Creative experiences and brain clocks

Participants and demographics

This report included a total of 1,467 healthy participants. We used data from 1240 healthy controls (HCs) from the EuroLaD EEG Consortium¹², to train the SVM models, including participants from global settings. Specifically, 724 HCs were from the global north (Turkey, Greece, Italy, the United Kingdom, and Ireland) (age range 17–91 years, mean age 45.7 ± 22.6 years, 48.8% female/male distribution), while 516 were from the global south (Cuba, Colombia, Brazil, Argentina, and Chile) (age range 18–89 years, mean age 46.4 ± 18.5 years, 50% female/male distribution) (Fig. 1a-c, full demographics in Table 1)^55,56. Inclusion criteria required normal cognitive function and no history of disease. Study protocols were approved by each contributing institution’s Institutional Review Board (IRB), and all participants provided informed consent following the Declaration of Helsinki. The complete demographics are presented in Table 1.

We used independent datasets already published, consisting of 232 participants from studies about creativity expertise (study 1) and pre/post-learning (study 2, Fig. 1c). For study 1, we compared experts and control non-expert participants, considering domains related to dance⁴⁷ (N = 46, Argentina), music^49,57 (N = 58, Canada), visual arts⁴⁸ (N = 30, Germany), and gaming²⁶ (N = 62, Poland). Both expert and non-expert participants were age-, sex-, education-, and geography-matched (Table 1). Participants in the gaming group were also matched by working memory capacity (Table S1 in Supplementary Information). Participants in the dancing group were matched by abstraction capacity, working memory, attention, verbal inhibitory control, and verbal working memory (Table S5 in Supplementary Information). Study 2 consisted of measurements before and after a short-term learning session of gaming^29,58 (N = 27, Poland). For all participants considered in this work, sex was defined as the self-reported biological sex. See Table 1 for full demographics and sections Expertise criteria and Pre/post-learning design in Methods for details of expertise criteria and training design. All datasets used in this study were obtained from participants who provided informed consent following the Declaration of Helsinki. The study protocols, including image acquisitions and data collection procedures, were reviewed and approved by the Institutional Review Boards of each contributing institution.

Study 1. Expertise criteria

Tango dancers

A total of 46 right-handed Argentinean participants⁴⁷ completed a self-assessment questionnaire consisting of 20 items designed to measure their level of expertise⁴⁷. Expert and beginner tango dancers were recruited from three tango schools in Buenos Aires — DNI, Flor de Milonga, and Divino Estudio del Abasto Tango school. Exclusion criteria included no past neurological or psychiatric history reported. Inclusion criteria included being right-handed, verified by the Edinburgh Inventory, with normal or corrected-to-normal vision. The self-assessment questionnaire covered various areas, including tango practice, general dance experience, and formal tango instruction. Participants were classified according to their expertise level (Table S2 in Supplementary Information, item 17). We then built two coarser groups, where participants were assigned to the expert tango dancers’ group (N = 23), or the non-experts’ one (N = 23), using the median months of formal tango instruction as criteria, i.e., >12 months for expert tango dancers. Demographic data are presented in Table 1. The detailed questionnaire items can be found in Table S2 in Supplementary Information. This dataset consisted of 4.4 min resting-state EEG recordings. Ethics for this study was approved by the Comité de Ética of Universidad de San Andrés / CONICET, Buenos Aires, Argentina.

Musicians

This group included a total of 62 right-handed participants⁴⁹. The expert musicians consisted of 31 participants with 5+ years of experience playing a musical instrument, including professional and amateur musicians (23 and 8 participants each, respectively). The group included various musical expertise across different instruments (e.g., string instruments, percussion) and singing. The non-expert group (non-musicians) consisted of healthy participants who did not play any musical instrument. The information was retrieved from the OMEGA questionnaire using self-declared musical expertise and years of experience. Musicians were selected based on self-reported experience of playing a musical instrument for five or more years. Four participants were excluded due to data quality concerns, leaving a final sample of 29 experts and 29 non-experts. Demographic data are presented in Table 1. This dataset consisted of 5-min resting-state MEG recordings from the OMEGA database^49,57. MEG recordings and questionnaires are available through the OMEGA repository⁵⁷. The years of education were calculated by translating the maximal educational instruction (e.g., a PhD degree) into cumulative years. Ethics for this study were approved by the Research Ethics Board of Montreal Neurological Institute & Hospital (McGill U.), Montréal, Canada.

Visual artists

A total of 34 right-handed participants were initially recruited for the study, with 17 visual artists and 17 non-artists. Recruitment occurred through social media, posters in locations related to the topic (e.g., universities, art schools, and art institutions), and word of mouth. After initial email contact, candidates were screened for inclusion and exclusion criteria and asked about their artistic background. Suitable participants were then invited to schedule an EEG session, during which a more detailed questionnaire on their art practice and interests was completed. Exclusion criteria included psychiatric, neurological, or cardiac conditions, and hairstyles or accessories (e.g., dreadlocks) that could impair EEG data acquisition. Following exclusions due to low data quality and non-compliance with inclusion criteria⁴⁸, the final sample consisted of 30 participants: 15 artists (experts) and 15 non-artists (non-experts). As part of the selection criteria, the artists were required to have completed at least three years of university-level academic art education, with specific training in drawing. In contrast, participants in the non-artist group had no formal drawing training and did not engage in drawing regularly. Participants were informed about the data collected and signed informed consent before the session. Compensation was provided, although the specific form or amount is not mentioned in the article. Demographic data are presented in Table 1. This dataset consisted of 2 min resting-state EEG recordings⁴⁸. Ethics for this study was approved by the Ethics Commission of Humboldt-Universität zu Berlin, Germany.

Gaming

In this study, 62 right-handed male subjects were included²⁶. All participants completed an online questionnaire on demographics, education, and video game experience. The online questionnaire was administered via the GEX platform (GEX Immergo, Funds Auxilium Sp. z o.o), which gathered demographic information, education status, and detailed data on video game habits. As part of the questionnaire, participants provided their Battle.net ID, allowing verification of their StarCraft II league ranking and recent gameplay activity. All participants were right-handed males, had no history of neurological illness or psychoactive substance use, and were matched on undergraduate-level education. Working memory capacity was assessed using a modified online version of the operation span (OSPAN) task, with participants required to maintain at least 85% accuracy on the math component to be included. Two participants were excluded due to poor MRI data quality. All participants gave written informed consent and received monetary compensation for their participation. The expert group (N = 31) met the following criteria: (a) experienced in real-time strategy video games and StarCraft II, (b) played real-time strategy video games at least 6 hours/week for the past 6 months, (c) spent over 60% of gameplay time on StarCraft II, and (d) actively played in the last two seasons, ranked in one of the StarCraft leagues (Gold, Platinum, Diamond, Master, Grandmaster). The non-expert group (N = 31) had: (a) less than 6 hours of real-time strategy play and (b) less than 8 hours per week of total video game play (any genre) in the past 6 months. Only males were recruited due to the lack of female participants with sufficient video game experience. Demographic data are presented in Table 1. Further details about video experience, intellectual level, and other demographic variables can be found in Table S1 in Supplementary Information. As our brain clock models were trained from EEG FC, and this data were not available for participants, we estimated EEG FC from DTI structural connectivity using whole-brain modeling^27,59. Previous simulated FC from empirical DTI²⁶ successfully revealed connectivity differences between players and non-players²⁷. We also reported the structural differences between groups in Supplementary Information Figure S6. Ethics for this study was approved by the Research Ethics Committee, SWPS University of Social Sciences & Humanities, Warsaw, Poland.

Study 2. Pre/post-learning design

A total of 24 right-handed participants were considered for the study in a short learning paradigm using video games^29,58. Participants were initially recruited online via a covert questionnaire. All participants reported normal or corrected vision, normal hearing, and were right-handed. Exclusion criteria included any history of neurological or psychiatric disorders, head injuries, surgeries, brain tumors, current medication use, or more than five hours of video gaming per week in the prior six months, especially RTS or FPS games. Only participants who completed both sessions and met training requirements were included. Each eligible participant received compensation of approximately 184 USD. Participants engaged in StarCraft II gaming sessions in a controlled laboratory environment at the NeuroCognitive Research Center, SWPS University in Warsaw. Before the first session, each participant completed an introductory training session with a StarCraft II coach to familiarize them with the game’s core concepts and basic mechanics. The training lasted 30 hours in total, spread over 3 to 4 weeks, with participants playing between 5 and 10 hours per week. Telemetric variables were extracted from StarCraft II replays using the Python libraries sc2reader (available at https://github.com/ggtracker/sc2reader) and PACanalyzer (available at https://github.com/Reithan/PACAnalyzer), which enable the retrieval of information from various StarCraft II resources. We used the Actions Per Minute (APM) as a metric of performance, as it is one of the strongest predictors of StarCraft II performance and skill development⁴⁰. APM reflects cognitive, motor, and decision-making speed, increasing as participants gain experience. EEG data were collected during two lab sessions lasting up to two hours each, which included instructions, electrode setup, and an attentional blink task. Only those who completed both sessions and met the training requirements were included in the final analysis. The attentional blink task involved rapid serial visual presentation (RSVP) of letters at the center of the screen. Participants were instructed to detect and report two target letters (T1 and T2) appearing in the stream, with T2 occurring shortly after T1, to measure the transient lapse in attention characteristic of the attentional blink effect. T1 corresponded to a green capital letter (vowel or consonant), and T2 to a black “X” presented at lag 1, 2, or 7 after T1. Participants answered two yes/no questions: whether a vowel (T1) and/or an “X” (T2) appeared. From the test, we reported the T1 and T2 reaction times and accuracy. EEG recordings were acquired before and after 30 hours of training. The total playing time was monitored, and participants were strictly prohibited from playing outside the lab. Demographic data are presented in Table 1. This dataset consisted of 23 min of EEG recordings during an attentional blink task⁶⁰. This task was included to assess generalized attention and temporal processing improvements post-training. Although not directly associated with BAGs, improvements in attentional blink performance may reflect learning-induced changes⁶⁰.

We included an active control group (N = 12) trained in Hearthstone. This group was part of the original study design⁵⁸, with identical recruitment, training time, and inclusion/exclusion criteria as the StarCraft II group. The Hearthstone game was selected due to its more rule-based and turn-based mechanics, with limited improvisation and creative play compared to StarCraft II’s real-time decision-making⁵⁸. Ethics for this study was approved by the Research Ethics Committee, SWPS University of Social Sciences & Humanities, Warsaw, Poland.

M/EEG acquisition, preprocessing, and connectivity

M/EEG data were processed offline using a custom-built automated pipeline (Fig. 1d). The pipeline integrates a mesh model tailored to various electrode arrays and performs source space estimation. Details on acquisition parameters (acquisition time, electrode numbers, sampling rate), are provided in Table S3 in Supplementary Information. The full set of pre/post-processing procedures is available in Supplementary Information sections 1.1-5, with a summary provided below.

The M/EEG signals were re-referenced to an average reference and were resampled to a uniform sampling rate of 512 Hz. EEG preprocessing included re-differentiation, removal of muscle and eye movement artifacts, identification and interpolation of bad channels, and normalization. Source reconstruction was conducted using standardized Low-Resolution Brain Electromagnetic Tomography (sLORETA). Brain regions were defined according to the Automated Anatomical Labeling (AAL) atlas³⁸, including only the 78 cortical regions (regions listed in Supplementary Table S4). All MEG data were obtained from a public access data repository (OMEGA⁵⁷). Preprocessing included low-pass filtering, artifact removal, and co-registration of MEG with anatomical images. MEG source estimation was performed using an atlas-based beamforming approach. A dipole-based forward model and beamformer approach were used to estimate time courses for 78 AAL³⁸ regions with adjustments for signal polarity. We filtered the M/EEG signals between 8 and 40 Hz using a 3rd-order Bessel filter and then computed Pearson’s correlation between pairs of brain regions, resulting in 78×78 functional connectivity matrices. MEG FCs for musicians and non-musicians were adjusted to the average EEG FC connectivity of a subset of participants from the global north (age range 26−30 years).

The 8–40 Hz band was selected to cover alpha to low-gamma rhythms, which are linked to creative and attentional processes⁴⁸ as well as brain age^12,61. Although many features derived from M/EEG correlate with brain age (e.g., power spectral density, entropy, kurtosis)³⁹, we focused on functional connectivity for specific reasons⁸. Functional connectivity is a robust marker for assessing brain aging, particularly when considering diverse datasets^8,12,61. It also allows for comparability with fMRI-based brain clock estimations⁸. By using a mesh model and a common source space, we ensured standardized brain mapping across participants⁸. This approach offers a balanced trade-off between spatial and temporal resolution, which is important in multi-site data. In any case, future studies should explore additional M/EEG-derived metrics³⁹ to further enrich brain age modeling.

Additionally, we assessed the Overall Data Quality (ODQ) of the EEG recordings using the method developed by Zhao et al.⁶², to discard possible effects of data quality on BAGs’ computation. Despite their differences, BAGs derived from different functional sources can be compared using harmonization steps. The use of a similar approach and a common brain parcellation³⁸ allowed us to compare EEG and MEG in the same source space.

Brain clock models and brain age gaps

To improve the model’s robustness and generalizability⁶³, we applied data augmentation over the EEG functional connectivity matrices of the training dataset (Fig. 1d). Using the augmented data, we trained SVMs to predict chronological age^8,9 (Fig. 1d). A 5-fold cross-validation scheme with up to 15 repetitions was used, and the model performance was assessed by Pearson’s correlation coefficient and the Mean Absolute Error (MAE) between predicted and real chronological ages in the test sets. Feature importance was defined as the absolute value of SVR weight coefficients, averaged across cross-validation folds and repetitions. Detailed methods are provided in Supplementary Information sections 1.6-7. The BAGs were calculated by subtracting the actual chronological age from the SVM-predicted brain age. In the out-of-sample validation, these values were then corrected by regressing the chronological age. The regression slopes and intercepts were estimated from the training data^8,9. The model performance was not assessed using age bias correction as this step can artificially increase the model performance. We finally normalized the BAGs, subtracting the average BAGs within each domain.

Expertise scores and BAGs standardization

To compare different expertise scores and their associations with BAGs, we converted BAGs and expertise scores into z-scores to group them on the same scale. This step allows the merging of all the expertise scores on a common scale, ensuring that measures, such as hours per week and years of experience, are expressed under the same scale for consistent comparisons.

Brain maps and Neurosynth associations

From the training data, we correlated nodal functional connectivity with chronological data, representing the brain areas more vulnerable to age, i.e., negatively correlated with age (Fig. 4). We subsequently captured the differences in nodal strength between (a) expert vs non-expert participants across all domains and (b) pre/post-learning connectivity. We reported these differences using Cohen’s D effect size. We then used Neurosynth²¹, an automated meta-analytical tool, to explore the cognitive processes linked to altered connectivity associated with creative experiences. We obtained association maps for 89 cognitive terms, which were then parcellated using the AAL atlas. We correlated the changes in nodal strength described above with the Neurosynth association maps, reporting the absolute strength of correlations.

To account for potential spatial autocorrelation in brain maps, we conducted Spin Test analyses⁶⁴ using the BrainSMASH Python library (https://brainsmash.readthedocs.io/)⁶⁵. We applied this method to all correlation analyses involving cortical surface data. The Spin Test generates spatially constrained null models by randomly rotating the spherical projection of cortical surface maps while preserving their spatial structure. We used 10,000 permutations to generate null distributions of correlation values. Empirical correlations were then compared to these null distributions to obtain p-values corrected for spatial autocorrelation. We additionally applied false discovery rate (FDR) correction across all comparisons.

Graph theoretical analyses

We quantified functional network properties associated with creative experiences using tools from graph theory⁴¹. Functional connectivity matrices were binarized after applying proportional thresholds ranging from 0.02 to 0.1, in steps of 0.01, keeping the highest functional connectivity values after thresholding. We reported the mean value across the whole range of thresholds to minimize the arbitrariness of choosing a single value⁴¹. From the binarized matrices, we computed the global and local efficiency.

Global efficiency, E, is a measure of network integration and can be related to generalized information processing, that is, coordinated activity throughout the brain. Global efficiency is based on paths and was defined as⁶⁶

where E_i is the nodal efficiency, n is the total number of nodes, and d_ij is the shortest path between nodes i and j. Global efficiency ranges between 0 and 1. Higher values indicate highly integrated networks, where nodes can easily transmit information, while values closer to 0 suggest poor integration and more difficult node-to-node communication.

Local efficiency, L, is a measure of network segregation, capturing the efficiency of information transfer within local neighborhoods of the network, which reflects specialized information processing within clusters of interconnected nodes. It is computed as follows⁶⁶

where k_i is the degree of node i, Γ_i is the set of neighbors of node i, and d_jh represents the shortest path length between nodes j and h within this neighborhood. Local efficiency ranges from 0 to 1. Higher values indicate efficient communication within local neighborhoods, while values closer to 0 suggest weaker local connections.

Graph theoretical analyses were conducted using the Brain Connectivity Toolbox for Python⁴¹.

Generative whole-brain model

We used a whole-brain model^27,59 to explore associations between BAGs and biophysical coupling, specifically the global coupling parameter, G, of the model. Global coupling represents the overall conductivity of fibers, reflecting the strength of interregional communication.

Our model integrates structural and functional connectivity with regional dynamics, using a network of 78 brain areas defined by the AAL parcellation. The local brain activity was simulated using the normal form of a supercritical Hopf bifurcation (Stuart-Landau oscillators), which can shift the system’s behavior from self-sustained oscillations (limit cycle) to a stable fixed point.

The model’s key parameters include the global coupling and the local bifurcation parameter, a_i, which determines whether a region i exhibits noise-induced oscillations (a_i < 0), self-sustained oscillations (a_i > 0), or critical behavior (a_i ≈ 0). We set a_i = 0.01 for all regions, following previous studies²⁷. Regions were also subjected to uncorrelated Gaussian noise with a standard deviation of β = 0.1. The complete set of equations consisted of

Here, x(t) represents the real component of the M/EEG-like signals, and y(t) represents the imaginary component. Regions with a_i > 0 exhibit self-sustained oscillations at a frequency f_i = w_i/2π, set to 10 Hz across all nodes. The brain regions are coupled via an empirical structural connectivity, M, where each entry M_ij indicates the strength of the connection between regions i and j. We used diffusion tensor imaging (DTI) structural connectivity data from the gaming expertise design²⁶.

Using the linear approximation of the Hopf model⁵⁹, static functional connectivity was predicted using analytical estimations without running extensive simulations. These estimations can be applied under weak noise conditions and small nonlinearities⁵⁹.

Model fitting

We generated EEG functional connectivity from DTI structural connectivity and the Hopf model⁶⁷. We swept parameter G between 0 and 3 in steps of 0.1. We compared the simulated and empirical functional connectivity matrices using the structural similarity index (SSIM = 1, perfect fit)^18,68. In the gaming expertise design, the model was fitted to a subsample of age-matched participants from the training data (24.5 ± 1.0 years). The simulated functional connectivity matrices and G values were used for subsequent analyses. We validated our results by comparing the BAGs between expert and non-expert video game players across different values of G, finding consistent results across the entire parameter range (Fig. S3 and S6 in Supplementary Information). For the remaining groups, we used the average structural connectivity matrix across participants of the gaming expertise design (average across experts and non-experts). We then fitted individual G values using participants’ empirical functional connectivity matrices. The results of the model fitting are presented in Figs. S4, 5 in Supplementary Information.

Statistical analyses and visualization

Pairwise t-tests for independent samples (expertise) and paired t-tests (pre/post-learning) with a p-value of 0.05 were used to determine significance, plus the 95% confidence intervals and degrees of freedom. Cohen’s D was calculated to provide effect sizes, with values interpreted as small (0.2 < |D| < 0.5), moderate (0.5 <| D| <0.8), large (0.8 < |D| <1.2), and huge (|D| > 1.2). Cohen’s f² was also used to measure effect size in correlations, with values interpreted as small (0.02 < f² < 0.15), moderate (0.15 < f² < 0.35), large (0.35 < f² < 0.5), and huge (f² > 0.5). For categorical variables, like sex, we proceeded with Chi-squared tests. Pearson’s correlation was used to assess statistical relationships, and all p-values from multiple correlations were corrected for false discovery rate (FDR) using the Benjamini-Hochberg method. For generating the brain plots, we used the BrainNet Viewer Toolbox⁶⁹. Word cloud plots were generated with the wordcloud 1.9.3 Python package (https://github.com/amueller/word_cloud). We used one-sided Mann-Whitney U tests in relatively small sample sizes, which limited the suitability of parametric alternatives.

Sensitivity analyses

Sensitivity analysis 1: To control for potential confounding variables, we conducted an ANCOVA with BAGs as the dependent variable and the expertise (group), age, sex, and education as covariates. This analysis was performed separately for each creative domain (tango dancers, musicians, visual artists, and gaming). In addition, normality tests and matching using parametric and non-parametric tests are reported in Table S7 in Supplementary Information.

Sensitivity analysis 2: We analyzed the effects of age bias correction, comparing the experts and non-experts with and without the correction.

Sensitivity analysis 3: We repeated the spatial correlation analyses between age vulnerability maps and connectivity-based effect sizes separately for each domain of expertise. For the tango dancers, musicians, and visual artists groups, we used the EEG-derived age vulnerability maps; for the gaming group, the age vulnerability map was constructed using MEG data from musicians and non-musicians.

Sensitivity analysis 4: We investigated whether the BAGs are associated with domain-specific expertise levels. Participants were classified as BAG-younger or BAG-older based on the overall BAG distribution, merging experts and non-experts. Specifically, individuals with BAG values below the 35th percentile were labeled as BAG-younger, while those above the 65th percentile were labeled as BAG-older. We then compared expertise scores between these two groups using one-sided Mann-Whitney U tests and FDR correction.

Sensitivity analysis 5: To assess the impact of channel density on model performance, we computed the slope of predicted versus chronological age using linear regression, separately for 64- and 128-channel EEG datasets, in both the training and creativity datasets.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.