A dorsomedial prefrontal cortex-based dynamic functional connectivity model of rumination

The study addresses how trait rumination, marked by persistent negative, self-focused thoughts, relates to resting-state brain connectivity and whether a generalizable predictive brain marker can be developed. Prior work implicates the default mode network (DMN), particularly its dorsal medial subsystem and the dorsomedial prefrontal cortex (dmPFC), in internally oriented processes central to rumination. However, findings on static connectivity have been inconsistent, and no precise predictive model applicable to new individuals exists. The authors hypothesize that the temporal variance of dynamic functional connectivity—reflecting stability versus variability over time—will better capture ruminative tendencies than static measures, and that dmPFC-centered connectivity may be key. They set out to build and validate a dynamic connectivity model across multiple subclinical datasets and test its relevance to depressive symptoms in a clinical sample.

Rumination has been linked consistently to altered DMN connectivity, with theories emphasizing its role in self-referential thought, autobiographical memory, and affective processing. The DMN is heterogeneous, with dorsal medial and medial temporal subsystems supporting distinct cognitive functions; the dorsal medial subsystem (including dmPFC) is theorized to underpin high-level reflective appraisal akin to ruminative thinking. Prior reports suggest dmPFC as a "dorsal nexus" modulating depression-related connectivity, with heightened dmPFC connectivity uniquely associated with major depressive disorder (MDD). Yet static connectivity studies show mixed results (both increases and decreases) in MDD. Dynamic features of rsfMRI have yielded generalizable neuromarkers for depression, pain, and attention in other contexts, but predictive models specifically tying dynamic connectivity variance to rumination are lacking. Limited prior work has examined connectivity variance in rumination, motivating a predictive modeling approach focusing on dynamic features.

Design and datasets: The authors used five datasets: three independent subclinical rsfMRI datasets for training, validation, and independent testing (Study 1 n=84; Study 2 n=61; Study 3 n=48), a supplementary subclinical dataset (n=60) collected later, and one clinical dataset of adults with MDD (Study 4 n=35). Behavioral measures: Rumination was assessed via the Ruminative Response Scale (RRS) subscales (brooding, depressive rumination, reflective pondering). Studies 2–3 used the Korean version (19 items). Depression was measured using BDI-II (Studies 1 and 4) and CES-D (Studies 2 and 3). Anxiety was measured with STAI-X (Studies 2 and 3). Imaging acquisition and preprocessing: Study 1 used a 3T Siemens Trio; rsfMRI multi-band EPI (TR=460 ms), T1 MPRAGE normalization; preprocessing in SPM8 included removal of initial volumes, motion correction, MNI normalization, smoothing (8 mm FWHM), band-pass filter (0.008–0.1 Hz), and nuisance regression (spikes, 24 motion parameters, 5 white matter and 5 CSF components). Studies 2–3 used a 3T Siemens Prisma; EPI (TR=460 ms), normalization in SPM12/FSL, smoothing (5 mm FWHM), similar band-pass and nuisance regression; Study 3 additionally regressed thought sampling cues and speaking effects. Study 4 used a 3T Siemens Verio (TR=2500 ms), with SPM12 slice timing correction and comparable preprocessing. Feature construction: Twenty predefined DMN seed regions (including left/right hemispheres; ref. 35) were used. Whole-brain target regions were defined by a modified Brainnetome parcellation (246 original parcels plus 34 additional midbrain/brainstem/cerebellar parcels) totaling 280 ROIs. Dynamic Conditional Correlation (DCC; ref. 36) time series were computed between each seed and each of the 280 ROIs. For each seed-target pair, the temporal variance of the DCC time series was extracted as the feature, yielding 280 features per seed. Modeling: For each of the 20 seeds, separate Lasso regression models predicted each RRS subscale (brooding, depressive rumination, reflective pondering), totaling 60 models (20 seeds × 3 subscales). In Study 1 (training), models used leave-one-participant-out cross-validation, with model performance assessed by Pearson correlation between predicted and actual scores. Significance was determined via one-sided permutation tests (10,000 iterations) and controlled with FDR q<0.05 (p<0.006). The number of selected features per Lasso model was constrained to the sample size (n=84) to avoid additional hyperparameter searches. Model selection and testing: Models significant in training were tested in Study 2 (validation) and, if significant, further tested in Study 3 (independent test) using the same one-sided permutation procedure. Static connectivity comparison: Parallel models using static connectivity features were trained/tested; none survived correction. Robustness analyses: The authors examined training with alternative datasets, combined datasets (Studies 2+3, n=109), and varying numbers of predictors (n_feature=84 vs N_feature=109). Only the dmPFC-based model of depressive rumination consistently generalized when feature counts exceeded ~80. Virtual lesion analysis: To quantify feature importance, each of the 84 non-zero-weight features in the final model was removed one at a time (independent tests), and the drop in prediction performance (Δcorr = r_full − r_reduced) was computed. Regions consistently reducing performance across both test datasets were deemed important; 21 regions met this criterion. Refined model and clinical test: A refined model comprising the 21 important features was applied to Study 4 (MDD, n=35) to predict BDI-II scores, compared with the full 84-feature model. Supplementary subclinical dataset: Two 14-min resting runs before and after an affective movie were acquired; the model’s predictive performance was evaluated for pre- and post-movie resting runs.

• Generalizable marker: Among 60 candidate models, only the dmPFC-based predictive model for the RRS depressive rumination subscale generalized across datasets. Performance: training (Study 1) r=0.342, p=0.001; validation (Study 2) r=0.240, p=0.037; independent test (Study 3) r=0.288, p=0.025 (one-sided permutation; 95% CIs reported in paper). Removing an outlier in Study 3 maintained significance (r=0.276, p=0.028). • Dynamic > static: Models using dynamic connectivity variance features succeeded; models using static connectivity features did not survive testing. • Model composition: The final dmPFC-based model contained 84 non-zero predictive connections (38 positive, 46 negative weights). Positive weights imply higher temporal variance between dmPFC and target region predicts higher depressive rumination; negative weights imply lower variance (more stability) predicts higher rumination. Positively weighted regions spanned subcortical structures (basal ganglia, hippocampus, thalamus), dorsal attention, frontoparietal, and visual networks. Negatively weighted regions included visual cortex, subcortical (thalamus), DMN nodes, and cerebellar regions, indicating whole-brain distribution. • Construct validity: In Study 1, cross-validated predictions correlated with BDI (r=0.269, p=0.007), RRS total (r=0.301, p=0.003), and other RRS subscales. In Study 2, only the target RRS depressive rumination subscale was significant (r=0.240, p=0.037). In Study 3, depressive rumination (r=0.288, p=0.025) and brooding (r=0.316, p=0.014) were significant. • Important features: Virtual lesion analysis identified 21 regions consistently important across both test datasets (overall importance pattern correlation r=0.628, p<0.001). Top contributors included left inferior frontal gyrus (lIFG), right inferior temporal gyrus (rITG), and left cerebellar crus I; other important regions included right IPL and subcortical/brainstem structures. • Clinical relevance: In MDD patients (Study 4, n=35), the full 84-feature model did not predict BDI-II (r=0.150, p=0.391), whereas the refined 21-feature model significantly predicted BDI-II (r=0.431, p=0.010, one-sided permutation). However, the model failed to generalize to three additional MDD datasets (n=21, 57, and 22), likely due to measurement/site differences. • Additional dataset: In a supplementary subclinical dataset (n=60), the model predicted depressive rumination only post-movie (r=0.228, p=0.040) and not pre-movie (r=−0.038, p=0.613), suggesting state induction may enhance detectability. • Feature count sensitivity: Successful generalization required at least a certain number of predictors (≥ ~80) in combined-data training analyses.

The findings support the hypothesis that dynamic connectivity variance, specifically dmPFC-centered interactions, captures core features of ruminative cognition. The dorsal medial subsystem of the DMN appears critical for high-level, self-referential, verbally mediated appraisal processes. Sustained positive connectivity between dmPFC and right IPL/TPJ may reflect persistent social-cognitive evaluations of the self, while variability in dmPFC coupling with regions such as the left IFG (language/appraisal), striatum, and superior parietal areas suggests altered attentional and evaluative dynamics in rumination. More stable dmPFC-visual cortex connections (negative weights) are consistent with perceptual decoupling during internally oriented thought. Identification of 21 key features spanning multiple networks indicates that rumination involves distributed, cross-network dynamics beyond the DMN alone, including frontoparietal and cerebellar systems. The refined model’s ability to predict depressive symptom severity in an MDD sample suggests a continuum connecting subclinical rumination and clinical depression, while failures to generalize across heterogeneous clinical datasets underscore sensitivity to measurement/site factors. Overall, dynamic variance in connectivity provides meaningful information beyond static measures for neuromarkers of rumination.

This work develops and validates a dynamic functional connectivity-based neuromarker of rumination centered on dmPFC interactions. The dmPFC-based model for the depressive rumination subscale generalizes across multiple independent subclinical samples, and a refined 21-feature model predicts depressive symptom severity in one clinical MDD dataset. The study advances mechanistic understanding of rumination by highlighting distributed, cross-network dynamics—especially within the dorsal medial DMN subsystem—and demonstrates the utility of dynamic connectivity variance for prediction. Future research should refine temporal characterization (e.g., frequency/coherence), increase sample sizes and scan durations, harmonize acquisition across sites, incorporate both trait and state measures of rumination, and test task-based paradigms to specify cognitive processes linked to dmPFC dynamics. Such efforts may guide targeted interventions (pharmacological or psychological) aimed at modulating dmPFC-centered networks.

• Seed selection bias: Focusing on DMN seeds may introduce selection/confirmation bias, though it leverages extensive prior literature and improves interpretability. • Scan duration and reliability: Resting-state runs (6–10 min) may yield limited reliability for individual differences. • Effect sizes and sample sizes: Small effect sizes and modest samples raise the possibility of Type I errors; additional datasets produced mixed results. • Temporal metrics: Variance captures magnitude of fluctuations but not frequency or speed of changes; richer temporal metrics (frequency, coherence) and modalities with higher temporal resolution (e.g., MEG) may be beneficial. • Trait vs. state: Models targeted trait rumination; lack of concurrent state measures during scanning limits inference about momentary rumination. • Generalizability sensitivity: Performance depended on training dataset composition and minimum number of predictors; interactions among modeling choices and feature counts are unclear. • Clinical generalization: MDD clinical testing involved a small sample (n=35); the model did not generalize to other clinical datasets with differing acquisition parameters (e.g., phase encoding, manufacturer). Clinical datasets lacked RRS measures, preventing direct rumination prediction in patients. • Cognitive specificity: Resting-state design limits specification of dmPFC-supported cognitive functions; task-based fMRI is needed for finer-grained mapping.