Hyperdirect insula-basal-ganglia pathway and adult-like maturity of global brain responses predict inhibitory control in children

Inhibitory control is fundamental to children’s self-regulation and cognitive development. Here we investigate cortical-basal ganglia pathways underlying inhibitory control in children and their adult-like maturity. We first conduct a comprehensive meta-analysis of extant neurodevelopmental studies of inhibitory control and highlight important gaps in the literature. Second, we examine cortical-basal ganglia activation during inhibitory control in children ages 9–12 and demonstrate the formation of an adult-like inhibitory control network by late childhood. Third, we develop a neural maturation index (NMI), which assesses the similarity of brain activation patterns between children and adults, and demonstrate that higher NMI in children predicts better inhibitory control. Fourth, we show that activity in the subthalamic nucleus and its effective connectivity with the right anterior insula predicts children’s inhibitory control. Fifth, we replicate our findings across multiple cohorts. Our findings provide insights into cortical-basal ganglia circuits and global brain organization underlying the development of inhibitory control.


Stanford_Child cohort: participants
A total of 78 fifth grade children were recruited from the East Palo Alto Ravenswood School District and South San Jose Unified School District as part of a longitudinal mindfulness training study. Only baseline data (before training) was used in the current study. Behavioral data and head motion in scanner were screened to ensure high quality of behavioral and fMRI data used in the data analysis. Behavioral and head motion criteria are explained in the behavioral analysis and fMRI preprocessing sections. Two children did not complete the scan. Thirty-eight children were excluded because of excessive head motion (N=33) and/or poor behavioral performance (N=5). The final dataset includes 38 participants (12 female, all right handed with no history of neurological or psychiatric disorders, 9-12 years of age, mean 11.03). No child was under stimulant medication. Other neuropsychological scores were summarized in Supplementary Table S5.

Stanford_Child cohort: behavioral paradigm and analysis
Each child performed two runs of the SST. Subjects were instructed to respond as quickly as possible to green arrows (Go Signal) by clicking with their right pointer or middle finger based on the direction of the arrow. In 25% of the trials, after a variable delay, the green arrow turned red (Stop Signal), indicating that the subject should cancel their response. The delay between the Go Signal and the Stop Signal, the SSD varied across trials in a step-wise fashion and adjusted dynamically to the subject's performance: beginning at 165ms, it decreased by 33ms for a failed stop, and increased by 33ms for a successful stop. There were 2 blocks of 96 trials each, in which 32 stop trials were randomly distributed across the block. Each Go Signal was preceded by a jittered inter-trial-interval, and a fixation cross for a duration of 500ms. For each subject, "Go Accuracy" and "Stop Accuracy" were calculated from the proportion of correct trials in Go trials and Stop trials, respectively. "Go RT" and "Failed Stop RT" were averaged reaction times (RT) in all correct Go trials and all failed Stop trials, respectively. SSRT was calculated from the distribution of Go RTs and the probability of failed Stop trials using an integration method, which is based on the Race Model 1 . Participants with less than 80% accuracy on Go trials, or with greater than 80% or less than 20% accuracy on the Stop trials, or with longer RT in unsuccessful stop trials than go trials, in either fMRI run were excluded from further analysis to ensure accurate estimation of the SSRT.

Stanford_Child cohort: MRI data acquisition
The fMRI data in the Stanford_Child cohort was collected in a single session at the Richard M Lucas Center for Imaging at Stanford University. Images were acquired on a 3T GE Signa scanner using an 8-channel head coil. Each participant was instructed to stay as still as possible during the scanning, and inflatable pillows were placed around the child's head in order to further minimize head movement. Functional images of 29 axial slices, parallel to the anterior/posterior commissure line and covering the whole brain, were acquired using a T2*weighted gradient-echo spiral in-out pulse sequence 2 with the following parameters: slicethickness = 4.0mm, repetition time (TR) = 2000ms, echo time (TE) = 30ms, flip angle = 80º. A high-order shimming method was used prior to data acquisition to reduce blurring and signal loss arising from field inhomogeneities. High-resolution T1-weighted images were acquired using a spoiled-gradient-recalled inversion recovery three-dimensional (3D) MRI sequence with the following parameters: TR = 8.4ms, TE = 1.8ms, flip angle = 15°, FOV = 22cm, matrix=256x192.

fMRI preprocessing
Functional MRI data from all three datasets were preprocessed using SPM8. The preprocessing pipeline included realignment, slice-timing correction, normalization to MNI space, and smoothing using a 6mm full-width half-maximum Gaussian kernel to decrease spatial noise. Maximum displacement was calculated based on parameters from realignment procedure. Subjects whose mean scan-to-scan movement were less than 0.5mm 3 and maximum displacement exceeded 5 mm 4 in either run were excluded from the analysis.

Classification analysis
We examined whether voxel-wise activation pattern within STN ROIs could differentiate between Go and SuccStop. To do so, we applied multivariate classification using the linear support vector machine algorithm (C=1) from an open-source library -LIBSVM (http://www.csie.ntu.edu.tw/~cjlin/libsvm/) as the classification tool. The contrast images of Go and SuccStop in each dataset were used as features in the classification analysis. The performance of classification was evaluated using a leave-one-subject-out cross-validation procedure. In this procedure, a pair of Go and SuccStop images from one subject was selected as a test set. The remaining participants' data was used to train a classifier, which was then applied to the test set to predict whether the images are Go or SuccStop. This procedure was repeated N times (N is the number of subjects in each dataset), with each subject's data used exactly once as a test set. The average prediction accuracy across all test sets is termed as the cross-validation accuracy. A permutation procedure was used to infer statistical significance of the cross-validation accuracy for each cluster and the accuracy difference between clusters in each dataset. Specifically, in each permutation, data labels were randomly shuffled. Crossvalidation accuracies from 500 permutations were used to construct the empirical null distribution in each dataset from which p-values for cross-validation accuracies were obtained.

Task-modulated effective connectivity and behavioral correlation analysis
Task-modulated effective connectivity was computed using seed-based generalized psychophysiological interactions method (gPPI). Seeds were selectively placed in the rAI, rIFC, rMFG and rPreSMA, as we were testing a hyperdirect pathway model, which projects from cortical regions to the STN. The gPPI model consisted of a physiological variable (the raw time series of a seed), multiple psychological variables (hemodynamic response function convolved main effect of condition of interest, Go, SuccStop and UnsuccStop), and multiple interaction variables (deconvolved raw time series of the seed multiplied by main effect of condition of interest, and then convolved with the hemodynamic response function) 5 . Task-modulation effect was computed by subtracting beta values for interaction variables between task conditions (e.g. SuccStop versus UnsuccStop). We computed task-modulated effective connectivity of the STN and computed its correlation with individual's SSRT using Pearson's correlations. Multiple linear regression was used to examine whether any brain-behavior relationships could be driven by other potential confounds such as age, gender and head motion.

Replication with different head motion regressors
We conducted additional analysis to examine whether brain-behavior regression results are stable when controlling for mean rather than maximum frame-wise displacement.

Replication using volume repair analysis
We conducted additional analysis to censor scan volumes with excessive head motion using a similar procedure commonly implemented in previous developmental neuroimaging studies 6,7 . Specifically, volumes with head motion exceeding 0.5 voxels or spikes in global signal exceeding 5% were interpolated using adjacent scans. Then, we conducted the same analyses to examine brain-behavior relation.

Replication with restricted age range in adult samples
We conducted additional analysis to examine whether a restricted age range (18-26 years old) in the adult samples impacts our findings. In the OpenfMRI_Adult1 dataset, only 11 adults are under 27 years old, which has low power in the group level analysis and no voxel survived at corrected threshold (p<0.01, FDR corrected). In the OpenfMRI_Adult2 dataset, there are 22 adults under 27 years old, which allows us to test whether our findings depends on the age range of adult samples.

Replication with additional neuropsychological measures as regressors
We conducted additional analyses to examine whether other potential confounds may explain individual differences in developmental population. We included additional regressors that measure children's cognitive development in the regression analyses, such as Sequential Processing, Simultaneous Processing, Learning Ability, Planning Ability and Mental Processing Index from Kaufman Assessment Battery for Children (KABC) 8 .

Replication using SST fMRI data from the ABCD study
We analyzed SST fMRI data from the first 500 typically developing children (TDC) ages 9-11 in the ABCD study (https://abcdstudy.org/). We downloaded minimal preprocessed data from the ABCD study (two SST sessions per subject), and then normalized it to the Montreal Neurological Institute (MNI) 2mm template (91x109x91) and smoothed the resulting images using a 6mm Gaussian kernel. We then created a task design matrix using the event and onset data provided by the ABCD study and ran the same GLM and gPPI analyses as we did on the Stanford_Child, OpenfMRI_Adult1 and OpenfMRI_Adult2 datasets. Because of missing data or event conditions, we could not complete GLM analysis on 17 participants. Among the remaining 483 participants, we excluded 240 participants because their head motion displacement was over 5mm (the criterion we used in our main analysis). We excluded individuals who had poor behavioral performance (less than 80% accuracy on Go trials, or with greater than 80% or less than 20% accuracy on Stop trials, or with longer RT in unsuccessful stop trials than go trials) and individuals who have outliers in key behavioral measure (i.e. SSRT) and brain measures (e.g. NMI, STN activation and PPI weight). Outliers was defined by more than 3 standard deviations away from the mean. The final sample included 186 TDC (9-11 years old, 104F/82M).

Similarity of brain activation patterns in children and adults during Go and Unsuccessful Stopping
We conducted additional analyses to examine spatial correlation between Child and Adult on Go and Unsuccessful stopping. We found that the spatial activation pattern on Go trials was correlated between Stanford_Child and OpenfMRI_Adult1 cohorts (r=0.36, Supplementary Figure S1) and between Stanford_Child and OpenfMRI_Adult2 cohorts (r=0.41, Supplementary Figure S1). Error-elicited spatial activation pattern was correlated between Stanford_Child and OpenfMRI_Adult1 cohorts (r=0.43, Figure S2) and between Stanford_Child and OpenfMRI_Adult1 cohorts (r=0.31, Figure S2). Overall, the correlation coefficients on Go and Unsuccessful Stop trials was much lower in comparison to spatial correlation during Successful Stop trials (r=0.69 and r=0.71, Figure 3D).
Although STN activation was inversely correlated with SSRT in our child cohort, this relationship was not observed in the two OpenfMRI adult cohorts. Previous findings in adults regarding the relationship between STN and SSRT have also been mixed 17,18,44,25 . Further research at higher field strengths and more precise electrophysiological studies are needed to resolve these discrepancies.

STN activation in relation to inhibitory control in adults
We repeated the same analyses to examine the relation between STN activation and SSRT in the two adult datasets. We did not find a significant correlation between activation in STN and SSRT (p's>0.05, Pearson's correlation, Supplementary Table 11, S12).
It is not completely clear why a significant correlation between rAI-STN connectivity and SSRT emerged in children but not in the two adult cohorts. One possibility is that children may be more dependent on Stop-signal driven modulation of the hyperdirect cortical-STN pathway because of immature proactive and anticipatory control mechanisms. Adults, on the other hand, may better implement proactive control mechanisms, which is less dependent on the hyperdirect cortical-STN pathway. Further research is required to test this hypothesis.

Cortical-STN connectivity in relation to inhibitory control in adults
We repeated the same analyses to examine whether task-modulated functional connectivity between the rAI and STN predicted SSRTs in the two adult datasets. We did not find a significant correlation between task-modulated connectivity between the rAI and STN and SSRT (p>0.05, Pearson's correlation).

Replication with different head motion regressors
Including mean, rather than max, frame-wise displacement, does not have a significant impact in the regression analyses. Multiple linear regression analyses, that included SSRT as the dependent variable and STN activation, age, gender, and mean head motion displacement as independent variables, showed that task-modulated connectivity between the rAI and rSTN was the best predictor for SSRT after controlling effect of age, gender, and head motion (p=0.006) (Supplementary Table 13).

Replication using volume repair analysis
Neural maturity indices in children related to individual inhibitory control abilities We found a significant negative correlation between the NMI and SSRT using a reference map derived from the OpenfMRI_Adult1 reference map (r=-0.36, p=0.03, Cohen's d=0.77) (Supplementary Figure 3). We replicated this finding using the OpenfMRI_Adult2 reference map (r=-0.41, p<0.05, Cohen's d=0.89) (Supplementary Figure 3). To further examine whether this relationship was driven by other potential confounds, we conducted multiple linear regression with SSRT as the dependent variable and NMI, age, gender, and maximum head motion displacement as independent variables. We found that the NMI is the most robust predictor (OpenfMRI_Adult1: p<0.05; OpenfMRI_Adult2: p<0.05) (Supplementary Table 14).
Children's STN activation related to inhibitory action control ability We found a negative correlation between STN activation in stopping and SSRT in the Stanford_Child cohort (lSTN: r=-0.38, p=0.02, Cohen's d=0.82; rSTN: r=-0.36, p=0.03, Cohen's d=0.77) (Supplementary Figure 4). To further examine whether this relationship is driven by potential confounds, we conducted multiple linear regression with SSRT as the dependent variable and STN activation, age, gender and maximum head motion displacement as independent variables. We found that both rSTN and lSTN activation was the significant predictor (ps<0.05) after controlling for effects of age, gender, and head motion (Supplementary Table 15).
Cortical-STN connectivity related to inhibitory control ability We found that stop signal-modulated connectivity between the rAI and rSTN, but not the lSTN, during stopping was significantly correlated with SSRT (r=-0.45, p=0.004, Cohen's d=1.0) in children (Supplementary Figure 5), such that increased rAI-rSTN connectivity was associated with faster SSRTs. Connectivity between other prefrontal nodes and STN were not significantly correlated with SSRT (p>0.05). Multiple linear regression analyses, that included SSRT as the dependent variable and STN activation, age, gender, and maximum head motion displacement as independent variables, confirmed that task-modulated connectivity between the rAI and rSTN was the best predictor for SSRT after controlling effect of age, gender, and head motion (p=0.007) (Supplementary Table S16).

Neural maturity indices in children related to individual inhibitory control abilities
We computed children's neural maturity index (NMI) using a reference map derived from the subset (18-26 years old) of OpenfMRI_Adult2 reference map. We found a significant negative correlation between the NMI and SSRT (r=-0.43, p=0.007, Cohen's d=0.95) (Supplementary Figure 6). To further examine whether this relationship was driven by other potential confounds, we conducted multiple linear regression with SSRT as the dependent variable and NMI, age, gender, and maximum head motion displacement as independent variables. We found that the NMI is the most robust predictor of children's SSRT (p<0.01) (Supplementary Table 17).

Replication with additional neuropsychological measures as regressors
After adding additional neuropsychological measures from KABC as regressors, we replicated our findings: (i) the NMI remains to be the most robust predictor (Supplementary Table 18), (ii) the STN activation is the most robust predictor (Supplementary Table 19), (iii) the effective connectivity between rAI and rSTN is the most robust predictor of children's SSRT (Supplementary Table 20).

III. SI Figures
Supplementary Figure 1. Brain-wide activation patterns elicited by motor execution (Go trials) in children were related to activation patterns in two different cohorts of adult. Each data point represents one voxel's activation (beta contrast of Go vs. baseline) in gray matter mask. Source data are provided as a Source Data file.
Supplementary Figure 2. Brain-wide activation patterns elicited by errors in children were related to activation patterns in the two different adult cohorts. Each data point represents brain activation in one voxel's (beta contrast of UnsuccStop versus SuccStop trials) in the gray matter mask. Source data are provided as a Source Data file.
Supplementary Figure 3. NMI in children was negatively correlated with SSRT (replication after volume repair analysis). This relationship was replicated using reference maps from the two adult cohorts. Each data point represents one child. Source data are provided as a Source Data file.
Supplementary Figure 4. STN activation during Stopping was negatively correlated with SSRT in children (replication after volume repair analysis). Each data point represents one child.
Source data are provided as a Source Data file.
Supplementary Figure 5. Effective connectivity between the rAI (seed) and rSTN (target) was negatively correlated with SSRT in children. No such relation was observed in the left hemisphere (replication after volume repair analysis). Each data point represents one child. Source data are provided as a Source Data file.
Supplementary Figure 6. NMI in children was negatively correlated with SSRT. The reference map was derived from the subset (18-26 years old) of the OpenfMRI_Adult2 cohorts. Data points represent individual children. Source data are provided as a Source Data file.
Supplementary Figure 8. STN activation during Stopping was negatively correlated with SSRT in children ages 9-11. Replication using data from the NIH-ABCD study. Each data point represents one child. Source data are provided as a Source Data file. Tables   Supplementary Table 1. Published studies and contrasts between task conditions included in meta-analysis of inhibitory control tasks in children. Most studies in children have used the Go/NoGo task.