Dynamic reconfiguration of functional brain networks during working memory training

The brain's functional architecture dynamically adapts to environmental demands across various timescales—from seconds during task performance to years during development. This adaptability is facilitated by the brain's modular structure, allowing for flexible switching between segregated and integrated information processing. Prior research suggests that simple, automated tasks rely on segregated brain organization, while complex tasks necessitate integration across multiple subnetworks. However, the impact of automating complex cognitive tasks on network integration remains unclear. Longitudinal fMRI studies, where participants are scanned repeatedly while mastering a task, can illuminate network adaptation during learning and automation. Previous studies on simple motor tasks showed increased network segregation with automation. This study investigated whether mastering a demanding working memory task (dual n-back) affects the balance between network segregation and integration. The researchers hypothesized that effortless performance of the demanding task would lead to increased network segregation, similar to simple motor tasks. They also questioned whether the breakdown of network segregation during challenging cognitive demands remains necessary once the task is automated, and whether stronger separation of cognitive control subnetworks would be observed when tracking dynamic brain network reorganization. To address these questions, participants underwent four fMRI scans while performing an adaptive dual n-back task over six weeks. An active control group performed the same task without the adaptive training to differentiate between training effects and repeated exposure effects.

Several studies have explored the dynamic reconfiguration of brain networks during cognitive tasks. Braun et al. (2015) demonstrated dynamic reconfiguration of frontal brain networks during executive cognition. Vatansever et al. (2015) investigated default mode dynamics for global functional integration. Cohen and D’Esposito (2016) examined the segregation and integration of distinct brain networks and their relationship to cognition. Shine et al. (2016) studied the dynamics of functional brain networks during cognitive task performance. Finc et al. (2017) explored the transition of the functional brain network related to increasing cognitive demands. Bassett et al. (2013, 2015) investigated task-based core-periphery organization and learning-induced autonomy of sensorimotor systems. Mohr et al. (2016) studied integration and segregation of large-scale brain networks during short-term task automatization. These studies, focusing on simpler tasks, suggested that increased network segregation and decreased integration might result from task automation. This study aimed to extend these findings to complex cognitive tasks involving higher-order functions like cognitive control, specifically examining the dual n-back working memory task.

Fifty-three healthy volunteers (26 female; mean age: 21.17) participated. After initial assessments (fluid intelligence testing), participants were randomly assigned to either an experimental group (adaptive dual n-back training) or a control group (non-adaptive 1-back task). Both groups underwent four fMRI scans during a 6-week period: before training (Naive), after 2 weeks (Early), 4 weeks (Middle), and 6 weeks (Late). The dual n-back task involved simultaneous visuospatial and auditory components. The adaptive version of the dual n-back task, used only for the experimental group’s training sessions, adjusted its difficulty level based on participant performance. fMRI data were acquired using a 3 Tesla MRI scanner. Structural and functional scans were obtained using T1-weighted and T2*-weighted sequences. Functional data preprocessing included slice time correction, motion correction, co-registration to T1w, and physiological noise regression (CompCor). Outliers were removed. Functional connectivity was calculated using Pearson’s correlation coefficients (positive correlations only) for a 264-region parcellation (Power et al., 2011) and weighted correlation for the dual n-back task data to control for HRF delays. Network modularity was assessed using a Louvain-like community detection algorithm (Blondel et al., 2008). Modularity was normalized against a null distribution of randomly rewired networks. Dynamic network analysis employed a multilayer community detection algorithm (Mucha et al., 2010) to assess recruitment and integration of 13 large-scale brain systems using functional cartography methods (Mattar et al., 2015). Multilevel modeling was used for statistical analysis.

Behavioral performance (d') significantly improved over training sessions in the experimental group, particularly for the 2-back condition, exceeding the control group's improvement. After training, the experimental group showed no significant performance difference between 1-back and 2-back conditions, suggesting task automation. Whole-brain network modularity was highest during rest, lower during the 1-back task, and lowest during the 2-back task (consistent across groups). However, in the experimental group, modularity significantly increased from the Naive to the Middle and Late sessions, indicating increased network segregation with training. Dynamic network analysis revealed two patterns of change following working memory training. The first pattern, observed in both groups, showed increased recruitment of the default mode system and increased integration between task-positive systems, which correlated with behavioral improvement. The second pattern, unique to the experimental group, involved nonlinear changes in default mode-frontoparietal integration and subcortical system integration. Specifically, increased frontoparietal recruitment and decreased frontoparietal-default mode integration were observed in the experimental group, which correlated positively with behavioral improvement.

The study's findings support the hypothesis that training on a demanding cognitive task leads to increased segregation of task-related functional brain networks. The increase in whole-brain modularity with training in the experimental group, coupled with preserved differences in modularity between 1-back and 2-back conditions, suggests that task automation is associated with enhanced network segregation. The two distinct patterns of dynamic network changes offer a nuanced understanding of how the brain adapts to cognitive training. The first pattern, shared by both groups, reflects general improvement, showing increased default mode autonomy and integration of task-positive systems. The second pattern, specific to the experimental group, represents the effects of task automation, with nonlinear changes in the interplay between default mode, frontoparietal, and subcortical systems. These findings highlight the complex interplay between network segregation and integration during cognitive learning, revealing distinct neural mechanisms underlying general performance improvements and task-specific automation.

This study provides evidence that mastering a complex cognitive task, like the dual n-back, leads to increased brain network segregation. This was reflected in increased whole-brain modularity and changes in the dynamic interplay between large-scale brain systems, notably the default mode, frontoparietal, and subcortical networks. These findings advance our understanding of brain plasticity and network dynamics during cognitive training. Future research could explore the precise neural mechanisms involved, potentially using effective connectivity analyses to clarify the causal relationships between different brain regions. Investigations into individual differences in learning and the time course of network reorganization following training of varying intensities would also be valuable.

The relatively small sample size may limit the generalizability of the findings. The study focused on a specific cognitive task, and further research is needed to examine whether these findings extend to other types of cognitive training. The use of only positive correlations in functional connectivity analyses might have missed important information contained in negative correlations. The cross-sectional nature of the fMRI data limits the ability to draw firm conclusions about the precise temporal dynamics of network reorganization.