Do dynamic features of resting-state fMRI functional connectivity reflect current depressive symptom severity in major depressive disorder?

Resting-state functional connectivity (FC) is closely linked to clinical features of MDD, including diagnostic status (MDD vs. healthy controls), and depressive symptom severity. While prior work has focused on static FC, recent approaches have proposed dynamic FC features as complementary markers of psychopathology-related brain-network organization. Coactivation Pattern (CAP) analysis is one such approach, capturing recurring whole-brain states and their dynamics, such as occurrences, dwell time, and transitions. While seminal literature suggests that these dynamic features reflect current depressive symptom severity, confirmatory evidence from well-powered studies is lacking. We hypothesized that the identified CAPs recapitulate large-scale network patterns from normative resting-state datasets, and that their dynamic features show systematic correlation with current symptom severity.

Hence, we analysed resting-state fMRI data from N = 578 patients with MDD from the FOR2107 cohort. Time series were extracted using a cortical (Schaefer200) atlas, yielding 200 ROIs. After exploring the geometric separability (silhouette), and stability (Hungarian centroid matching) of different cluster solutions, clustering into CAPs was performed by k-means clustering of functional profiles into k = 5 CAPs. To enhance interpretability of CAPs, network-correspondence with the Yeo7 large-scale resting-state network parcellation was determined using the Network Correspondence Toolbox. In a next step, dynamic features were computed, including fractional occupancy, transition probability matrices, and entropy of Markov Trajectories. To account for compositional data structure of fractional occupancy and transition probabilities, isometric log-ratio transformation was performed. Current depressive symptom severity of all patients was assessed using the Hamilton Depression Rating Scale (HDRS). To examine whether dynamic CAP features predicted depressive symptom severity, a two-step modelling approach was used to balance interpretability and predictive performance. First, a linear regression model, controlling for age, sex, medication, and scanner site (Marburg vs. Münster) was adopted. Data-analyses are still on-going: In a second step, we plan to implement a Random Forest Regressor using nested cross-validation, treating the two different sites as discovery- and validation sample.

Preliminary results provided functional CAP profiles that significantly recapitulated large-scale resting-state networks. CAP1 was characterised by significant positive correlation with the Salience/Ventral/ Attention, the Dorsal Attention, and the Somatomotor Network. CAP2 correlated with the Somatomotor, Visual, and Limbic Network, and CAP 3 with the Salience/ Ventral/ Attention, Control, and Default Mode Network. CAP4 was correlated with the Control Network. Lastly, CAP 5 corresponded to the Limbic, Control, and Default Mode Network. Linear modelling of the correlation of dynamic CAP features and depressive symptom severity yielded no significant associations, hence not supporting prior findings. However, Random Forest results will provide a clearer picture, accounting for non-linear relationships and high-dimensional feature interactions.

Despite increasing research interest in the role of dynamic FC features in depressive pathophysiology, the current literature is limited by small samples and methodological heterogeneity. Therefore, it remains unclear which of the previously reported results reflect robust neurobiological signals across different samples while adopting slightly different modelling strategies. Establishing which effects are reproducible is a necessary step toward integrating dynamic FC metrics into neurophysiological models of depression.

References

No references