Decoding the cascading architecture of creative ideation: an ERP study | BMC Psychology

Participants

The present study recruited 150 healthy participants (age: 20.28 ± 2.26 years; 42 males) from a university in northern China. All participants were right-handed, had normal or corrected-to-normal vision, and reported no history of neurological or psychiatric disorders, major medical disorders (e.g., cardiac illness, stroke, or cancer), or head injuries. For the 12 h prior to their lab appointments, participants were instructed to abstain from alcohol, caffeine, and other stimulants and to get sufficient rest. Written informed consent was obtained from all participants before the study, and they were compensated after completing all tasks. This work was approved by the institutional review board of the local ethics committee and conducted in accordance with the latest version of the Declaration of Helsinki.

Task and procedure

To better capture the spontaneous nature of CI and to allow for a more precise estimation of its cognitive stages, this study adopted a computerized version of the AUT, following procedures established in previous studies [45, 46]. The experimental procedure was programmed using E-Prime 2.0. Before the formal experiment, four practice trials were conducted to ensure that participants fully understood the experimental instructions and procedures. The materials included 30 different everyday objects, each consisting of two Chinese characters (e.g., “umbrella” corresponding to “yusan/雨伞”, see Table S1 for all 30 objects). For each object, participants were instructed to engage in two types of thinking based on the cue words “common” or “creative.” In the common condition, they were explicitly instructed to think of a highly common, ordinary, or typical use, while in the creative condition, they were explicitly instructed to think of a highly unusual, original, or creative use. For the purpose of the present study, in the creative condition, this task instructed participants to select the most creative ideas to increase the evaluation process during the CI. Therefore, the applied approach strongly emphasized the originality aspect of creativity, in line with prior work [21]. Each object was presented once under both cue conditions in a pseudo-randomized order, resulting in 60 trials.

The timing of the task is shown in Fig. 1. Initially, a fixation point (“+”) appeared at the center of the screen for 10 s. This was followed by a preparation period lasting 5 s, during which the cue word “common” or “creative” was shown, indicating the types of thinking to be performed during the subsequent task. Then, the generation stage was displayed for a maximum of 30 s, or until the participants pressed the “Enter” button on the computer keyboard as soon as they had an idea that they believed to be the most suitable response, thus stopping their thinking. Participants then rated their ideas using a five-point Likert scale (1 = Not original at all, 5 = Very original) within 4 s. This setup was designed to ensure participants’ active engagement in the task and the generation of high-quality ideas [19, 21]. Next, participants verbally reported the ideas they had generated (e.g., “An umbrella could be used as a fruit basket”), and a recording device recorded these reports. Afterward, the recordings were transcribed into text. Then, participants were allowed to rest at their own pace. When ready to proceed, they pressed the “Enter” key to begin the next trial. The self-paced procedure was designed to better capture the spontaneous nature of CI and its distinct stages [21]. The presentation order of objects and conditions was pseudo-randomized. The experiment consisted of four blocks, each containing 15 trials. The total testing time was approximately 35 min.

Overview of the timing of AUT. The blue rectangle marks the time window (−500 ms to 2000 ms) for the EEG analysis in this study. The corresponding butterfly chart of the grand average ERP (including creative and common conditions) is presented in the bottom panel

According to a previous study [45], the participants’ originality was assessed by six external expert raters (3 males) independently. The raters were blind to the experimental conditions and the purpose of the study. Ratings were made on a Likert-type scale from 1 (not original) to 10 (very original). This procedure is a common approach in creativity research [21, 45]. Trials exceeding 30 s in reaction time or lacking responses were deemed invalid and removed from all analyses. For each condition (creative and common), the ratings were averaged across raters and items, yielding two originality scores per participant. Raters achieved good reliability for both conditions (Cronbach’s α: creative = 0.826; common = 0.917).

EEG data recording and preprocessing

In a room with dim lighting and reduced sound, participants sat on a comfortable chair with the stimuli presented on a monitor approximately 60 cm in front of them. Before EEG recording, the experimenter reminded the participants to minimize head, eye, and other movements from the fixation point (“+”) presentation to the moment of pressing the button to end thinking, to prevent artifacts. Outside of these stages, participants were free to move their eyes.

EEG data were recorded using a 64-channel Ag/AgCl scalp electrode placed according to the International 10–20 system (Brain Products GmbH, Germany, passband: 0.01–100 Hz; sampling rate: 1000 Hz). Two additional sensors were placed on the outer canthus and beneath the left eye to record horizontal (hEOG) and vertical (vEOG) electrooculography, respectively. The FCz was used as a recording reference, and AFz was used as a ground electrode. All impedances were kept below 5 kΩ.

The raw EEG data were preprocessed using the EEGLAB 2021.1 toolbox [47] on the MATLAB 2021b platform. Continuous EEG data were off-line high-pass filtered at 0.5 Hz and low-pass filtered at 30 Hz using the pop_eegfiltnew function (default options: zero-phase Hamming-windowed sinc FIR filter). The data were referenced offline to the average of the two mastoid electrodes (the original reference electrode was renamed “FCz”) and then downsampled to 500 Hz. Epochs were defined from − 500 ms to 2000 ms relative to stimulus onset. Trials contaminated by eye blinks, muscle, cardiac, and other transient artifacts were corrected using an independent component analysis (ICA) using the built-in function pop_runica of the EEGLAB toolbox. Artifactual components were carefully visually detected based on the time course, topography, and power spectral density of the components. Baseline correction was performed using the period from − 500 ms to 0 ms. Electrodes with excessively noisy signals were interpolated using spherical spline interpolation [48]. Epochs with amplitudes exceeding ± 100 µV were discarded. The average trial number was 27.64 (SD = 1.36) in the creativity condition and 29.47 (SD = 0.72) in the common condition. In line with previous research in the creativity field, and to obtain ERP data less influenced by reference electrodes, a common average reference was applied to the data before subsequent analyses according to relevant studies [49,50,51].

It should be noted that the analysis time window from − 500 to 2000 ms was chosen based on prior research [16, 45], as well as the temporal characteristics of the ERP butterfly plot observed in the present study (see Fig. 1). This time window sufficiently captures the key temporal dynamics of the ERP components relevant to CI. Beyond 2000 ms, the amplitude of evoked potentials returns to near-baseline level, suggesting that most phase-locked neural activity has subsided. Extending the epoch beyond 2000 ms would provide limited benefit while increasing the risk of introducing noise, potentially compromising data quality and interpretability. Furthermore, TGA was also conducted using a broader epoch window from − 500 to 4000 ms, which fully encompasses the average response time in the common condition (RT = 3.643 s). As shown in Fig. S4, decoding performance notably declines after 2000 ms, further reinforcing the decision to focus on the 2000 ms post-stimulus interval.

Univariate analysis

According to a previous study [16], an ERP univariate analysis was performed before MVPA. ERPs were calculated by averaging trial epochs per condition for each participant. Based on previous studies on conceptual expansion [32,33,34], as well as the grand-averaged waveforms obtained in the present study (see Fig. S2), this study investigated the N400 (390–440 ms) and LC (450–950 ms). The average amplitude (µV) was calculated for statistical analysis.

Multivariate pattern analysis

To track the temporal evolution of CI, this study applied MVPA to the AUT EEG data collected under both creative and common trials.

First, to investigate the decodable time range of CI, this study utilized the support vector machine (SVM) function with a linear kernel (regularization parameter C = 1) from the scikit-learn package ( to perform a time-by-time decoding analysis. The voltage values from all 60 electrodes were used as input features. To enhance decoding efficiency and performance, the EEG data were downsampled by averaging every 5 time points, compressing the original 1250 time points (−500 ms to 2000 ms) into 250 time points. The time-by-time decoding procedure was as follows: (1) For the issue of imbalance in the number of trials between conditions, an undersampling method was employed. The condition with fewer trials was identified, and trials from the other condition were randomly removed to ensure an equal number of trials in both conditions. Then, for each condition, the trials were randomized, and pseudo-trials were constructed by averaging every 5 trials to improve the signal-to-noise ratio and decoding accuracy. (2) Next, a 5-fold cross-validation (CV) was performed, using 4/5 of the samples for training and the remaining 1/5 for testing. (3) To obtain robust decoding accuracy, steps (1) and (2) were repeated 100 times, and the average accuracy was taken (see Fig. S1 for a visualization of these procedures). This procedure was applied to all 150 participants.

Second, to examine whether the cognitive architecture of CI is serial, parallel, or cascading, a TGA was performed. The procedure for TGA is similar to the time-by-time decoding analysis, with the key difference being that TGA trains a classifier at each time point and performs generalization testing across all time points, resulting in a temporal-generalization matrix that visually reveals the temporal evolution of the neural representation of CI. The diagonal of the matrix corresponds to the results of the time-by-time decoding analysis, while the off-diagonal results reflect the temporal-generalization characteristics of the neural representations [17, 44].

Subspace analysis

As a supplement to the TGA, this study conducted a subspace analysis to (1) quantify the dynamics of neural subspaces and (2) reveal the temporal specificity of different stages during CI. A subspace refers to a low-dimensional projection of the neural state space, representing a limited dimension within a larger state space [52, 53]. For instance, the response of each neuron or voxel can be viewed as a subspace of the neural population. In this study, principal component analysis (PCA) was used to define and visualize the low-dimensional subspaces that represent CI.

A matrix X was defined with dimensions corresponding to the 2 conditions (creative/common) × n electrodes (i.e., 2 × 60). The matrix X represents the averaged electrode activity across all trials for each condition. Each column of X was mean-centered, resulting in zero-mean columns. According to a previous study [54], for subspace visualization (Fig. 5A), this study ignored the time-varying information by averaging the data across all time-points above chance (i.e., ~ 450–2000 ms, see Fig. 3D), followed by principal component analysis (PCA). To obtain principal components (PCs), eigendecomposition was performed on the covariance matrix \(\mathrm X^{\mathrm T}\mathrm X=\mathrm W\wedge\mathrm W^{\mathrm T}\) , where each column of W is a unit-length vector representing the weights of each PC, and Λ is a diagonal matrix containing the corresponding eigenvalues. The data from each time point was projected into the subspace by computing T = XW (T is usually called a PC score), where X is the matrix (2 × 2) for a single time-point, representing the projection of electrode activity patterns for each condition onto the subspace defined by the first two PCs, PC1 and PC2.

To visualize the neural population dynamics of subspace, PCA was applied separately to three data matrices X_T1, X_T2, and X_T3 (obtained by averaging data within each time window; see Section Temporal-generalized approach results) to estimate each time window subspace (Fig. 5B).

To further quantify the dynamics of the neural subspace, the principal angle (PA) was computed for each participant by performing PCA separately for each condition. PA measures the alignment between subspaces [54, 55]. A larger PA indicates greater misalignment and separation between subspaces. PA was computed using the method proposed by [56]: singular value decomposition (SVD) was applied to the inner product matrix \(\mathrm{W}_{\mathrm{T1}}^\mathrm{T}\mathrm{W}_{\mathrm{T3}}=\mathrm{P}_{\mathrm{T1}}\mathrm{CP}_{\mathrm{T3}}^{\mathrm{T}}\) , where W_T1 and W_T3 are weight matrices for the T1 and T3 subspaces obtained via PCA, with size n_sensors × 2. The matrix C is a diagonal matrix whose diagonal elements are the ranked (from small to large) cosines of the principal angles θ₁ and θ₂: \(C=diag\left(\text{cos}\left(\theta_1\right),\text{cos}\left(\theta_2\right)\right)\) . The first PA was reported in the present study. Similarly, PAs for the T1-T2 and T2-T3 subspaces were also computed (see Table 1; Fig. 5C).

Statistical analysis

For the behavioral data, a paired samples t-test was performed to compare the reaction time and originality scores for the creative and common conditions, respectively.

For the time-by-time decoding analysis, a cluster-based permutation test (CBPT) was employed to assess statistical significance at each time point while controlling for multiple comparisons [57]. The steps were as follows: (1) By comparing accuracy against the chance (i.e., 0.5), a one-sample t-test (α = 0.05) was performed at each time point to identify significant clusters; (2) The sum of t-values within each cluster was computed as the clustering statistic; (3) A null distribution was constructed using the Monte Carlo method with 5,000 permutations; (4) By comparing the cluster statistic to the null distribution, the p-values (0.05, one-sided) were obtained. The same procedure was applied to test the significance of the TGA results.

For ERP univariate analysis, the statistical analysis also employed the aforementioned CBPT, with the following differences: a paired-sample t-test was used in step (1). A cluster was formed if at least two adjacent electrodes were included (α = 0.05, two-sided). The adjacency of electrodes was determined using the “triangulation” method [58]. In step (4), a two-tailed test was applied.

To assess the significance of the PA differences between time windows, a bootstrapping procedure was implemented to construct a bootstrap distribution of PA values for two subspaces within the same time window. The PA between time windows was then compared to determine if it was significantly greater than the PA within the same time window. The bootstrapping was iterated 5,000 times. During each iteration, trials were resampled twice, the subspaces for the same time window were computed, and the PA between them was calculated. This procedure was performed for each time window. If the PA between time windows (e.g., T1-T3) fell outside the top 2.5% of the bootstrap distribution of PA within time windows (e.g., T1 and T3), it was considered significant.