analyses to quantify global topology and nodal centrality across the 246 regions. Finally, motivated by convergent hub effects emerging from these primary analyses, we conducted an exploratory follow-up seed-based similarity analysis centered on the identified parietal hub to quantify rumination-to-mindfulness reconfiguration of its whole-brain connectivity profile.
Materials and methods
Participants
A total of 55 RD (both unipolar and bipolar depression) and age- and sex-matched 41 HCs were recruited at the Second Xiangya Hospital in China from 2021 to 2024. The general inclusion criteria for the two groups were: 1) aged 18–60 years; 2) at least 6 years of education, ensuring the ability to independently complete all scales and assessments; and 3) right-handed. Moreover, RD patients were needed to fulfill the additional following inclusion criteria: Diagnosis of major depressive disorder (MDD) or bipolar II disorder, and now in remission from the last depressive episode, according to the Diagnostic and Statistical Manual of Mental Disorders Fifth Edition (DSM-V) conducted by two attending psychiatrists; And for MDD patients, a total Hamilton Depression Rating Scale-17 (HAMD-17) score ≤ 7 for at least 2 months; For bipolar II disorder patients, both a Young Mania Rating Scale score and a HAMD-17 score maintained at ≤7 for at least 2 months.
Four RD patients were excluded because of excessive head motion (defined as a mean framewise displacement [FD] >0.2 mm) in either the resting or mindfulness states. Ultimately, 51 patients (43 with MDD and 8 with bipolar II) and 41 HC were enrolled in the final study. The study was approved by the Ethics Committee of the Second Xiangya Hospital (approval number: 2021S022) and registered on ClinicalTrials.gov (NCT05585047). Written informed consent was obtained from all participants.
Measures
Rumination was assessed using the Chinese version of the Ruminative Responses Scale (RRS) (Han & Yang, Reference Han and Yang2009; Nolen-Hoeksema & Morrow, Reference Nolen-Hoeksema and Morrow1991), which includes 22 items and three subscales (depression, brooding, and reflection) rated from 1 to 4.
Six-state task
Before scanning, participants provided four distressing autobiographical events using three cue words for each event. During fMRI, they completed six consecutive 6 minutes states in a fixed order: resting state I, negative autobiographical recall, rumination, mindfulness, distraction, and resting state II. The full task instructions for each state are shown in Supplementary Figure S1. These instructions were adapted from the RRS and have been used and validated for effectively inducing rumination (Chen et al., Reference Chen, Chen, Shen, Li, Li, Lu, Zhu, Fan and Yan2020; Jia et al., Reference Jia, Chen, Du, Tang, Ma, Ning, Zou, Zuo, Li, Cui, Deng, Fu, Fu, Huang, Li, Lian, Liao, Liu, Lu and Liu2023). Each state consisted of four sequential 90-s screens. During the two resting states, participants viewed a fixation cross. In the negative autobiographical recall block, participants were presented with the four sets of personalized cue words and were instructed to vividly recall the corresponding negative events. In the rumination block, participants viewed four reflective prompts encouraging elaborative self-focused thinking about the recalled events. In the mindfulness block, participants received a brief standardized in-scanner instruction to observe their thoughts and feelings without judgment, attend closely to the breath, acknowledge that even negative emotions are part of human experience, and notice bodily sensations. No extensive formal mindfulness training was provided prior to scanning. In the distraction block, participants imagined neutral external scenes, such as raindrops sliding down a window, clouds in the sky, the smile of the Mona Lisa, and a small boat sailing on a river. After each state, participants completed a brief 4-point questionnaire on thought content and phenomenology (Supplementary Table S1), which was used as a manipulation check. We derived a negative thought-content index as item 6 minus item 5.
MRI acquisition
Images were acquired using a 3.0 Tesla United Imaging uMR 790 scanner equipped with a 32-channel head coil at the Second Xiangya Hospital. We acquired T1-weighted magnetization-prepared rapid-acquisition gradient-echo structural images (repetition time (TR) = 6.97 ms, echo time (TE) = 3.1 ms, flip angle = 10°, field of view (FOV) = 230 mm × 256 mm, slice thickness = 1 mm, slice gap = 0 mm, voxel size = 1.33 mm × 1 mm × 1 mm, and scan time = 3 minutes and 12 seconds). And we acquired functional magnetic resonance imaging (fMRI) data using T2-weighted echo planar imaging sequences (TR = 700 ms, TE = 30 ms, flip angle = 52°, FOV = 210 mm × 210 mm, slice thickness = 2.5 mm, voxel size = 2.5 mm × 2.5 mm × 2.5 mm, scan time = 6 minutes, simultaneous multi-slice acceleration factor = 7).
Data preprocessing
The T1-weighted structural images were nonlinearly registered to the standard Montreal Neurological Institute space to obtain the transformation matrix. Prior to co-registration with the corresponding structural images, the functional magnetic resonance imaging (fMRI) signals underwent preprocessing steps including slice-time adjustment and motion artifact correction. A series of nuisance covariates were regressed out, which covered linear, quadratic, and cubic drift, 24 motion parameters (six translational parameters, six rotational parameters, and their first-order derivatives), as well as the mean cerebrospinal fluid and white matter signal. Functional data were spatially smoothed using a low-pass Gaussian filter with a cutoff frequency of 0.12 Hz. To project the Fan 246 atlas (Fan et al., Reference Fan, Li, Zhuo, Zhang, Wang, Chen, Yang, Chu, Xie, Laird, Fox, Eickhoff, Yu and Jiang2016), which contains 210 cortical and 36 subcortical regions, into each participant’s native space; we used the inverse transformation matrix derived from T1-image normalization. Cortical region of interest (ROIs) were then assigned Yeo 17-derived network labels and grouped into eight broader cortical networks: DMN (default A/B/C/D), FMN (control A/B/C), dorsal attention network (DAN; dorsal attention A/B), ventral attention network (VAN; ventral attention), SN (salience), visual network (VIS; visual peripheral/central), somatomotor network (SMN; somatomotor A/B), and limbic network (LIM; limbic-1/2) (Fan et al., Reference Fan, Li, Zhuo, Zhang, Wang, Chen, Yang, Chu, Xie, Laird, Fox, Eickhoff, Yu and Jiang2016; Yeo et al., Reference Yeo, Krienen, Sepulcre, Sabuncu, Lashkari, Hollinshead, Roffman, Smoller, Zöllei, Polimeni, Fischl, Liu and Buckner2011). Fan atlas subcortical parcels were summarized separately as a ninth category (SUBC).
Functional network construction
Upon completion of the preprocessing pipeline, the mean time series corresponding to each region of interest (ROI) was retrieved for every participant. Pearson correlation coefficients were calculated between all pairs of ROI time series, resulting in a 246 × 246 correlation matrix per participant. Each of these matrices was composed of 30,258 non-redundant edges, in line with the protocols of previous investigations (DeSerisy et al., Reference DeSerisy, Ramphal, Pagliaccio, Raffanello, Tau, Marsh, Posner and Margolis2021; Dvorak et al., Reference Dvorak, Hilke, Trettin, Wenzler, Hagen, Ghirmai, Müller, Kraft, Reif and Oertel2019). Ultimately, the correlation coefficients were subjected to Fisher’s Z-transformation to achieve normalization. As a sensitivity analysis, the same procedure was repeated using the Shen 268 atlas (Shen, Tokoglu, Papademetris, & Constable, Reference Shen, Tokoglu, Papademetris and Constable2013) without cerebellum nodes, and the results are reported in the Supplementary.
Demographic and clinical characteristics and subjective report analyses
Demographic and clinical variables were analyzed in R 4.3.1 (https://cran.rstudio.com/bin/windows/base/old/4.3.1/). Group differences in continuous variables (e.g. age, years of education, HAMD-17 and RRS total and subscale scores) were examined using independent-samples t tests or Mann–Whitney U tests, depending on distributional assumptions. Categorical variables (e.g. sex) were compared with chi-square tests.
To test whether the paradigm successfully modulated subjective negative thought content, we first conducted a repeated-measures ANOVA on the negative thought-content index, with State (rest I, negative autobiographical recall, rumination, mindfulness, distraction, rest II) as a within-subject factor and Group (RD, HC) as a between-subject factor. When Mauchly’s test indicated a violation of sphericity, Greenhouse–Geisser corrected degrees of freedom and p values were reported. Given our a priori interest in the rumination-to-mindfulness transition, we additionally conducted a planned repeated-measures ANOVA restricted to the rumination and mindfulness blocks, with State (rumination, mindfulness) as a within-subject factor, and Group as a between-subject factor. This was complemented by within-group paired comparisons between rumination and mindfulness and by between-group comparisons within each block.
Network-based statistic
For each participant and each of the six task states, we constructed a 246 × 246 Fisher Z-transformed functional connectivity matrix as described above. Network-based statistics (NBS) were used to identify connected components of edges that differed between RD and HC. Group comparisons were performed separately for each state using NBS V1.2 toolbox (Zalesky, Fornito, & Bullmore, Reference Zalesky, Fornito and Bullmore2010) in MATLAB 2020b.
At each edge, we fit a general linear model in which group (RD vs HC) was the main predictor and age, sex, years of education, and mean FD for that state were included as covariates. The design matrix therefore contained one group regressor and four nuisance regressors, and two group contrasts were specified to test both directions of effect (RD > HC and HC > RD). For each contrast, the resulting edgewise t-statistics were thresholded at a primary threshold corresponding to p < 0.001. Suprathreshold edges that formed connected components were then tested for significance at the cluster level using permutation testing with 5,000 permutations. In each permutation, group labels were randomly shuffled, and the maximal component size (number of edges) under the null hypothesis was recorded. A family-wise error (FWE)-corrected p-value was obtained for each observed component based on its size relative to this null distribution, and components with p_FWE < 0.05 were considered significant.
Graph theoretical analysis
We used GRETNA V2.0.0 toolbox based on MATLAB 2020b (Wang et al., Reference Wang, Wang, Xia, Liao, Evans and He2015) to characterize the topological organization of large-scale functional networks across the six experimental states. For each participant and each state, a 246 × 246 Fisher Z-transformed functional connectivity matrix was first obtained as described above. These matrices were then converted into binary graphs by thresholding at network sparsities from 0.02 to 0.40 in steps of 0.01, ensuring identical edge density across participants at each sparsity level (Watts & Strogatz, Reference Watts and Strogatz1998). On these sparse matrices, we computed a set of global and nodal graph theoretical metrics (Rubinov & Sporns, Reference Rubinov and Sporns2010). At the global level, three metrics were calculated to reflect the functional integration of the whole-brain network, including the global efficiency, characteristic path length, and normalized characteristic path length. Whereas the local efficiency, clustering coefficient, and normalized clustering coefficient were computed to indicate the functional segregation of brain network. The small-worldness index was also computed to describe the balance between functional segregation and integration, defined as the ratio of the normalized clustering coefficient to the normalized characteristic path length. At the nodal level, the degree and betweenness centrality were assessed to index the centrality of each brain region; the nodal efficiency and the nodal shortest path length were computed to indicate the integration of each node, while nodal clustering coefficient and nodal local efficiency were defined to index local clustering of individual nodes.
All graph metrics were computed at every sparsity and, for each participant, state and metric, summarized by calculating the area under the curve across this sparsity range to obtain a single sparsity-independent value for each metric and state. To compute the two normalized metrics (i.e. the normalized clustering coefficient, normalized characteristic path length), 100 degree-preserving random graphs were generated for each participant, state, and sparsity level using a random rewiring procedure. The mean clustering coefficient and mean characteristic path length of these random graphs were then used to normalize the clustering coefficient and characteristic path length of the empirical graphs.
Seed-based connectivity pattern similarity
To quantify rumination-to-mindfulness reconfiguration at the node showing altered topology in the primary analyses, we conducted an exploratory seed-based pattern similarity analysis. For each participant, we extracted the node connectivity profile vector separately for the rumination state and mindfulness state. We then computed Pearson correlation between these two vectors to index similarity of the node connectivity pattern across states, and applied Fisher R to Z transformation.
Statistical analysis
Group differences in graph metrics were assessed separately for each state. For global graph metrics, we fit general linear models with the area under the curve value as the dependent variable, group (RD vs HC) as the main predictor, and age, sex, years of education, and mean framewise displacement for that state as covariates. For nodal graph metrics, the same model was fit at each node and p-values for the group effect were corrected for multiple comparisons across nodes within each state using the false discovery rate.
To further characterize the functional relevance of altered nodes in the graph analysis, we applied the same procedure to nodal graph metrics that showed significant group differences. For each such node, partial correlations were computed between the nodal area under the curve value and RRS total and subscale scores, again separately within the RD and HC groups and controlling for age, sex, years of education, and mean framewise displacement in the corresponding state. P values were corrected for multiple comparisons using the false discovery rate within each set of nodal correlations.
Results
Demographic and clinical characteristics
After exclusion for excessive motion, the final sample comprised 51 RD patients (43 MDD, 8 bipolar II) and 41 HC. Groups did not differ in age, sex, or years of education. As expected, RD patients showed higher residual depressive symptoms and rumination scores than HC at baseline (Table 1).
Demographic information and clinical characteristics
Table 1. Long description
This table compares demographic and clinical characteristics, rumination scores, and head motion data in different states between healthy controls (HC, n = 41) and (RD, n = 51). Age: HC = 21.73 ± 3.10 years, RD = 23.53 ± 5.22 years, P = 0.133. Male/Female ratio: HC = 12/29, RD = 13/38, P = 0.164. Education: HC = 15.10 ± 2.13 years, RD = 14.65 ± 2.29 years, P = 0.332. HAMD-17 score: HC = 1.63 ± 1.79, RD = 3.10 ± 2.29, P = 0.001. Age of onset: RD only = 21.2 ± 5.6 years. Number of episodes [median (P25, P75)]: RD only = 1 (1, 2). Medication status at enrollment (discontinued/not discontinued): RD only = 15/36. Previous psychotherapy history (yes/no): RD only = 3/48. Below are the Rumination Scores (RRS) and its subscales. RRS-Total: HC = 34.85 ± 9.50, RD = 46.78 ± 16.50, P = 0.000. RRS-Brooding: HC = 9.10 ± 3.15, RD = 10.48 ± 4.19, P = 0.076. RRS-Reflection: HC = 8.78 ± 2.67, RD = 11.02 ± 3.63, P = 0.001. RRS-Depression: HC = 16.98 ± 4.93, RD = 24.96 ± 9.73, P = 0.000. Below are the Head Motion (mean FFD, mm) in each tasks: Resting state I: HC = 0.05 ± 0.01, RD = 0.06 ± 0.02, P = 0.124. Autobiographical recall: HC = 0.05 ± 0.01, RD = 0.06 ± 0.03, P = 0.046. Rumination: HC = 0.06 ± 0.01, RD = 0.07 ± 0.03, P = 0.019. Distraction: HC = 0.06 ± 0.01, RD = 0.07 ± 0.03, P = 0.002. Mindfulness: HC = 0.05 ± 0.01, RD = 0.07 ± 0.03, P = 0.013. Resting state II: HC = 0.05 ± 0.03, RD = 0.08 ± 0.05, P = 0.007.
Notes: RRS was a four-point scale (Never = 1, Sometimes = 2, Often = 3, Always = 4) consists of three dimensions: RRS-Brooding (items 5, 10, 13, 15, and 16), and RRS-Reflection (items 7, 11, 12, 20, and 21), RRS-Depression (includes items 1, 2, 3, 4, 6, 8, 9, 14, 17, 18, 19, and 22).
HAMD, 17-item Hamilton Depression Scale; n, number of participants; M, median; P25, 25th percentile; P75, 75th percentiles, standard deviation; x, mean; FFD, Frame-to-frame displacement.
Values in bold indicate statistical significance.
a P-values obtained by Mann–Whitney U test.
b P-values obtained by Pearson’s chi-square test.
c P-values obtained by independent samples t-test.
Task-related subjective ratings
Across states, the negative thought content index differed by State and was higher in RD than HC (State and Group main effects, both p < 0.01), with no State by Group interaction (p = 0.32; Figure 1). Focusing on rumination-to-mindfulness, Negative thought index decreased from rumination to mindfulness (F = 39.71, p < 0.001, ges = 0.12) and remained higher in RD than HC across both blocks (F = 9.31, p = 0.003, ges = 0.07), with no interaction (F = 1.47, p = 0.23). RD showed a higher negative thought index score than HC during rumination (p = 0.003) and mindfulness (p = 0.042).
Negative thought index across the six task states in RD and HC.
Mean negative thought index (item 6 − item 5 from the state questionnaire, higher values indicating relatively more negative than positive thoughts) is shown for remitted depression (RD) and healthy control (HC) groups across the six states: resting state I (Rest1), negative autobiographical recall (NegAuto), rumination, mindfulness, distraction, and resting state II (Rest2). Lines represent group means and error bars denote standard errors of the mean.
Figure 1. Long description
A line graph with the Y axis labeled Negative thought index ranging from minus 1 to 1 and the X axis labeled State with six categories. A legend on the right identifies two groups: H C in blue and R D in red.
Both groups follow a similar trend across the states, but the R D group consistently maintains a higher negative thought index than the H C group at every point.
* Rest 1: R D is near minus 0.4 while H C is near minus 1.2.
* Neg Auto: Both groups peak here, with R D reaching approximately 1.2 and H C reaching approximately 0.6.
* Rumination: R D remains high at approximately 0.8 while H C drops to near 0.
* Mindfulness: R D drops to minus 0.3 and H C drops to minus 0.8.
* Distraction: Both groups reach their lowest points, with R D at minus 1.0 and H C at minus 1.5.
* Rest 2: Both groups show a slight increase, with R D at minus 0.6 and H C at minus 0.7.
Vertical error bars representing S E M are present for each data point.
To improve interpretability, we additionally examined the two constituent items separately. Both positive thought content (item 5) and negative thought content (item 6) showed significant state effects in HC (Q5: F = 7.04, p < 0.001; Q6: F = 23.93, p < 0.001) and RD (Q5: F = 10.31, p < 0.001; Q6: F = 31.72, p < 0.001). As a manipulation check, negative thought content increased significantly from Rest1 to negative autobiographical recall in both HC (t = −6.27, p < 0.001) and RD (t = −4.76, p < 0.001), supporting the effectiveness of the induction phase. From rumination to mindfulness, negative thought content decreased significantly in both HC (t = 4.13, p < 0.001) and RD (t = 5.31, p < 0.001).
Network-based statistic
State-dependent group differences emerged only during mindfulness and the subsequent resting state II. During the mindfulness state, RD showed a hyperconnected component (77 edges, 51 regions, component-level p_FWE = 0.030; Figure 2a). This component was primarily anchored in the left inferior parietal lobule (IPL) and involved connections among regions belonging to the VAN, DAN, FPN, DMN, and SMN networks (Figure 3a). Notably, the left IPL (MNI coordinates −56, −49, 38) exhibited the highest degree in the network, with 28 incident edges. During resting state II, a second hyperconnected component was observed (106 edges, 85 regions, component-level p_FWE = 0.016; Figure 2b), indicating more widespread alterations across large-scale brain networks (Figure 3b). Within this resting state II component, the left IPL again emerged as one of the highest-degree hubs, with 12 altered connections, a degree matched only by a right insular node (MNI coordinates 33, 14, −13, degree = 12) and a right inferior temporal node (MNI coordinates 54, −31, −26, degree = 12). Across both the mindfulness state and resting state II, all altered connections incident on the left IPL linked this DMN hub to nodes in DAN, VAN, and SMN (see Supplementary Tables S4 and S5).
Network-based statistic results (RD > HC).
The figure shows the single supra-threshold connected component in which RD exhibited greater functional connectivity than HC during the mindfulness state (a) and resting state II (b). Node color indicates Yeo 7 network affiliation as shown in the color bar (0 = subcortical or unlabeled regions, 1 = VIS, 2 = SMN, 3 = DAN, 4 = VAN, 5 = LIM, 6 = FPN, and 7 = DMN). Node size is proportional to the number of supra-threshold edges incident on that node within the component, so nodes with more altered connections appear larger. Edges represent functional connections that are significantly stronger in RD than in HC and line thickness reflects the absolute magnitude of the between-group effect. L, R, A, and P denote left, right, anterior, and posterior views, respectively. RD: remitted depression participants; HC: healthy controls.
Figure 2. Long description
Two panels, a and b, each containing eight glass brain views showing a supra-threshold connected component.
Panel a, Mindfulness State. Displays a network of nodes and edges. The largest node is located in the posterior region of the left hemisphere. Edges are concentrated between the left and right hemispheres and within the posterior regions.
Panel b, Resting State I I. Shows a more dense and widespread network compared to panel a. Large nodes are visible in both the left and right hemispheres, particularly in the temporal and parietal regions.
Common features across both panels.
* Node Color. Based on Yeo 7 network affiliation. 0 is subcortical or unlabeled, 1 is V I S, 2 is S M N, 3 is D A N, 4 is V A N, 5 is L I M, 6 is F P N, and 7 is D M N.
* Node Size. Proportional to the number of supra-threshold edges incident on that node.
* Edges. Lines connecting nodes represent significantly stronger functional connectivity in R D than in H C. Line thickness reflects the absolute magnitude of the effect.
* Views. L is left, R is right, A is anterior, and P is posterior.
* Color Bar. A horizontal scale at the bottom of each panel ranging from 0 to 7, transitioning from light green to dark orange.
Between-network distribution of altered functional connections in remitted depression.
(a) Mindfulness and (b) resting state II show 9 × 9 matrices summarizing the number of altered edges within the significant NBS component for RD > HC. Rows and columns correspond to nine large-scale groups: DMN, FPN, LIM, SN, VAN, DAN, SMN, VIS, and SUBC. Cortical ROIs were assigned Yeo 17-derived network labels and grouped into the broader cortical networks shown here, whereas Fan atlas subcortical parcels were summarized separately as SUBC. Each cell indicates the number of edges connecting nodes from the corresponding pair of groups; color intensity and overlaid numbers represent edge counts. RD, remitted depression; HC, healthy controls.
Figure 3. Long description
Two side-by-side 9 by 9 matrices labeled a Mindfulness and b Resting state I I. Both matrices share the same axes representing nine networks: D M N, F P N, L I M, S N, V A N, D A N, S M N, V I S, and S U B C. A color scale at the bottom indicates edge numbers from 0 in white to over 20 in dark blue.
Panel a Mindfulness R D greater than H C:
* S N row: 1 with D M N, 3 with S N.
* V A N row: 8 with D M N, 3 with F P N, 4 with S N, 1 with V A N.
* D A N row: 5 with D M N, 4 with F P N, 1 with S N, 1 with V A N.
* S M N row: 19 with D M N, 8 with F P N, 4 with S N, 4 with D A N.
* S U B C row: 2 with S N, 3 with V A N, 4 with S M N, 2 with S U B C.
Panel b Resting state I I R D greater than H C:
* F P N row: 6 with D M N.
* L I M row: 3 with D M N.
* S N row: 1 with D M N, 4 with F P N, 2 with L I M, 1 with S N.
* V A N row: 5 with D M N, 6 with F P N, 5 with S N.
* D A N row: 8 with D M N, 5 with F P N, 1 with S N, 2 with V A N.
* S M N row: 23 with D M N, 7 with F P N, 4 with S N.
* V I S row: 5 with D M N, 1 with L I M, 1 with D A N.
* S U B C row: 1 with F P N, 8 with L I M, 3 with S N, 4 with D A N.
Nodal graph metrics
Consistent with the network-based results, nodal effects during mindfulness were confined to the same left IPL node. Compared with HC, RD showed lower degree centrality (t = −4.079, p_FDR = 0.0248), lower betweenness centrality (t = −4.077, p_FDR = 0.0249), and lower nodal efficiency (t = −4.196, p_FDR = 0.0162) at this node. No other nodes survived correction in the mindfulness state. Within the RD group, lower IPL degree centrality and nodal efficiency during mindfulness were associated with higher rumination (RRS total: r = −0.33 and r = −0.31; RRS depression subscale: r = −0.35 and r = −0.34; all FDR-corrected) (Figure 4). No significant associations were observed in HC.
Associations between left inferior parietal lobule nodal graph metrics during mindfulness and rumination.
Scatter plots show associations between nodal degree centrality, betweenness centrality, and nodal efficiency of the left inferior parietal lobule node (MNI coordinates −56, −49, 38) during the mindfulness state and rumination scores. Panels (a–c) show associations with RRS total scores, including degree centrality (a), betweenness centrality (b), and nodal efficiency (c). Panels (d–f) show associations with RRS depression subscale scores, including degree centrality (d), betweenness centrality (e), and nodal efficiency (f). Each point represents one participant, with remitted depression (RD) and healthy control (HC) groups plotted separately and fitted with group specific linear regression lines. The y-axes show corresponding raw RRS depression, whereas the reported correlation coefficients are partial correlations controlling for age, sex, years of education, and head motion.
Figure 4. Long description
A multi-panel layout with two rows and three columns.
Top row (panels a, b, c) y-axis: R R S total raw scores ranging from 20 to 80.
Bottom row (panels d, e, f) y-axis: R R S depression raw scores ranging from 10 to 40.
Column 1 (panels a and d) x-axis: Degree centrality (left I P L) from 10 to 40.
Column 2 (panels b and e) x-axis: Betweenness centrality (left I P L) from 0 to 200.
Column 3 (panels c and f) x-axis: Nodal efficiency (left I P L) from 0.18 to 0.27.
Data points:
- H C (Healthy Control): Blue circles with a blue linear regression line showing a near-flat or slightly negative slope.
- R D (Remitted Depression): Red triangles with a red linear regression line showing a steeper negative linear decrease.
Statistical values in top-right of each panel:
- Panel a: R D r equals minus 0.33, p equals 0.025. H C r equals minus 0.16, p equals 0.341.
- Panel b: R D r equals minus 0.17, p equals 0.272. H C r equals minus 0.15, p equals 0.370.
- Panel c: R D r equals minus 0.31, p equals 0.034. H C r equals minus 0.17, p equals 0.312.
- Panel d: R D r equals minus 0.35, p equals 0.017. H C r equals minus 0.15, p equals 0.374.
- Panel e: R D r equals minus 0.18, p equals 0.223. H C r equals minus 0.11, p equals 0.520.
- Panel f: R D r equals minus 0.34, p equals 0.022. H C r equals minus 0.15, p equals 0.383.
During the distraction state, RD showed reduced nodal local efficiency in a single frontal node located in the left middle frontal gyrus. During resting state II, RD showed higher betweenness centrality and nodal efficiency in a node in the left lateral occipital cortex, as well as higher nodal efficiency in the left superior temporal gyrus, right superior temporal gyrus, and left insula (see Supplementary Table S6 for full statistics).
Global graph metrics
At the global level, no significant group differences survived FDR correction in any of the task states other than resting state II. During resting state II, RD showed higher global efficiency and small-worldness, together with altered clustering properties, relative to HC (global efficiency: t = 2.58, p_FDR = 0.037; clustering coefficient: t = −2.35, p_FDR = 0.037; normalized clustering coefficient: t = 2.58, p_FDR = 0.037; small-worldness index: t = 2.43, p_FDR = 0.037).
IPL connectivity pattern similarity from rumination to mindfulness
To quantify the rumination to mindfulness transition at the IPL hub, we computed IPL seed-based connectivity pattern similarity between the rumination state and mindfulness state connectivity profiles. After adjusting for age, sex, education, and state-specific motion (FD in rumination and mindfulness), RD showed lower pattern similarity than HC (Group effect: β = −0.151, SE = 0.0477, t = −3.17, p = 0.002).
Discussion
The present findings indicate that network abnormalities in RD are not uniformly expressed across mental states, but emerge most clearly at a clinically relevant transition point, namely when individuals attempt to move from deliberate rumination into a mindful mode of processing. Although the relative negative thought-content index decreased from rumination to mindfulness in both groups, RD continued to report relatively more negative than positive thought content. Across NBS, graph theoretical analyses, and connectivity profile similarity, the clearest group effects during mindfulness converged on a left posterior IPL node that showed broader coupling with distributed systems, lower nodal centrality and efficiency, and less preservation of its whole-brain connectivity profile from rumination to mindfulness. Together, these findings suggest that, even during remission, RD is characterized by a persistent alteration in state-dependent network control, expressed as inefficient and less-consistent hub reconfiguration during mindfulness attempts following rumination, which may represent a neural mechanism of ongoing ruminative vulnerability.
This pattern is consistent with cognitive inhibition accounts of rumination, which propose that depressed individuals have difficulty in removing no-longer-relevant negative information from ongoing mental activity and disengaging attention from mood-congruent material (Kaiser et al., Reference Kaiser, Snyder, Goer, Clegg, Ironside and Pizzagalli2018; Rnic et al., Reference Rnic, Battaglini, Jopling, Tracy and LeMoult2023; Yaroslavsky, Allard, & Sanchez-Lopez, Reference Yaroslavsky, Allard and Sanchez-Lopez2019). Previous work indicates that depression and trait rumination are associated with impaired updating of internally maintained information in the presence of irrelevant negative stimuli and with difficulties suppressing or discarding negative material that should be ignored (Bruning, Mallya, & Lewis-Peacock, Reference Bruning, Mallya and Lewis-Peacock2023; Roberts et al., Reference Roberts, Schreiner, Pocius, Dillahunt, Farstead, Feldman, Bessette, Kaufman, Slattery, Jacobs, Jago, Crowell, Watkins and Langenecker2024). In the present paradigm, mindfulness immediately followed deliberate rumination on negative autobiographical material, a transition that likely places strong demands on shifting attention away from elaborative self-referential processing and toward a nonjudgmental monitoring stance (Brewer, Reference Brewer2019; Raes & Williams, Reference Raes and Williams2010; Watkins & Nolen-Hoeksema, Reference Watkins and Nolen-Hoeksema2014). Within this framework, the persistence of more negative thought content in RD together with IPL-centered network inefficiency and reduced rumination-to-mindfulness reconfiguration may be consistent with less-efficient implementation of control processes that support disengagement and updating after rumination.
Our findings suggest that abnormalities were instantiated in an IPL-centered subnetwork linking a DMN-anchored posterior parietal node to other large-scale systems. The hub identified here corresponded to Brainnetome ROI 141, an anatomically defined posterior IPL parcel in the A40c/PFm territory (Fan et al., Reference Fan, Li, Zhuo, Zhang, Wang, Chen, Yang, Chu, Xie, Laird, Fox, Eickhoff, Yu and Jiang2016). Tractography-based parcellation of the human left IPL has shown that PFm constitutes a distinct posterior subregion with characteristic anatomical and resting-state connectivity patterns, supporting a heteromodal role at the interface of self-referential and attentional processes (Wang et al., Reference Wang, Fan, Zhang, Liu, Jiang, Zhang, Yu and Jiang2012). More broadly, the IPL and adjacent temporoparietal cortex have been described as major network hubs involved in multimodal integration and higher-order cognition (Igelstrom & Graziano, Reference Igelstrom and Graziano2017), and functional work further indicates substantial specialization within IPL subregions across attentional, semantic, and social domains (Numssen, Bzdok, & Hartwigsen, Reference Numssen, Bzdok and Hartwigsen2021). In the present data, altered edges incident on this node extended predominantly to somatomotor and pericentral regions, together with dorsal attention, ventral attention or salience related, insular, and cingulate parcels, consistent with the IPL and adjacent temporoparietal regions acting as heteromodal hubs at the intersection of default mode, attention, and control systems (Braga, DiNicola, Becker, & Buckner, Reference Braga, DiNicola, Becker and Buckner2020; Igelstrom & Graziano, Reference Igelstrom and Graziano2017; Morawetz et al., Reference Morawetz, Riedel, Salo, Berboth, Eickhoff, Laird and Kohn2020). This region has been implicated in reorienting attention to salient events and in integrating internally generated and externally driven information (Heise et al., Reference Heise, Albouy, Dolfen, Peeters, Mantini and Swinnen2025; Jiang, Wang, & Wan, Reference Jiang, Wang and Wan2022; Numssen, Bzdok, & Hartwigsen, Reference Numssen, Bzdok and Hartwigsen2021), as well as in self-related and social cognition (McCalley et al., Reference McCalley, Sanderson, Dowdle, Padula, Siddiqi, Ellard and Taylor2025). Non-invasive stimulation studies further show that modulating IPL or neighboring angular gyrus can influence DMN coupling with executive systems and internally oriented thought (Eldaief, Halko, Buckner, & Pascual-Leone, Reference Eldaief, Halko, Buckner and Pascual-Leone2011; Jargow et al., Reference Jargow, Zwosta, Korb, Ruge and Wolfensteller2021). Within dual network models of cognitive control, FPN and cingulo-opercular circuits initiate, adjust, and maintain task sets over time (Dosenbach et al., Reference Dosenbach, Fair, Miezin, Cohen, Wenger, Dosenbach, Fox, Snyder, Vincent, Raichle, Schlaggar and Petersen2007; Friedman & Robbins, Reference Friedman and Robbins2022; Wood & Nee, Reference Wood and Nee2023; Yu et al., Reference Yu, Han, Wu, Wei, Zhang, Ding and Wen2023). In this context, our findings suggest that control operations relevant to disengaging from rumination may require coordination through an IPL hub that interfaces between default mode and attentional and sensorimotor systems during mindfulness attempts.
The findings also show partial continuity with prior work on mindfulness-based treatment and psychotherapy. In recurrent depression, van der Velden et al. reported that MBCT altered salience-network connectivity during depressive rumination, and that this neural change mediated improvement in the ability to sustain and control attention to body sensations (van der Velden et al., Reference van der Velden, Scholl, Elmholdt, Fjorback, Harmer, Lazar, O’Toole, Smallwood, Roepstorff and Kuyken2023). Reviews of mindfulness-based interventions similarly implicate insular and cingulate regions, particularly in relation to interoception, salience processing, and emotion regulation, while meta-analytic evidence on psychotherapy highlights treatment-related change in anterior cingulate or paracingulate and insular regions (Marwood, Wise, Perkins, & Cleare, Reference Marwood, Wise, Perkins and Cleare2018; Young et al., Reference Young, van der Velden, Craske, Pallesen, Fjorback, Roepstorff and Parsons2018). The convergence with the present study lies in the broader circuitry rather than in the location of the dominant hub. Our significant component also involved insular and cingulate parcels, indicating overlap with salience-related and bodily monitoring systems. The key distinction concerns the timescale and inferential scope of the neural processes examined. Prior studies examined training-related neural plasticity after weeks of structured intervention, whereas the present paradigm captured an acute, untrained attempt to shift from rumination to mindful monitoring. Under these conditions, the most robust abnormality was localized to a posterior IPL hub within the DMN, suggesting that the transition from self-referential elaboration to mindful monitoring may be constrained by inefficient parietal hub reconfiguration.
More broadly, depression and rumination have been associated with altered connectivity within and between default mode and control and attention systems (Marchitelli et al., Reference Marchitelli, Paillère Martinot, Trouvé, Banaschewski, Bokde, Desrivières, Flor, Garavan, Gowland, Heinz, Brühl, Nees, Papadopoulos Orfanos, Paus, Poustka, Hohmann, Holz, Vaidya and Fröhner2024; Wei et al., Reference Wei, Dong, Ding, Luo, Wang, Baeken and Wu2024; Zhang et al., Reference Zhang, Zhang, Wang, Lei, Jiang, Xiong, Chen, Zhang, Zhao, Wang, Zhang, Xu, Zhai, An, Li, Hao and Liu2025), and mindfulness training is linked to altered DMN coupling (Rahrig et al., Reference Rahrig, Vago, Passarelli, Auten, Lynn and Brown2022). A key feature of the present results is the state dependence and symptom relevance of the IPL effects. Although NBS revealed distributed alterations across large-scale networks, the most robust and rumination-linked effects converged on the IPL during mindfulness. This specificity is consistent with the possibility that a DMN-anchored parietal hub involving the IPL may play a central role in the transition from rumination to mindfulness in remission (Sun et al., Reference Sun, Wang, Zheng, Wang, Wang and Xue2025).
Alterations also persisted into the post-task resting state, where RD showed a more widespread hyperconnected component and global graph differences consistent with greater integration and altered clustering. Post-task resting connectivity is known to retain traces of preceding cognitive or affective states, potentially reflecting consolidation or continued processing of recent experience (Gaviria et al., Reference Gaviria, Rey, Bolton, Delgado, Van De Ville and Vuilleumier2021; Linke et al., Reference Linke, Haller, Xu, Nguyen, Chue, Botz-Zapp, Revzina, Perlstein, Ross, Tseng, Shaw, Brotman, Pine, Gotts, Leibenluft and Kircanski2023; Tambini, Ketz, & Davachi, Reference Tambini, Ketz and Davachi2010). One interpretation is that RD does not fully return to baseline rest after the internal mentation and regulation sequence, but remains in a partially reconfigured state. However, this account remains provisional because we did not measure cognitive control processes during resting state II and the post-task effects were less tightly linked to rumination scores than the mindfulness-state IPL metrics.
These findings have potential clinical implications. First, they suggest that even in remission, individuals with depression may rely on a less-efficient parietal hub configuration when attempting to regulate ruminative mentation with mindfulness strategies, consistent with the persistence of ruminative vulnerability beyond symptomatic recovery (de Klerk-Sluis et al., Reference de Klerk-Sluis, Huijbers, Locke, Spijker, Spinhoven, Speckens and Ruhe2022; Wang, Ren, Gao, & Yu, Reference Wang, Ren, Gao and Yu2022). From a biomarker perspective, IPL-centered connectivity and nodal efficiency during mindfulness and early post-task rest may index a latent vulnerability dimension related to ruminative style in remission, beyond residual symptom severity. If replicated in larger longitudinal samples, such measures could help identify RD patients who remain at higher risk for relapse despite symptomatic improvement. Second, our findings raise the possibility that neuromodulation targeting parietal hubs might complement existing prefrontal-focused approaches for reducing rumination and preventing relapse. Non-invasive stimulation of IPL and neighboring angular gyrus has been shown to modulate connectivity within and between DMN and control networks and to influence internally oriented cognition (Williams et al., Reference Williams, Numssen, Guerra, Kopal, Bzdok and Hartwigsen2024). Combining such stimulation with MBCT or other interventions that explicitly train disengagement from rumination may help normalize the inefficient hub configuration observed here and enhance the implementation of regulatory strategies in daily life. Finally, the multi-state paradigm used in the present study may provide a useful platform for future mechanistic trials, in which IPL-centered network indices during the rumination-to-mindfulness transition are measured before and after treatment or are used as potential targets in real-time neurofeedback or closed-loop stimulation designs.
Several limitations should be acknowledged. First, because the present study was cross-sectional, the observed differences in network reconfiguration cannot establish causal relationships between IPL-centered abnormalities, symptom improvement, or vulnerability to future relapse. Although we controlled for state-specific mean framewise displacement and excluded participants with excessive motion, residual motion-related confounding may not be fully ruled out. The task states were presented in a fixed order to model a clinically meaningful progression from autobiographical recall to rumination and then to a mindfulness attempt. However, this design does not permit complete separation of state-specific effects from carryover or passage-of-time effects. Relatedly, the present study did not directly assess state rumination or participants’ adherence to the rumination instructions. Therefore, variability in the extent to which participants engaged in the intended rumination process cannot be entirely ruled out. Participants received only brief mindfulness instructions, and we lacked more detailed measures of strategy implementation or state engagement. Thus, the findings reflect neural organization during attempts to adopt a mindful stance rather than the effects of extensive mindfulness training. The post-state ratings were brief in-scanner measures of thought content and phenomenology, but they may have been affected by recall bias and may not have fully captured moment-to-moment fluctuations in attention, including episodes of mind wandering during the instructed states. Future studies could use online thought probes or button-response markers to capture such fluctuations more directly. We also did not include independent behavioral measures of inhibitory control or cognitive shifting, limiting direct tests of the proposed mechanism. Finally, the present sample size was moderate for whole-brain network-level inference, and replication in larger samples and in longitudinal studies tracking depressive episodes or mindfulness training will be important to confirm the stability and generalizability of the IPL-centered findings (Poldrack et al., Reference Poldrack, Baker, Durnez, Gorgolewski, Matthews, Munafò, Nichols, Poline, Vul and Yarkoni2017).
In conclusion, this six-state fMRI paradigm showed that the neural signature of RD becomes most evident at the moment a ruminative mode must be disengaged and replaced by a mindful mode of processing. During this transition, RD expressed a broad increase in coupling among default mode, attention, control, and somatomotor systems that was anchored in the left IPL, yet the same node showed reduced centrality and efficiency and a less-preserved connectivity profile from rumination to mindfulness. This combination is consistent with a more diffused but less-optimized coordination pattern during regulatory effort. The resulting IPL-centered transition signature may therefore represent a candidate neural marker of persistent ruminative vulnerability during remission. Future longitudinal studies tracking network changes across depressive episodes, remission, and mindfulness-related training will be required to determine whether this signature reflects a stable vulnerability marker, a modifiable regulatory process, or a consequence of prior depressive episodes.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0033291726106126.
Data availability statement
Data are available from the corresponding authors upon reasonable request.
Acknowledgments
We would like to thank all the patients who participated in our study.
Author contributions
Conceptualization: H.H., K.L., and Y.J.; Data curation: S.S., M.M., A.D., F.M., Y.P., S.G., and H.L.; Formal analysis: H.H. and K.L.; Funding acquisition: S.S., Y.Z., and Y.J.; Investigation: S.S., M.M., A.D., F.M., Y.P., S.G., and H.L.; Original draft: H.H. and K.L.; Resources: Z.L. and Y.J.; Supervision: Z.L. and Y.J.; Writing - review and editing: Z.W., M.W., J.L., B.L., Z.L., and Y.J.
Funding statement
This work was supported by the STI2030-Major Projects (2021ZD0202000 to Yan Zhang), the National Natural Science Foundation of China (82101612 and 82471570 to Yumeng Ju), the Science and Technology Innovation Program of Hunan Province (2024RC3056 and 2025JK2137 to Yumeng Ju), the Natural Science Foundation of Hunan Province (2025JJ60832 to Su Shu), and the Scientific Research Launch Project for new employees of the Second Xiangya Hospital of Central South University. The sponsors had no role in the study design, data collection and analysis, interpretation of the results, and writing of the report.
Competing interests
The authors declare no conflicts of interest.
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation and with the Helsinki Declaration of 1975, as revised in 2008.
Open access
