Transcranial direct current stimulation targeting the bilateral IFG alters cognitive processes during creative ideation

Participants

Eighty-one undergraduate students participated in this study (27 males; mean age = 21.53 ± 1.78 years). All participants were right-handed with normal or corrected-to-normal vision, and none had a history of psychiatric illness. Informed consent was obtained from all participants to ensure that they fully understood the experimental procedures and objectives. All collected behavioral data were included in subsequent statistical analyses. However, data from nine participants were excluded from the construction of the prediction model and further analyses because of unqualified EEG signals, as determined through manual inspection using the open-access MATLAB toolbox EEGLAB⁵¹. This study conformed to the principles of the Declaration of Helsinki and was approved by the Ethics Committee of East China Normal University (HR2-0032-2022).

General procedure and tasks

Upon arrival, participants provided written informed consent and were randomly assigned to one of three groups: anodal stimulation on the left IFG and cathodal stimulation on the right IFG (L+R−), reverse stimulation protocol (L−R + ), or Sham stimulation. Before stimulation, participants completed the Runco Ideational Behavior Scale (RIBS)⁵² and the Self-Assessment Manikin Scale (SAM)⁵³. Then, after the application of electrode gel and an electrode cap, they received 15 min of tDCS or sham stimulation. A second application of electrode gel was specifically targeted at the electrode where tDCS was administered. This approach allowed us to implement tDCS efficiently and collect EEG data within a short timeframe. The participants then sequentially performed the Stroop task and CRA after a short break. Following tDCS and the tasks, participants completed the SAM again and reported to the experimenter whether they felt any stimulation on their scalp. The specific configurations of tDCS and tasks applied in this study are described in detail below. Figure 5 illustrates the general procedure used in this study. The software used for electric field modeling in this study (see Fig. 5) was SimNIBS⁵⁴. Most parameters were set to their default values except for current density, electrode position, and electrode size (for details, see “Transcranial direct current stimulation (tDCS) configuration”).

Note. SAM Self-Assessment Manikin scale; RIBS Runco Ideational Behavior Scale, CRA compound remote association task.

The Stroop Task⁵⁵ is widely used for efficiently assessing executive control abilities. Participants were asked to perform the Stroop task to investigate whether tDCS altered their executive control abilities. Generally, the Stroop task requires participants to determine whether the color of a word matches its semantic meaning. Participants were instructed to press the “J” key if there was a match, and the “F” key if there was no match. For instance, if the word “yellow” is displayed in red or blue, participants should press “F.” Conversely, if “yellow” is shown in yellow, they should press “J.” Performance on the Stroop task was quantified based on the reaction time and proportion correct. In this study, the Stroop task consisted of 72 trials, with 32 trials presenting words for which the color matched the meaning (congruent) and 40 trials for which it did not (incongruent). The inteR−stimulus interval was set at 500 ms. We measured Stroop task performance based on the reaction time of the correctly answered trials.

The Compound Remote Associates Task¹⁸ is a widely utilized task that focuses on one aspect of creativity: convergent thinking ability. Unlike open-ended divergent thinking tasks, the CRA is closed-ended but still demands creative cognition. This characteristic allows the investigation of different creative thinking processes while avoiding biases due to answer variability. This study employed a well-established Chinese version of the CRA⁵⁶.

In the CRA, participants were presented with three Chinese characters and were required to identify another character that could form a meaningful compound with each presented character. For example, the characters (also can be deemed words) “词” (word), “运” (transport), and “带” (bring) can be combined with “动” (move) to create “动词” (verb), “运动” (sports), and “带动” (drive). The participants in our study completed 20 CRA trials. Each trial lasted 15 s after a 1 s fixation period. The participants indicated their readiness to answer by pressing a key, after which they entered their responses into an answer box. CRA performance was quantified based on the proportion correct and reaction time for correct answers.

To quantify the relative semantic distance (RSD) for each CRA question, we first utilized a Python-based toolbox, Gensim⁵⁷, to calculate the semantic distance between the three cue words and the correct answer using the Word2Vec algorithm⁵⁸. The corpus employed in our study was obtained from Tencent Corp.⁵⁹, comprising approximately eight billion Chinese words, each represented by a 200-dimensional vector. For each CRA trial, six semantic distances were computed, three of which were the semantic distances between the cues (SD_w) and three of which were the semantic distances between each cue and answer (SD_b). These distances were used to calculate the RSD, as Eq. (1) depicted:

note. where RSD refers to the relative semantic distance, SD_b refers to the semantic distance between the answer and each CRA cue, and SD_w refers to the semantic distance between the CRA cues.

Transcranial direct current stimulation (tDCS) configuration

In this study, a portable, battery-powered 1 × 1 low-intensity transcranial DC stimulator (Soterix Medical, New Jersey) was used to deliver a constant current of 1.5 mA for 15 min, including a 30 s ramp-up time. Stimulation was administered using a pair of saline-soaked sponge electrodes, each measuring 7 × 5 cm². This duration and current density of tDCS have been shown to be effective at influencing creative thinking²⁵. The locations of the left and right IFG were determined as F7 and F8, respectively, according to the 10–20 EEG system.

Electroencephalographic data collection and preprocessing

EEG data were recorded using a Neuroscan EEG system (Compumedics, Australia) with electrodes positioned at FP1, F3, F7, FC3, C3, CP3, P3, P7, T7, and O1 on the left side and their corresponding locations on the right side. Additional electrodes were placed at Fz, FCz, Cz, CPz, and Oz following the 10–20 system. The sampling rate was set at 1000 Hz, and no downsampling was applied in subsequent analyses.

Initially, the EEG data were bandpass-filtered between 1–50 Hz. All data were then re-referenced to the average EEG signals collected from the aforementioned electrodes. Electrodes exhibiting excessively noisy signals, as manually assessed for each participant, were interpolated from neighboring electrodes using spherical spline interpolation. Subsequently, independent component analysis (ICA) was applied to correct for ocular and muscle artifacts.

For each trial, the last 2000 ms of the EEG signals prior to the participant’s response were extracted. In cases for which participants failed to provide an answer within 15 s, signals from 13 to 15 s were used. The data for each trial were then manually inspected and any unqualified trials were discarded from further analysis. It is important to note that participants with fewer than 10 qualified trials were excluded from further analysis, which is consistent with the approach taken in a previous study⁶⁰.

To account for potential superficial correlations in EEG signals due to volume conduction effects, we averaged the signals from the electrodes within the same lobe. Specifically, signals from FP1, F3, and F7 were averaged to represent the left frontal lobe, CP3 and P3 the left parietal lobe, T7 and P7 the left temporal lobe, and O1 the left occipital lobe, as there were no other electrodes in this study located in the left occipital lobe. A similar averaging process was applied to electrodes in the right hemisphere.

Eight signals corresponding to the bilateral frontal, parietal, temporal, and occipital lobes were obtained. From these, 28 functional connectivity (FC) values ((8 × 8 − 8)/2) were calculated using the phase-locking value (PLV) for each trial. PLVs were computed during each trial and during the fixation period preceding each trial. To determine task-related PLV changes, we compared the Fisher Z-transformed task PLVs with the reference PLVs for each trial.

Data analysis

To investigate whether participants could identify the group to which they were allocated, χ² tests were performed. A one-way ANOVA of the RIBS scores was then performed to compare the differences in daily life creativity among the three groups. For the SAM, two repeated-measures ANOVAs were performed across time and groups for valence and arousal. For the Stroop task, we conducted two one-way ANOVAs to investigate the effect of tDCS on the accuracy and reaction time of correct trials.

To explore the impact of tDCS on the bilateral IFG while accounting for triaL−level differences in the CRA task, we constructed four mixed-effect models: two linear and two quadratic models. These models assessed the accuracy of the CRA responses and reaction times. In these models, tDCS group, RSD, gender, Stroop task performance (reaction time for correct trials), and the interaction between tDCS group and RSD were included as fixed effects. We are not interested in the possible effect of practice and fatigue, and so individual differences in the present study, therefore, the order of CRA trials and participant identities were treated as random effects. In the quadratic models, if the quadratic terms (i.e., the square of RSD or the interaction between the square of RSD and the tDCS group) demonstrated significant effects, a breakpoint regression analysis was conducted to confirm either a U-shaped or an inverted U-shaped relationship between RSD and CRA performance. The breakpoint in these analyses was determined using the Robin Hood procedure, as described in a previous study⁶¹. Additionally, if the interaction term between the square of the RSD and the tDCS group was significant, we performed breakpoint regression separately for groups in which the simple effect of RSD² was significant. The rationale for this separate model is that a significant interaction term suggests that the breakpoint for each group may differ, making it inappropriate to summarize all data using a single breakpoint regression model.

The methodology used to construct the models is illustrated in Fig. 6. Broadly, we aggregated trials within each group and employed nested ensemble classification to predict the correctness of the CRA responses. Owing to the inherent randomness of ensemble classification, the same algorithm can produce slightly different optimal models for each iteration. To address this, we constructed 80 models for each group for comparison. The number of models was determined using G*Power. For a 3 (group) × 2 (bands) two-way ANOVA, with an effect size set at 0.25 (medium size) and both alpha and beta error probabilities set at 0.01, a total of 444 samples were required. Therefore, 480 models (2 × 3 × 80) were constructed to satisfy this criterion. In each prediction iteration, random under-sampling was used to balance the number of correct and incorrect trials. Specifically, 20 correct and 20 incorrect trials were randomly selected as test data, whereas the remaining data were undersampled before being used for model training.

Twenty correct and twenty incorrect trials were randomly chosen as test data. Bayesian optimization and 10-fold cross-validation were used to identify the optimal model in each iteration. Eighty models were finally obtained for each group.

Bayesian optimization and 10-fold cross-validation were used to identify the optimal model for each iteration. A model was considered for further analysis only if it significantly predicted the correctness of the CRA trials in both the training and test data, as determined by a permutation test performed 1000 times for generalizability. For each group, a maximum of 1600 iterations were conducted to ensure the statistical robustness of the classification models (with a significance threshold set at 80/1600 = 0.05). This modeling process was performed separately for both the alpha (8–12 Hz) and beta (13–30 Hz) frequency bands.

After completing the modeling process, we first fit the model to all of the datasets to test the specifications of the model. For example, 80 models based on alpha band oscillation from L+R− group were fitted in all six data sets (3 groups and 2 bands), resulting 80 × 6 classifications. One-sample t-tests were conducted to investigate whether the area under the ROC curve (AUC) was significantly higher than the estimated level. We then conducted a two-way ANOVA to compare the AUCs across the different frequency bands and groups. This analysis helped to assess the predictive accuracy of the model under various conditions. To identify stable predictors within each group, we performed a series of paired t-tests. These tests compared the weight of each functional connectivity (FC) against the median weight of the FCs in each model, with Bonferroni correction applied to account for multiple comparisons. FCs with weights significantly higher than the median weight of the model were considered stable predictors.

Finally, we tested the lateralization of the bilateral frontal lobes in the prediction models by comparing the weighted degree of each frontal lobe. The weighted degrees were obtained by summing the weights of the FCs connected to the bilateral frontal lobes. We then compared these weighted degrees across groups and hemispheres. This comparison was performed using two separate mixed ANOVAs: one for the alpha band (8–12 Hz), and the other for the beta band (13–30 Hz). The between-subjects (model) factor was the different groups, and the within-subjects (model) factor was the hemisphere.