Dynamic reconfiguration of functional brain networks during working memory training

The brain’s functional architecture adapts over multiple time scales to meet environmental and cognitive demands. Modularity enables switching between segregated and integrated processing, but it remains unclear whether the integration typically observed during demanding tasks reflects task complexity per se or the lack of automation. Prior work shows increased integration during high-load tasks relative to rest or simple motor tasks, suggesting costly global workspace engagement. Yet, when simple motor tasks are trained to automaticity, increased segregation and subsystem autonomy have been observed. The central question is whether a complex cognitive task, once mastered, can be executed within a more segregated network, reducing the need for costly integration. To test this, the study examined how mastering a demanding working memory task (dual n-back) over 6 weeks alters the balance between segregation and integration, whether demand-related modularity changes persist after automation, and whether autonomy and interactions among systems central to working memory (frontoparietal and default mode) dynamically shift with training.

Evidence indicates that modularity decreases (integration increases) as cognitive demand rises during n-back tasks and that simple motor tasks maintain segregated organization. Longitudinal studies of visuomotor training report increased autonomy and segregation in task-relevant systems (e.g., motor and visual; default-mode), suggesting automation may foster segregation. However, these findings come from relatively simple tasks. The literature on higher-order cognitive control implicates frontoparietal and default-mode systems in both competitive and cooperative interactions that vary with task demands. The Global Workspace Theory posits that complex tasks engage integrated networks, while automated tasks rely more on segregated modules. The gap addressed here is whether intensive training on a complex working memory task shifts functional organization toward segregation and how large-scale systems dynamically reorganize during automation.

Design and participants: Fifty-three healthy right-handed adults (18–28 years) were recruited; 46 completed four fMRI sessions over 6 weeks and met quality criteria. Participants were matched by sex and randomized post-baseline to experimental (adaptive dual n-back training) or active control (single-modality 1-back) groups (n=23 each; no baseline differences in age or fluid intelligence). Sessions occurred at Naive (pre-training), Early (2 weeks), Middle (4 weeks), and Late (6 weeks). Ethics approval and written informed consent were obtained. Training/intervention: Experimental group completed 18 supervised sessions (30 min each) of adaptive dual n-back (auditory letters and visuospatial locations simultaneously). Difficulty increased when ≥80% correct, decreased when error >50%. Control group completed 18 sessions of single-modality 1-back (auditory or visuospatial, alternating order). In-scanner task: At each fMRI session, all participants performed a dual n-back with alternating 1-back and 2-back blocks (20 blocks total; 30 s/block; 12 trials per block; 25% targets). Instruction screens preceded blocks (4 s). Participants responded with lateralized button presses for auditory and visual targets. Resting-state scans (10:10) were acquired at the start of each session. Behavioral measures: Primary outcome was d′ (signal detection theory), combining sensitivity and bias; values computed per subject, session, and condition (averaged across modalities). Penalized reaction time (pRT) served as a secondary measure. MRI acquisition: GE 3T MR750 with 8-channel coil. T1w FSPGR BRAVO (1 mm isotropic). BOLD EPI: TR=2000 ms, TE=30 ms, flip=90°, FOV=192 mm, matrix 64×64, 42 slices, voxel 3×3×3 mm, 0.5-mm gap; 5 dummy scans. Task run: 11:30 (340 volumes). Resting-state: 10:10 (305 volumes). Preprocessing: BIDS structured. fMRIPrep 1.1.1 for slice timing correction (AFNI 3dTshift), motion correction (FSL MCFLIRT), BOLD–T1w coregistration (FreeSurfer bbregister), normalization (ANTs), and CompCor (a/tCompCor). Framewise displacement computed. Denoising via Nilearn/Nistats: regressors included six aCompCor components, 24 motion parameters (with derivatives and quadratics), outlier volumes (FD >0.5 mm; DVARS >±3 SD) and derivatives, task effects and derivatives, and linear trends; band-pass 0.008–0.25 Hz. Four high-motion participants excluded (mean FD >0.2 mm and >10% outlier volumes in any session). Parcellations and connectivity: Primary functional parcellation: Power 264 ROIs; robustness tested with Schaefer 300-ROI parcellation. Static functional connectivity (FC) computed as Pearson correlations between ROI mean time series, retaining positive correlations, Fisher z-transformed. For task FC, a weighted correlation accounted for HRF by convolving block regressors, filtering positive segments, and weighting concatenated block time series accordingly. Static matrices computed for rest, 1-back, and 2-back. Dynamic FC computed per block, yielding 20 layers per session. Static modularity: Louvain-like optimization of Newman–Girvan modularity Q with configuration null model. To control for network strength dependence, modularity values were normalized by the mean modularity from 100 degree-preserving rewired null networks. Dynamic modularity and cartography: Multilayer community detection (generalized Louvain; Mucha et al.) across the 20 task blocks (layers). Parameters: structural resolution γ set per prior work; interlayer coupling ω=1 within same condition, ω=0.5 across conditions. For each subject and session, 100 runs were performed to build module allegiance matrices P (fraction of runs×layers two nodes co-assigned). Recruitment for each of 13 large-scale systems computed as within-system mean of P; integration as between-system mean of P. To remove size bias, recruitment/integration were normalized by permutation-based nulls (1000 random ROI-to-system label permutations). Statistics: Multilevel models (MLM) accounted for nesting (trials within sessions within participants where appropriate). For behavior (d′), predictors were group, condition, session, and their interactions. For baseline modularity (Naive), condition was the predictor. For training effects on static modularity and for dynamic recruitment/integration, predictors were group, session, and interactions. Likelihood ratio χ² tests assessed effects; planned contrasts and paired/two-sample t tests reported where relevant. Multiple comparison corrections applied as stated (Bonferroni for selected tests; FDR for system-wide panels).

Behavioral performance: - Significant session × condition × group interaction on d′: χ²(3)=9.39, p=0.02. Largest gain in experimental group for 2-back from Naive to Late: +43.2% d′, paired t(20)=9.17, p<0.0001 (Bonferroni-corrected). Control group 2-back increased by 24.3%, paired t(20)=-6.45, p<0.0001 (Bonferroni-corrected). Between-group difference in 2-back improvement: t(20)=-4.12, p=0.004 (Bonferroni-corrected). In 1-back, experimental +12.2%, t(20)=-3.18, p=0.02 (Bonferroni-corrected); control not significant, t(22)=-1.91, p=0.28; between-group difference ns, t(39.64)=-0.52, p=0.47. After training, experimental showed no performance difference between 1-back and 2-back, t(20)=0.02, p=0.98; control retained a difference, t(20)=4.91, p=0.0016 (Bonferroni-corrected). Baseline static modularity (Naive): - Main effect of condition: χ²(2)=84.13, p<0.00001. Rest > task: β=0.20, t(88)=-11.37, p<0.00001. 2-back < 1-back: β=-0.08, t(296)=-2.60, p=0.01. Training effects on static modularity: - Main effects: session χ²(2)=19.40, p=0.0002; group χ²(1)=6.62, p=0.01. No session×group interaction χ²(1)=1.44, p=0.69; no session×condition χ²(1)=1.50, p=0.68. Increases from Naive to Middle: β=0.15, t(114)=2.61, p=0.01; Naive to Late: β=0.24, t(114)=4.05, p=0.0001. Experimental group higher mean modularity (M=3.09) than control (M=2.87). Paired tests Naive→Late: experimental increased in 1-back t(20)=-3.66, p=0.006 (Bonf.), and 2-back t(20)=-3.33, p=0.013 (Bonf.); control changes ns. - Change in modularity did not correlate with behavioral change in experimental 2-back (r=0.08, p=0.71). Dynamic network reconfiguration: - Frontoparietal (FP) recruitment: session × group interaction χ²(3)=9.03, p=0.028; experimental increased from Early→Late, β=0.07, t(120)=-2.892, p=0.027 (Bonf.); control ns. - Default-mode (DM) recruitment: main effects of session χ²(3)=24.17, p<0.0001 and group χ²(1)=3.96, p=0.046; interaction ns χ²(3)=2.66, p=0.48. Largest increase Naive→Late β=0.09, t(123)=5.00, p<0.0001. Experimental > control overall, t(165.6)=-3.03, p=0.003. - FP–DM integration: session × group interaction χ²(3)=14.25, p=0.0025. Decreased Naive→Late only in experimental: β=0.07, t(120)=4.37, p=0.0002 (Bonf.). Between-group differences in intermediate intervals: Naive→Early t(120)=2.16, p=0.03; Early→Middle t(120)=-2.70, p=0.02 (Bonf.). Experimental showed an inverted U-shaped trajectory; control showed the opposite pattern. Other systems (main session effects, FDR-corrected): - Increased recruitment: salience and auditory systems. - Increased integration among task-positive systems: FP–salience, dorsal attention–salience, dorsal attention–cingulo-opercular. - Decreased integration between DM and task-positive systems: DM–salience, DM–cingulo-opercular; also decreases for memory–somatomotor and DM–auditory. Behavior–network relationships (Naive→Late): - Δd′ (2-back minus 1-back) correlated with increased recruitment: DM r=0.33, p=0.03; salience r=0.34, p=0.03 (uncorrected). Correlated with increased FP–salience integration r=0.35, p=0.02; and with decreased DM–FP r=-0.31, p=0.04 and DM–salience r=-0.41, p=0.006 (uncorrected). Similar patterns observed with pRT (opposite sign due to metric). Subcortical dynamics and early learning: - Experimental group showed nonlinear (inverted U) changes in subcortical integration with dorsal attention, ventral attention, cingulo-opercular, and auditory systems (increase Naive→Early then decrease). Opposite pattern for subcortical–DM integration (decrease then increase). Group differences also noted for cingulo-opercular couplings with memory and uncertain systems, and dorsal attention–somatomotor. - Early-phase (Naive→Early, experimental) behavioral gains positively correlated with increased integration among multiple systems (e.g., dorsal attention–somatomotor, dorsal attention–subcortical, FP–somatomotor, dorsal attention–cingulo-opercular, salience, and DM); negatively with subcortical–cingulo-opercular integration. Overall, early integration supports performance before automation; later training associated with greater segregation of DM and reduced DM–FP coupling.

The study asked whether mastering a demanding working memory task reduces the need for global integration, enabling performance within more segregated modules. Consistent with prior work, modularity decreased with higher task demands relative to rest, confirming that challenging conditions engage more integrated configurations. Critically, across six weeks of training, whole-brain modularity increased during task performance, indicating a shift toward segregation even while performing the demanding 2-back condition. Dynamic analyses revealed that training increased recruitment (autonomy) of the default-mode system in both groups and of the frontoparietal system specifically in the experimental group. Integration between default-mode and frontoparietal systems decreased with training only in the experimental group, following a nonlinear trajectory, aligning with the transition from controlled to automated processing. Behaviorally, greater improvements were associated with higher default-mode (and salience) recruitment and reduced default-mode coupling with task-positive systems, suggesting that improved performance is supported by enhanced within-system stability and decreased cross-system coupling, especially between DM and FP. Early in training, increases in integration across multiple task-positive and subcortical couplings related to performance gains, implying that integration supports initial learning before automation. Later, segregation dominates, reflecting more efficient, less costly processing within specialized systems. These findings refine theories of cognitive control and global workspace by demonstrating that complex, once-mastered tasks can be executed within a more segregated architecture while retaining demand-dependent modulation.

Intensive working memory training leads to more segregated functional network organization during task performance. Static analyses showed increased whole-brain modularity across sessions, and dynamic analyses revealed increased recruitment of default-mode (both groups) and frontoparietal (experimental) systems, along with decreased default-mode–frontoparietal integration in the trained group. Behavioral improvements related to greater default-mode and salience recruitment and reduced coupling between default-mode and task-positive systems, while early-stage gains were supported by increased integration among task-positive and subcortical systems. Together, these results suggest a training-driven transition from integration (supporting initial learning) to segregation (supporting automated performance) in large-scale brain networks. Future research should examine finer-grained temporal dynamics (e.g., daily changes), causal interactions via effective connectivity, dose–response effects of training intensity, and larger cohorts to resolve group-by-session effects and individual differences.

- Group-by-session interaction for static modularity was not significant, and control participants also showed modest behavioral gains and some network reconfiguration, indicating possible effects of repeated exposure and limited power to detect between-group differences. - Scanning occurred every two weeks; finer temporal sampling might capture rapid neuroplastic changes during early learning phases. - Correlations between modularity change and behavior were not significant, suggesting static measures may be insensitive to individual differences compared to dynamic metrics. - Analyses are based on functional connectivity; causal directionality between systems (e.g., subcortical influence on frontoparietal/default-mode dynamics) was not assessed. Effective connectivity analyses would help infer mechanisms. - Positive-only FC edges were primarily analyzed; although robustness checks included signed networks, edge-thresholding choices can influence graph metrics. - Sample comprised healthy young adults; generalizability to other populations (older adults, clinical groups) is untested.