Dynamic switching between brain networks predicts creative ability

Creative thinking is thought to emerge from the interaction of spontaneous associative processes and deliberate cognitive control, largely subserved by coupling between the Default Mode Network (DMN) and Executive Control Network (ECN). Prior neuroimaging work links creative ability to DMN and ECN engagement and to static functional connectivity within and between these networks. However, static approaches cannot capture the dynamic coordination required to move between generative and evaluative phases of creative thought. The study asks whether individual differences in creativity are predicted by dynamic switching between DMN–ECN segregation and integration and whether optimal creative performance reflects a balanced (inverted-U) relationship in these dynamics. The authors test these hypotheses across large, multi-center resting-state datasets and validate them during an in-scanner creative task, aiming to establish a robust, network-dynamic neural marker specific to creative ability.

Extensive literature implicates the DMN in internally directed cognition (e.g., memory activation, mind wandering, associative processes) that supports idea generation, and the ECN in controlled retrieval, flexibility, and evaluative processes. Static connectivity studies have shown that interactions within and between DMN and ECN predict creative performance at rest and during tasks, and that DMN–ECN coupling reflects coordination between generative and evaluative processes. Dynamic connectivity studies report temporal variability within DMN/ECN linked to creativity, and broader work on network segregation/integration suggests that balanced dynamics benefit cognition (e.g., memory). Yet, direct evidence connecting dynamic DMN–ECN state switching (segregated vs integrated) to individual creative ability has been limited. Dual-process and dynamic frameworks of spontaneous thought predict that shifting between DMN–ECN segregation and integration—reflecting generation and evaluation—should be beneficial, potentially with an optimal balance (neither excessive segregation nor integration).

Design and samples: Meta-analytic network neuroscience approach using raw data harmonization across 10 independent resting-state fMRI datasets from Austria, Canada, China, Japan, and the USA (final N = 2433; ages 16–58). Divergent thinking ability was assessed outside the scanner (primarily Alternate Uses Task; multiple raters per dataset when available), with creativity scored by originality alone or originality+fluency depending on dataset. Six datasets (N = 908) also included intelligence. Imaging preprocessing: Structural and functional MRI were preprocessed with fMRIPrep 1.4.1rc1 (Nipype 1.2.0). Participants with excessive motion or data quality issues were excluded. Motion denoising included ICA-AROMA; FD and DVARS thresholds were applied. Dynamic functional analysis: Time series were extracted from DMN and ECN regions (Schaefer 300-parcel atlas; DMN/ECN subsets). Time-resolved functional connectivity was estimated using Multiplication of Temporal Derivatives (MTD) within sliding windows of 14 TRs. For each window, community structure (Louvain algorithm; 500 runs with consensus) yielded within-module strength z (W_t) and participation coefficient (B_t). A cartographic profile (2D histogram of W_t and B_t) per window was clustered via k-means (k=2; 100 restarts) into two states: integrated (higher B_t, lower W_t; Tin) and segregated (higher W_t, lower B_t; Tse). Metrics: (1) Switching frequency = number of Tin→Tse plus Tse→Tin transitions; (2) Balance = (Tin dwell − Tse dwell)/Total dwell (zero indicates perfect balance). Within each dataset, creativity–switching correlations controlled for gender, age, mean FD, and global signal; robust correlations were also computed. For balance, linear vs quadratic models (including squared term) were compared via ANOVA and AIC. Meta-analysis and moderators: Correlations (r) were transformed and combined using random-effects models (metafor, REML). Moderators examined included location (center), scanner type, creativity scoring method, sample size, age, scanning time, and gender ratio. Robust correlation-based meta-analysis was also conducted. Intelligence served as a control analysis across six datasets. Mega-analysis: All individual-level data were pooled after within-dataset Z-scoring of creativity, switching frequency, and balance. Linear mixed-effects models (nlme) tested relationships between creativity and switching and between creativity and balance (linear vs quadratic), with dataset-related effects modeled appropriately; model comparison via AIC and likelihood ratio tests. External task-fMRI validation: Independent sample (N=31) completed in-scanner Alternate Uses Task (AUT; creative ideation) and Object Characteristics Task (OCT; control). Events were aligned to ideation onsets; DMN–ECN dynamic states were computed analogously (Schaefer DMN/ECN ROIs; sliding windows; k=2). For each participant and task/run, switching frequency and balance were computed. Paired t-tests compared AUT vs OCT switching. Mixed-effects models tested associations between task performance (originality) and switching/balance (linear and quadratic).

• Creativity relates to DMN–ECN switching frequency: Significant positive correlations in 7/10 datasets. Random-effects meta-analysis: g = 0.174, 95% CI [0.08, 0.27], Z = 3.69, p < 0.001; heterogeneity Q(9)=20.67, p=0.014, I²=56.45%.
• Moderator effects: Location, scanner type, and scoring method jointly moderated the effect (QB=27.05, p<0.001). Subgroups—Austria (k=3): g=0.311 [0.17,0.46], p<0.001; China (k=4): g=0.114 [0.01,0.22], p=0.04. Magnetom (k=4): g=0.284 [0.18,0.39], p<0.001; Trio (k=5): g=0.11 [0.01,0.20], p=0.037. Scoring originality only (k=7): g=0.18 [0.09,0.28], p<0.001; originality+fluency (k=3): g=0.14 [0.01,0.27], p=0.042. Participant-related moderators (sample size, age, scanning time, gender ratio) were non-significant.
• Robust correlation meta-analysis: g = 0.149 [0.07, 0.23], z = 3.47, p < 0.001; heterogeneity Q(9)=16.08, p=0.065, I²=44.03%.
• Specificity to creativity vs intelligence: Switching–intelligence (6 datasets, N=908) not significant (g=0.023 [−0.042, 0.089], Z=0.699, p=0.485; Q(5)=4.076, p=0.539). In the same six datasets, switching–creativity was significant (g=0.20 [0.05,0.36], z=2.54, p=0.01; Q(5)=15.05, p=0.01). Difference between models significant (z = −2.069, p = 0.038).
• Inverted-U (balance) relationship: Quadratic model fit better than linear in three datasets (UG_S1, TKU, SLIM_S2). Meta-analysis of quadratic term: g = −0.07, 95% CI [−0.14, −0.01], z = 2.26, p = 0.024; Q(9)=16.72, p=0.053, I²=48.81%. Interpreted as optimal (moderate) balance between segregation and integration benefits creativity.
• Mega-analysis: Switching frequency predicts creativity (β=0.15, t(2419)=4.6, p<0.001). Quadratic balance term predicts creativity (β=−0.06, t(2418)=−3.81, p<0.001); nonlinear model better (AIC=6925.45) than linear (AIC=6931.19; χ²=7.74, p=0.005). Replication with Shen atlas DMN/ECN ROIs: g=0.09 [0.01,0.16], Z=2.18, p=0.029.
• Task-fMRI validation (N=31): AUT showed higher DMN–ECN switching frequency than OCT (0.14±0.05 vs 0.12±0.07), t(86)=2.59, p=0.011, d=0.28. Switching frequency did not correlate with AUT originality (β=0.02, t=0.04, p=0.971) or OCT originality (β=−0.24, t=−0.82, p=0.418). Balance showed a significant quadratic relation with AUT originality (β=−0.41, t=−2.62, p=0.011; nonlinear model AIC=53.82 better than linear AIC=58.46; χ²=6.64, p=0.01), but not with OCT originality.

Across large, diverse resting-state datasets, individuals who more frequently switch between segregated and integrated DMN–ECN states generate more original ideas, and those with a balanced interaction profile (neither excessive segregation nor integration) show the best creative performance. The absence of a relationship with intelligence and lower switching during a control task (OCT) suggest specificity of these dynamics to creative cognition rather than general ability. These findings refine network-level accounts of creativity by identifying dynamic patterns—frequent switching that maintains an optimal balance between spontaneous (DMN) and controlled (ECN) processes—as a neural marker of creative ability. This aligns with dual-process and dynamic frameworks proposing that creative thought cycles between generation and evaluation, requiring flexible but balanced coupling of DMN and ECN. The validation in task fMRI strengthens interpretability by demonstrating that creative ideation engages greater DMN–ECN switching and benefits from balanced coupling in situ.

This study establishes a robust association between creative ability and dynamic DMN–ECN interactions: higher switching frequency and an optimal balance between segregated and integrated states predict better divergent thinking. Results were consistent across multi-center resting-state datasets, replicated with alternative atlas parcellation, and validated during a creative task. These findings highlight a general neural marker of creative ability and support a dynamic framework of creative cognition. Future research should: (1) delineate subnetwork- and region-specific contributions within DMN and ECN across generative versus evaluative phases; (2) test causal manipulations (e.g., neuromodulation, training) to alter network dynamics and creativity; (3) expand to other creative domains and real-world creativity; (4) standardize acquisition and scoring to reduce heterogeneity; and (5) examine broader network interactions (e.g., salience, sensorimotor, visual) in multi-network dynamic models.

• Effect size is small, indicating DMN–ECN dynamics are one contributor among many (e.g., personality, attention, curiosity, self-beliefs). The directionality (ability shaping dynamics vs dynamics enabling ability) remains to be clarified.
• Meta-analytic pooling across heterogeneous centers introduces variability in scanners, protocols, and creativity scoring. Although moderators were modeled, more homogeneous large datasets would improve precision.
• The field debates required sample sizes for robust brain–behavior associations; while the present N is large and methods standard, even larger and more uniform datasets could improve replicability and reduce effect inflation.
• Parcellation choice (atlas resolution and DMN/ECN definitions) can affect associations; results were replicated with an alternative atlas but further validation across parcellations is needed.
• Analyses focused on DMN and ECN as whole networks, not finer-grained subnetworks or other networks that may contribute to creative cognition.