Twenty-seven participants were recruited in this study. They were all right-handed native Chinese college students with normal or corrected-to-normal vision. None of them reported any neurological disorders. Data from 2 participants were excluded due to falling asleep or exhibiting a significantly low accuracy rate (less than 50%) in the writing sequence processing task during MRI scanning. This left data of 25 participants (14 females, mean age = 23.04 years, SD = 2.93) for the final analyses.
Forty-two Chinese characters: 14 single-component structured (e.g., “车”), 14 left-right structured (e.g., “阳”), and 14 top-bottom structured (e.g., “拿”) were selected as the experimental stimuli. Single-component characters refer to those written by inner writing sequences of strokes, for example, horizontal strokes before vertical. Similarly, left-right structured and the top-bottom structured characters refer to those written by typical rules between strokes, left before right and top before bottom respectively, as shown in Figure 1. They were further divided into two subsets that were well-matched on character frequencies, and stroke numbers.
To simulate a real writing process, mini videos were made to animate the intact and scrambled writing process of each Chinese character. In the intact writing process, the order of each stroke complied with the sequences stipulated by the State Language Commission (2000). In the scrambled writing sequence, the order of each stroke was rearranged except the first and the last stroke, deliberately breaking its original integral writing rules (left before right, top before bottom, middle before surrounding, etc.) to violate the highest schematic information of writing sequences, as shown in Figure 2B.
The study was approved by the Institutional Review Board of Beijing Normal University. All participants signed a consent form before the experiment and received a debriefing form afterward. Before the formal experiment, participants performed a practice session of 12 trials. In each trial, a fixation cross was presented for 600 ms followed by a blank screen for 600 ms. Each static character was presented for 400 ms in gray allowing participants to retrieve its orthographic and semantic information so that activation associated with lexical processing that we were not interested in would remain similar across two experimental conditions (Yu et al. 2011). Then, the animation of the writing sequence was presented, in which the duration of each stroke mimicked writing progress by turning from grey to black, was 400 ms, and the integrated Chinese character was also presented in black for 400 ms at the end of each animation.
Participants were instructed to judge whether the writing sequence was correct or not when the blank screen was presented. A randomized inter-trial interval of 2-5 s was applied after presenting the entire writing process of each character. They were required to press a button with either their left or right index finger. The response fingers for correct or incorrect writing sequences were counterbalanced. The experimental procedure is shown in Figure 2A.
The formal experiment lasted approximately 60 minutes with 6 runs. In each run, participants watched 42 animations in a randomized order with 21 intact and 21 scrambled writing sequences. The stimuli for all 6 runs were identical but presented in different orders. After each run, participants could take a break as long as necessary. The functional scanning took about 50 minutes in total, followed by a 6-minute anatomical scanning for each participant.
Previous studies have found that processing Chinese character writing sequences might recruit executive functions as top-down modulation (Yu et al. 2011). Therefore, based on our meta-analytic decoding of the fMRI results, after MRI scanning, participants were called back to perform a task switching task (Kang et al. 2020) to verify the role of executive functions in processing of Chinese character writing sequences. Specifically, participants were required to judge the parity (odd or even) or the magnitude (larger or smaller than 5) of a digit presented in either a square frame or a rhombus frame, which served as a cue. Participant were told to judge the parity (odd or even) of the number following a square frame and to judge the magnitude (larger or smaller than 5) following a rhombus frame as quickly and accurately as possible. The mappings of cues and task requirements were counterbalanced across participants. During the task, each trial began with a fixation in the center of the screen (“+”) for 300 ms, followed by a blank screen for 200 ms. The geometric frame cue would then appear at the center for 500 ms, after which a single digit (1-4 and 6-9) would be presented in the middle of the cue. The digit and the cue would disappear together when any response was made. Then, a blank screen for 1000 ms would appear before the next trial.
The two types of cues were pseudo-randomized, resulting in switch and nonswitch conditions. The cue-task and the response-key mapping were counterbalanced among participants. Participants did a practice session to familiarize themselves with switching between tasks. We calculated switch cost in both accuracy and response times, i.e., difference between switch and nonswitch trials.
Functional images were acquired by a Siemens Prisma 3-T MRI scanner with an interleaved multiband EPI sequence with the following parameters: multi-slice factor = 2, TR = 1000 ms, TE = 29 ms, flip angle = 70°, FOV = 200 × 200 mm², matrix size = 64 × 64, resolution within slices = 3.1 × 3.1 mm², and slice thickness/gap = 4 mm/0.6 mm. There were 378 TRs for each experimental session in total for all conditions. High-resolution T1-weighted anatomical images were also obtained using the following scan parameters: TR = 2530 ms, TE = 2.27 ms, flip angle = 7°, FOV = 256 × 256 mm², matrix size =256 × 256, resolution within slices = 1.0 × 1.0 mm², slice thickness = 1 mm, and number of slices = 207.
The fMRI data pre-processing was conducted by the SPM 12 (Wellcome Department of Cognitive Neurology, London, UK, http://www.fil.ion.ucl.ac.uk/spm) based on MATLAB 2018b (The MathWorks Inc). The first 8 functional images (corresponding to the first filler trial) were removed due to the instability of the magnetic field. The data preprocessing procedure included the following steps: First, the slice timing corrections were performed. Next, head motion corrections were applied to exclude participants with translational head motion greater than 2 mm or 2°, and no participants were excluded. Besides, we coregisterered all anatomical images to the mean EPI image obtained during realignment. Then, all images were normalized to the EPI template on the basis of Montreal Neurological Institute (MNI) space template and resampled into 3-mm cubic voxels. Finally, 184 TRs (184 s) of data were approximately involved in the formal analysis averaged for each session of each condition.
Multivoxel pattern analysis
First, unsmoothed images were put into a general linear model (GLM) to conduct an individual first-level analysis for each participant to estimate the beta images for correct or incorrect writing sequences. Therefore, for each participant, there were 2 beta images for two experimental conditions, respectively. We generated a mask with voxels of higher than 30% probability based on the SPM original gray matter probability template. Within that mask, a whole-brain searchlight analysis (Kriegeskorte et al. 2006) was conducted by taking these voxels as centers of spheres with a 6-mm radius and calculated their beta values in 6 runs of all participants.
For each searchlight sphere, we applied the support vector machine (SVM, the binary linear vector algorithm in the fitcsvm function of the MATLAB Statistics and Machine Learning Toolbox, https://www.mathworks.com/help/stats/fitcsvm.html) to extract unique brain discrepancy maps (Wang et al. 2007). SVM has been proven to be valid in analyzing high dimensional data even if the sample size is small (LaConte et al. 2005). We used the beta images for 24 participants in two experimental conditions to train the classifier and test it on data from the last participant based on leave-one-out cross-validation, where the data from 24 participants were used for training and the remaining one participant’s data for testing (Alink et al. 2012; Haynes and Rees 2006). The same process was repeated for each participant so that the SVM assigned a weight to each voxel indicating its testing accuracy and an accuracy image could be obtained for each participant.
Finally, we spatially smoothed all images using an isotropic Gaussian kernel with a 6-mm full width at half-maximum. We executed the nonparametric one-sample t-test by the statistical nonparametric mapping toolbox (SnPM, https://warwick.ac.uk/snpm) to identify if the classification accuracy of each voxel is higher than the chance rate (50%) on the group level. The criteria were set to FWE corrected p < 0.05 on the voxel level and the variance smoothing was set to 0. We set the permutation to occur 5000 times.
Meta-analytic functional decoding
To explore the relationship between the brain mechanism of Chinese character writing sequences and specific psychological components of executive functions, we used the online Neurosynth Image Decoder (http://www.neurosynth.org, Rubin et al. 2017) to decode the MVPA filtered image. Based on our assumption, we chose 9 terms (including rule, sequence, spatial attention, action observation, motor imagery, expectation, executive function, shifting, inhibition, and updating) covering the writing sequence stimuli features and executive functions. The criteria were set to FDR corrected p < 0.01 automatically by the platform.
Effective connectivity analysis
To get the connectivity network of writing sequence perception, 10 nodes were defined by drawing spheres with the coordinates of the brain regions found in the MVPA as centers and radius of 6 mm. We extracted time-series of spheres by the Resting-State fMRI Data Analysis Toolkit (REST, https://www.nitrc.org/projects/rest/, (Song et al. 2011) and defined them as nodes in connectivity analysis. The node centers were the peak voxel in clusters with significantly higher than 50% classifying accuracy values for intact and scrambled writing sequences in the MVPA results.
We employed the extended unified structural equation model (euSEM) to construct an effective connectivity model (Gates et al. 2011). Group iterative multiple model estimation (GIMME) was used to automatically conduct model selection and fitting in practice. The SEM was only qualified to build relationships without time-series order between regions of interest (ROIs) with the assumption that the activation of ROIs was independent. However, the BOLD signals detected by fMRI had a sequential relationship with the time series. Therefore, Kim et al. (2007) developed a unified SEM (uSEM) for block-designed fMRI experiments that combine some lagged relationships with time series and that was also the basis of the euSEM along with its consideration of the task effects and bilinear effects. Therefore, it also fits for ER-designed fMRI studies (Gates et al. 2011).
For our study, Lagrange multiplier equivalents were used to carry out the model selection to free each path, after which we decided if the particular freed connection could improve the holistic model fitting of at least 75% of the participants. Afterward, the model was strictly pruned on the group level. In this step, the connections that did not exist in 75% of participants after being freed were removed. The paths that passed the group level pruning were then freed on the individual level based on the semi-confirmatory manner of freeing. Eventually, those connections that were no longer significant after other connections were freed were trimmed and tested if they fit the confirmatory model.
For model convergence, one participant was removed. Based on prior selected model fitting parameters describing reliability, 2 norms were fit in the final model: confirmatory fit index (CFI) > 0.90 and nonnormed fit index (NNFI) > 0.90 (Gates et al. 2011).
We separated the network into subdivisions by the Louvain community detection algorithm with added finetuning so that the network would be easier to read and evaluate (Rubinov and Sporns 2011). This algorithm was built to maximize the number of within-group edges and to minimize the number of between-group edges.
We then applied the optimal core-periphery subdivision algorithm to divide the connectivity network into a core and a periphery group. This approach maximized the edges in the core group and minimized the edges in the peripheral group (Rubinov and Sporns 2010). The cores were named as hubs of the brain network (Braun et al. 2015).
Subsequently, for the core hubs, we calculated their local efficiency, i.e., the inverse of the average shortest path connecting all neighbors of a particular vertex (Rubinov and Sporns 2010), also defined as a measure of how efficiently the node exchanges information in the network system (Latora and Marchiori 2001).
Pearson correlation between the local efficiency and the switching cost of the task-switching task was calculated to examine whether there is any potential relationship between neural correlates for processing Chinese character writing sequences and the shifting component of executive functions.
Data and Code Availability Statement
Data and codes are available and can be currently previewed on Mendeley (https://data.mendeley.com/datasets/8k2857yrch/draft?a=542beca1-a9fd-4cb1-8110-c9408d25f941). The dataset and codes will be open and accessible to readers online.