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

Abstract

Rumination is a cognitive style characterized by repetitive thoughts about one's negative internal states and is a common symptom of depression. Previous studies have linked trait rumination to alterations in the default mode network, but predictive brain markers of rumination are lacking. Here, we adopt a predictive modeling approach to develop a neuroimaging marker of rumination based on the variance of dynamic resting-state functional connectivity and test it across 5 diverse subclinical and clinical samples (total n = 288). A whole-brain marker based on dynamic connectivity with the dorsomedial prefrontal cortex (dmPFC) emerges as generalizable across the subclinical datasets. A refined marker consisting of the most important features from a virtual lesion analysis further predicts depression scores of adults with major depressive disorder (n = 35). This study highlights the role of the dmPFC in trait rumination and provides a dynamic functional connectivity marker for rumination.

Fig. 1. Analysis overview and prediction results.

a For the model development, we first predefined 20 seed regions within the DMN based on ref. . We then calculated the Dynamic Conditional Correlation (DCC) between each seed region and 280 Brainnetome-based parcels using rsfMRI data from 84 participants. Using the variance of DCC time-series data as input features, we trained predictive models of the Ruminative Response Scale (RRS) subscales. The “B” stands for the brooding subscale, “D” for the depressive rumination subscale, and “R” for the reflective pondering subscale. We used Lasso regression with leave-one-participant-out cross-validation. We then selected and tested only good-performing models on the next independent test datasets. b Among the initial 60, we selected seven predictive models that showed significant cross-validated prediction performance (q < 0.05, false discovery rate) in the training dataset (n = 84). Among the seven predictive models, we again selected one predictive model that showed significant independent prediction performance at p < 0.05 (one-sided permutation test) with the validation dataset (n = 61). The selected model was the dmPFC-based predictive model of depressive rumination. We finally tested the model on the last independent test dataset (n = 48) to evaluate the model’s generalizability. A red-dashed circle indicates the data point that was identified as an outlier (i.e., greater than three standard deviations away from the mean), which did not affect the significance after its removal (r = 0.276, p = 0.028, one-sided permutation test, 95% CI [−0.012, 0.579]).

Fig. 2. The dmPFC-based predictive model of depressive rumination and its predictions.

a The final model that showed generalizable prediction performances across two independent datasets was the dmPFC-based predictive model of the depressive rumination subscale of the RRS. The top panel shows the brain regions that had positive predictive weights and their large-scale network assignments. Positive weights mean the higher the DCC variances between the dmPFC and the regions, the higher the depressive rumination scores. The bottom panel shows the brain regions that had negative predictive weights and their large-scale network assignments. Negative weights indicate the higher DCC variances between the dmPFC and the regions, the lower the depressive rumination scores. The color bars represent the sign and magnitude of the model weights. The percentage on the pie chart indicates the proportion of region assignments to each large-scale network. A percentage of less than 5% is not shown in the graph. b To evaluate the divergent and convergent construct validity of our final model, we examined the correlations between the model prediction and other self-report questionnaires. The superscript k indicates a study that used the Korean translation version for the questionnaires. The questionnaires that showed significant correlations (p < 0.05, one-sided permutation test) were marked with green circles.

Fig. 3. Virtual lesion analysis results of the final model.

a Each block represents one brain region out of the 84 regions included in the final model. We grouped them into 10 large-scale functional networks. The color of the blocks indicates the difference in the prediction-outcome correlation between the full model and the reduced model after removing (i.e., virtually lesioning) the brain region from the model. Positive differences indicate that the region is important for the prediction. The boxes with red outlines show the regions with positive difference values across both independent test datasets. There were 21 important brain regions for the prediction. Note that the overall patterns of the region importance were similar across two independent test datasets (r = 0.628, p = 1e-10, two-sided, 95% CI [0.520, 0.956]). b The brain map shows the 21 brain regions important for the prediction across two test datasets. Mean difference values across two datasets were shown. The top three most important regions included the left inferior frontal gyrus (lIFG), right inferior temporal gyrus (rITG), and left cerebellum (lCb).

Fig. 4. Twenty-one important features of the predictive model.

a Circos plot showing 21 important regions identified with the virtual lesion analysis. The region names outside of the circos plot were from the Brainnetome nomenclature. The outer circle represents the color coding of large-scale networks, while the inner circle represents the mean values of DCC connectivity (equivalent to static connectivity). The line color inside the circos plot indicates the sign of the predictive weights based on DCC temporal variance (i.e., red: positive, blue: negative), while the line thickness indicates the magnitude of the predictive weights. b Relationship between the predictive weights and the DCC mean connectivity values of 21 important regions. c The scatter plot shows the testing results of the model on Study 4 clinical data from people with MDD (n = 35). It shows a significant correlation between the model prediction and the BDI-II score of r = 0.431 (p = 0.010, one-sided permutation test, 95% CI [0.115, 0.808]).