Transparency and Openness
This study adheres to the Transparency and Openness Promotion (TOP) Guidelines57. All key aspects of the research, including sample size determination, task manipulations, data exclusions, and measures, are reported. The data, analysis code, and research materials are available from the corresponding author upon request. The data were analyzed using R version 4.4.0 (R Core Team, 2023) and MATLAB 2019a (The MathWorks Inc., 2019). The study design and analysis were not pre-registered. The language in the introduction and discussion sections was refined with the assistance of generative artificial intelligence (e.g., ChatGPT) to ensure clarity and coherence.
Participants
A sample of 86 university students was recruited as subjects for this study. All participants were drawn from a normal university in China to ensure consistency in academic environment. Recruiting participants was particularly challenging due to the limited availability of design major students and COVID-19 pandemic-related restrictions. Therefore, we adhered to the widely accepted rule of thumb for experimental studies, which recommends a minimum of 30 participants per group. Consequently, the design group consisted of 44 participants, and the non-design group consisted of 42 participants. However, eight participants were excluded from the analysis: one non-design participant withdrew midway, two non-design participants failed to follow task instructions, one design participant had excessive head movement, another design participant’s work was deemed inconsistent with design expertise by five professional raters, and three participants were left-handed. In the end, 78 valid participants were obtained, comprising 39 design major students (32 females, aged 22.23 ± 1.74 years, grade 4.85 ± 1.46) and 39 non-design major students (30 females, aged 22.74 ± 1.62 years, grade 5.23 ± 1.25). All design-major participants were sophomores or above and had received an average of 4.65 ± 4.30 years of drawing training prior to university, as well as 4.85 ± 1.46 years of professional design training after entering university. Within the design major group, participants were distributed across product design (25.6%), visual communication design (23.1%), environmental design (23.1%), public art (15.4%), and digital media art (12.8%), reflecting a representative range of disciplines within design education. The non-design major group consisted of students in their second year or above, none of whom had received professional or amateur training in painting, photography, or other related arts, nor reported hobbies in these areas (e.g., regularly visiting painting exhibitions). Among the 39 non-design participants, 13 (33.3%) were from STEM-related disciplines (e.g., software engineering, computer science, neurobiology, mathematics education), and 26 (66.7%) were from non-STEM fields (e.g., psychology, education, accounting, sociology, business), ensuring a broad representation of academic backgrounds. Participants should have normal or corrected-to-normal vision and should not have any serious physical or mental illnesses. Participants were paid RMB 100 for their participation. All procedures were approved, and all participants provided informed consent prior to participating in the study, as approved by the East China Normal University Committee on Human Research Protection (UCHRP) (Code: HR 132-2020).
Experimental tasks and procedure
On arrival at the laboratory, participants signed an informed consent form and completed a demographic questionnaire. The experimenter then used a PowerPoint presentation to explain the tasks and their specific requirements for each stage, ensuring that participants fully understood before proceeding. Participants were then fitted with an fNIRS device to record brain activity at rest and while completing different creativity tasks, with the order of tasks randomized (see Figs. 1a, 2a). A camera recorded the creative process, and task progress was controlled using E-Prime 2.0. After the fNIRS session, participants provided recorded explanations of their creations. Following the completion of the four fNIRS tasks and their recorded explanations, participants performed a series of cognitive tests, including a 12-item short form of the Standard Raven’s Advanced Progressive Matrices (1993 version) to assess intelligence, and the Visuospatial Working Memory Span Task58 requiring sequential recall of locations on a 5 × 5 grid. The four creativity tasks were as follows:
Product Design Task (see Fig. 1b). This task was adapted from the paradigm developed by Kowatari et al. (2009) 18 and further refined for the present study. Participants were asked to design an innovative product—a “suitcase”—by following a set sequence of steps. The product could be improved in both function and appearance, potentially going beyond existing technological constraints. The more creative the product, the better. The task consisted of four phases. The first two phases comprised the idea thinking phase I and idea thinking phase II. In idea thinking phase I, participants analyzed the pain points and shortcomings of current suitcase designs, with two minutes allocated for silent reflection and two minutes for written articulation. In idea thinking phase II, they generated creative solutions to address the previously identified issues, again with two minutes for thinking and two minutes for writing. Next was the idea generation phase, in which participants conceptualized and sketched their design solutions over seven minutes, thinking and writing simultaneously. The final idea revision phase involved refining and optimizing the initial sketches, with three minutes allotted for revision. The duration of each phase was determined based on a pilot study with 10 participants, in which all were able to complete the tasks within this time. Our aim was to capture the initial stage of creative ideation rather than fully elaborated designs; therefore, a brief time window was considered sufficient and helped ensure focused, task-related responses.
Book Cover Design Task (see Fig. 1c). This task was adapted from Ellamil et al. (2012)53 and further refined for the present study. Participants were tasked with designing a book cover for Gu Long’s wuxia novel Happy Heroes by following a structured sequence of steps. After reading the book’s synopsis, participants were instructed to create a cover. The more creative the cover design, the better. The first two phases were idea thinking phase I and idea thinking phase II. In idea thinking phase I, participants identified the thematic concept of the cover, with two minutes for silent reflection and two minutes for writing. In idea thinking phase II, they explored visual strategies to communicate this theme, again with two minutes for thinking and two minutes for writing. This was followed by the idea generation phase, during which participants had seven minutes to conceptualize and sketch their design ideas, combining both visual and written elements. Finally, in the idea revision phase, participants refined and optimized their initial sketches, with three minutes allocated for adjustments. The time for each stage was also set based on 10 pilot participants.
Both the product design and book cover design tasks have been previously used in creativity research9,18,53 and are well-aligned with the training objectives of design-related visual arts education. Specifically, the product design task emphasizes both functional innovation and aesthetic refinement, while the book cover task engages visual communication and aesthetic expression. Together, these tasks provide appropriate and context-relevant measures of creative performance for design-trained students.
The verbal and visual divergent thinking task (see Fig. 1d). The Alternative Uses Task (AUT)59 was used to assess verbal divergent thinking. Participants were required to generate as many unusual and original uses as possible for common objects. In this experiment, the AUT task used the item “tire”. The visual divergent thinking test was adapted from the Picture Completion task, a subset of the Torrance Test of Creative Thinking figural version (TTCT-figural)60,61. Participants used four geometric shapes (i.e., one triangle, two circles, and one rectangle) to create meaningful pictures and then name their pictures. Each geometry figure could only be used once. The size of the figure could be altered, but not the shape. Each task had a total duration of 3 min and followed the “idea-button” paradigm62. Specifically, participants pressed the “Enter” key each time they thought of an idea and then recorded their answer by writing or drawing it on the answer sheet.
Assessment of performance on creative tasks
The performance of participants on the AUT and TTCT-figural tasks was assessed in terms of fluency and originality59. Fluency scores were calculated based on the total number of ideas generated. Originality scores were determined by evaluating the uniqueness of the ideas. Each idea was rated by five trained raters on a 5-point Likert scale, ranging from 1 (not original at all) to 5 (highly original). The inter-rater agreement was found to be satisfactory, with internal consistency coefficient (ICC, calculated as Cronbach’s α) values of 0.77 for the AUT task and 0.88 for the TTCT-figural task. The final originality score for each participant was calculated by averaging the ratings from the five raters. For the product design task, five independent raters, each with at least five years of experience in design arts, rated the products on a 5-point Likert scale across 10 dimensions9: functional novelty, functional usefulness, functional fluency, appearance novelty, aesthetics, concept novelty, elaboration, imagination, likeability, and overall rating. The ICC values for these dimensions were 0.77, 0.68, 0.88, 0.76, 0.72, 0.79, 0.79, 0.78, 0.78, and 0.78, respectively, indicating satisfactory inter-rater agreement across all dimensions. Similarly, for the book cover design task, five independent raters evaluated participants’ designs on a 5-point Likert scale across nine dimensions9: content novelty, content appropriateness, content fluency, drawing novelty, aesthetics, elaboration, imagination, likeability, and overall evaluation. The ICC values for these dimensions were also satisfactory, with scores of 0.66, 0.72, 0.78, 0.89, 0.87, 0.89, 0.84, 0.86, and 0.86, respectively. The inter-rater reliability for all dimensions in both tasks met the required standards, confirming the robustness of the evaluation process.
Propensity Score Matching
To address potential confounding covariates and ensure comparability between the art and control groups, propensity score matching (PSM) was conducted using the MatchIt package in R34. Missing data in the continuous covariates (e.g., IQ and visual working memory capacity) were imputed using their means. Propensity scores were estimated using logistic regression to predict group membership based on the covariates including age, grade, gender, IQ, and visual working memory capacity (VisWMC), following the foundational framework introduced by Rosenbaum and Rubin (1983)63. Initially, optimal pair matching was applied, in which the sum of the absolute pair distances between participants in the designer group and the control group (e.g., the non-designer group) was minimized64. The quality of the matching was evaluated using standardized mean differences (SMD) and variance ratios (VR) of each covariate. In psychology, SMD values65 smaller than 0.2 and VR values66 smaller than 2 indicated an acceptable balance. However, despite improvements, optimal pair matching did not achieve sufficient balance across all covariates. Therefore, optimal full matching, a matching algorithm that has been shown to outperform other matching algorithms in terms of balance65,67, was applied instead. The quality of matching was assessed again using SMD, VR, and eCDF statistics for all covariates. The matched dataset was then used for subsequent analyses. However, full matching resulted in a smaller effective sample size for the control group, which may reduce statistical power and affect the stability of the results. This methodological limitation should be taken into account when interpreting the findings.
fNIRS data acquisition
In this study, functional near-infrared spectroscopy (fNIRS) data were collected using the ETG-7100 system (Hitachi Medical Corporation, Japan). The device specifically assessed the absorption of near-infrared light at wavelengths of 695 and 830 nm, with a sampling rate of 10 Hz, enabling continuous monitoring and recording of concentration changes in deoxyhemoglobin (HbR) and oxyhemoglobin (HbO) in the brain. Three optode probe sets were used during the experiment: a 3 × 5 set, a 3 × 10 set, and a 4 × 4 set. The 3 × 5 probe set consists of 7 detectors and 8 emitters, with a minimum optode separation of 3 cm, supporting 22 measurement channels. The 3 × 10 probe set consists of 15 detectors and 15 emitters, with a minimum optode separation of 3 cm, supporting 47 measurement channels. The 4 × 4 probe set consists of 8 detectors and 8 emitters, with a minimum optode separation of 3 cm, supporting 24 measurement channels. The three probe sets were strategically placed to monitor bilateral prefrontal, parietal, temporal, and occipital regions, effectively achieving near-whole-brain coverage (see Fig. 2b). The registration of the probe sets was based on the 10–20 system. For the 3 × 5 probe set, the bottom row of optodes was aligned with the horizontal reference line, with the central optode corresponding to the nasion. For the 3 × 10 probe set, the central row was aligned with the sagittal reference line, with the channel at this row corresponding to FCz. For the 4 × 4 probe set, the bottom dorsal row of optodes was aligned with the horizontal reference line, with the central channel corresponding to Pz, and the middle column aligned with the sagittal reference line. Channel locations were further determined using a 3D digitizer and converted into Montreal Neurological Institute (MNI) space coordinates. The MNI coordinates of each fNIRS channel were then assigned to the corresponding regions using the Automated Anatomical Labelling (AAL) atlas. Specifically, each channel was automatically assigned to the nearest AAL region based on its MNI coordinates, and channels within the same AAL region were grouped to form regions of interest (ROIs). The MNI co-ordinates of the optode channels of a normal participant are shown in Supplementary Table 1.
fNIRS data analysis
Several pre-processing steps have been taken to reduce noise68,69. First, the principal component spatial filtering algorithm was applied to remove global components from the raw data. Next, we used correlation-based signal enhancement to remove motion artefacts. After correction, oxy-hemoglobin (HbO) and deoxy-hemoglobin (HbR) were found to be negatively correlated (i.e., corrected HbR values were equal to the products of corrected HbO and a negative coefficient70). Therefore, data analysis focused primarily on HbO rather than HbR. Additionally, poor channels were identified by visual inspection using NIRS time course plots, identifying channels with significantly higher variance compared to others in the same participant. The noise channels were removed and replaced with the average data of its four neighboring channels (e.g., if channel 34 is identified as a bad channel, it will be replaced with the average date of channels 30, 31, 37, and 38).
The fNIRS data analysis mainly includes three components: estimation of neural activation, estimation of neural coupling, and dynamic brain network analysis. Generally, the NIRS Statistical Parametric Mapping (NIRS_SPM) software package37,71 was used to analyze brain activation during different phases of the four creativity tasks and the resting baseline. The haemodynamic response function (HRF) low-pass filtering and wavelet minimum description length detrending algorithms were used, as required by “NIRS_SPM”. The general linear model (GLM) was used for the estimation of neural activation. Neural activation during different phases of the four creativity tasks and the resting baseline (each phase divided into an early half and a late half) was estimated through the following steps (see Fig. 2c): First, reference waves were established for each phase of the experimental tasks and the resting baseline (with each phase divided into an early half and a late half) to reflect the theoretical changes in HbO signals triggered by the experimental stimulus. Next, regression analyses were performed using the actual brain activation of each channel as independent variables to predict the changes in the reference wave. Each channel yielded a regression coefficient beta (β). The beta increment, obtained by subtracting the beta value of the resting baseline from the beta value of each phase of the experimental tasks, was used as an indicator of brain activation levels. Beta increments were then standardized across all participants using a Z-score transformation for each channel. Finally, the beta increment from each of the 93 channels was used as the dependent variable, with participant group (design and non-design group) as the independent variable. Independent samples t-tests with the weight calculated by the propensity score were carried out to examine the difference in neural activation between the groups. The false discovery rate (FDR) correction method was used to correct the results for multiple comparisons.
Neural coupling (NC) indicates the functional connectivity between different cerebral regions during the different stage of different tasks. Neural coupling strength was quantified using correlation coefficients (see Fig. 2c). Pearson correlations were used to assess the NC between the HbO signals from each pair of channels36. First, 8649 channel (CH) combinations (93 × 93 CHs) were identified. After exclusion of the redundant CH combinations [equal CH combinations (e.g., CH1-CH2 and CH2-CH1) or CH combinations of a single CH (e.g., CH1-CH1)], 4278 valid CH combinations remained. Fisher’s z-transformation was then used to convert the NC values (i.e., correlation coefficients). Similar to the analysis of the beta increment, a weighted independent samples t-test was performed for the NC group comparison, followed by correction for multiple comparisons using the FDR method.
A dynamic brain network analysis based on sliding windows and k-means clustering was conducted to characterize NC states during each creative task37,72,73 (Fig. 2c). Sliding windows were used to segment the NC data, with different time frames analyzed for the four creativity tasks. Specifically, for the product design and book cover design tasks, data from the first two thinking stages (2 min and 3 min, respectively) were concatenated before applying the sliding window approach. For the AUT and TTCT-figural tasks, the entire 3 min duration of each task was used for the analysis. The window size was set to 2 s, with a 0.5-second step increment, spanning the entire task duration. Within each time window, NC values were averaged, resulting in a series of NC matrices. These matrices were then averaged across participants, and a k-means clustering algorithm was applied in MATLAB to assess the similarity between windowed NC matrices and identify representative NC states. The similarity was quantified using Manhattan distance74. To determine the optimal number of states, various validity indices (e.g., the ratio of within-cluster distance to between-cluster distance) were calculated for a range of potential state numbers. These indices were plotted against the number of clusters, and the elbow criterion was applied to select the optimal number of NC states. According to this criterion, the selected state number corresponds to the “elbow” of the curve, balancing clustering precision with complexity75,76 (Supplementary Fig. 1). In this study, k = 3 was determined to be the optimal number of clusters for each creative task. To minimize the impact of random effects, 1000 iterations were conducted, and cluster centroids representing the three NC states (State 1, State 2, and State 3) across all participants were identified. These centroids, derived from the averaged NC matrices, were then used as initial centroids for individual-level clustering, enabling the extraction of dynamic NC states for each participant.
Each state was further analyzed using graph-based network metrics computed with the GRETNA toolbox in MATLAB77,78. In this analysis, channels were treated as nodes of the brain network, and the functional connections between them were defined as edges, forming weighted undirected networks. Then a dynamic sparsity threshold was employed to ensure that only the most significant connections in the NC matrices were retained, minimizing the influence of spurious or weak correlations. Unlike fixed thresholds, the minimum sparsity threshold Smin was dynamically calculated for each subject, ensuring that the threshold was tailored to the specific properties of the data. The sparsity range was set from Smin + 0.5 with a step size of 0.02, allowing for a comprehensive exploration of the network across varying sparsity levels. The brain network parameters analyzed included clustering coefficient (CC), average shortest path length (ASPL), global efficiency (globE), and local efficiency (locE). The CC reflects the degree of local interconnectivity or “cliquishness” in the network, with higher CC values suggesting more efficient information transfer among neighboring brain regions79. ASPL measures the harmonic mean of the shortest paths between all node pairs, with shorter ASPL values indicating more direct pathways for information transfer across distant brain regions80. GlobE assesses the overall efficiency and capacity of information transmission across the entire network, reflecting global integration, while locE measures the efficiency of information transfer within localized clusters or sub-networks of the brain. Higher values of globE and locE indicate that the brain network can process information more rapidly and in parallel, leading to more efficient cognitive processing77. Repeated measures ANOVAs were followed by post hoc pairwise comparisons using Bonferroni correction to explore differences in network parameters between the three states (State 1, State 2, and State 3) for each creative task.
Following the identification of NC states for each participant, various metrics were used to characterize these states36,76. These metrics included the ratio of each state, which indicates the proportion of total windows attributed to each state; the number of transitions between states, which reflects the frequency of state changes; and the dwell time of each state, representing the average duration spent in each state, calculated from the starting point of each occurrence. Several weighted independent samples t-tests were performed to examine the difference in the abovementioned metrics between design and non-design groups. All p-values were corrected using the FDR method.
Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
