Connectome‐Based Predictive Models of General and Specific Executive Functions

ABSTRACT

Executive functions, the set of cognitive control processes that facilitate adaptive thoughts and actions, are composed primarily of three distinct yet interrelated cognitive components: Inhibition, Shifting, and Updating. While prior brain research has examined the nature of different components as well as their interrelationships, fewer studies examined whole‐brain connectivity to predict individual differences for the three cognitive components and associated task scores. Here, using the Connectome‐based Predictive Modelling (CPM) approach and open‐access data from the Human Connectome Project, we built brain network models to successfully predict individual performance differences on the Flanker task, the Dimensional Change Card Sort task, and the 2‐back task, each putatively corresponding to Inhibition, Shifting, and Updating. We focused on grayordinate fMRI data collected during the 2‐back tasks after confirming superior predictive performance over resting‐state and volumetric data. High cross‐task prediction accuracy as well as joint recruitment of canonical networks, such as the frontoparietal and default‐mode networks, suggest the existence of a common executive function factor. To investigate the relationships among the three executive function components, we developed new measures to disentangle their shared and unique aspects. Our analysis confirmed that a shared executive function component can be predicted from functional connectivity patterns densely located around the frontoparietal, default‐mode, and dorsal attention networks. The Updating‐specific component showed significant cross‐prediction with the general executive function factor, suggesting a relatively stronger role than the other components. In contrast, the Shifting‐specific and Inhibition‐specific components exhibited lower cross‐prediction accuracy, indicating more distinct and specialized roles. Given the limitation that individual behavioral measures do not purely reflect the intended cognitive constructs, our study demonstrates a novel approach to infer common and specific components of executive function.

Using connectome‐based modeling, we predict novel individuals' executive function (EF) performance and characterize the functional networks underlying specific EF components—inhibition, shifting, and updating—as well as a general EF factor shared across them.

1. Introduction

Executive function (EF), also referred to as cognitive control or executive control, is the process of aligning one's cognition and behavior with current goals (Baddeley 2000; Diamond 2013; Miller and Cohen 2001; Seeley et al. 2007). EF may be categorized into three main processes (Miyake et al. 2000; Lehto et al. 2003): Inhibition, Shifting, and Updating. Inhibition involves actively suppressing the dominant response and is assessed through tasks like the Stroop task (Stroop 1935), the Flanker task (Eriksen and Eriksen 1974), and the Stop Signal task (Logan and Cowan 1984). Shifting, sometimes called cognitive flexibility, occurs when individuals switch between multiple tasks, measured through tasks like the Wisconsin Card Sorting Test (Berg 1948), the Dimensional Change Card Sort (Frye et al. 1995), and the Color/Shape Switching tasks (Hayes et al. 1998). Lastly, Updating involves continually integrating relevant information into our finite working memory while discarding those that are no longer useful, which can be measured via tasks like the N‐Back task (Kirchner 1958), the Letter Memory task (Morris and Jones 1990), and the Keep Track task (Yntema 1963).

The relationship between different forms of EF has been extensively studied in behavioral experiments. Using large batteries of cognitive tasks, Miyake et al. (2000) originally discovered that performance in all the tasks can be loaded onto the three factors: Inhibition, Shifting, and Updating. However, these three factors are intercorrelated. To refine the model, Friedman and Miyake (2017) replaced the original Inhibition factor with a general EF factor that is highly loaded by all tasks, resulting in a model with a better fit and more orthogonal factors, which they termed as the “Unity and Diversity” of EF (Friedman and Miyake 2017). The notion that different EFs are interdependent has also been reported in other studies. For instance, people with superior working memory performance are more likely to achieve a higher Inhibition score, suggesting that successful working memory may rely on actively inhibiting distractors irrelevant to one's goals (Conway et al. 2001; Kane and Engle 2003; Unsworth et al. 2004; Kane et al. 2007). Similarly, Shifting also displays a positive relationship with working memory (Baddeley et al. 2001; Emerson and Miyake 2003) and Inhibition (Mayr and Keele 2000; Koch et al. 2010).

Beyond the behavioral level, researchers have long strived to characterize the neural systems underlying each EF component. Earlier lesion studies indicated that damage to the prefrontal cortex may impair performance in different EF tasks, but the exact anatomical locations differed across studies and the type of process targeted (Aron et al. 2003, 2004; Floden and Stuss 2006; Barbey et al. 2013). Empirical studies and meta‐analyses of fMRI data have also revealed multiple brain regions such as the lateral prefrontal cortex, posterior parietal cortex, and dorsal anterior cingulate cortex with both similar and different roles across various EF factors (Assem et al. 2024; Derrfuss et al. 2005; Jamadar et al. 2010; Kim et al. 2012; McNab et al. 2008; Niendam et al. 2012; Rodríguez‐Nieto et al. 2022). Discrepant results across studies may lead to different conclusions being drawn about the relationship between EF components, such as the superordinate role of the Updating component (Lemire‐Rodger et al. 2019; Rodríguez‐Nieto et al. 2022) or the existence of a core cognitive control network in the brain (Niendam et al. 2012).

While univariate fMRI analyses inform us about the brain regions whose activity is modulated by task conditions, connectivity‐based approaches enable us to peek into the interaction between different brain regions and how they may be related to individual differences in task performance (Friedman and Robbins 2022; Menon and D'Esposito 2022). In the realm of EF, researchers have shown that the prefrontal cortex's global functional connectivity as well as structural connectivity profiles can predict one's EF performance (Cole et al. 2012; Smolker et al. 2015). Furthermore, performance in the three types of EF tasks may correspond with different sets of functional connectivity profiles: Inhibition was highly related to the connectivity strength between the frontoparietal and cingulo‐opercular networks (Deck et al. 2023); Shifting scores were correlated with the strength of connectivity within the cingulo‐opercular network (Reineberg and Banich 2016), between the default mode and the dorsal attention networks (Deck et al. 2023), and between the medial frontal and default mode networks (Chén et al. 2019). Updating performance was reflected in the frontoparietal network connectivity profiles (Reineberg and Banich 2016), and between the angular gyrus and ventral attention network (Reineberg et al. 2015). Another insightful study found that connectivity between the cerebellum and the frontoparietal network predicts a general EF score (Reineberg et al. 2015), but their analysis was limited to resting‐state data, and the use of an ICA approach restricted their focus to highly synchronized brain regions.

In this study, using Connectome‐based Predictive Modeling (CPM), we analyze whole‐brain connectivity features to study the unity and diversity of the EF task measures of Inhibition, Shifting, and Updating. The CPM method (Finn et al. 2015; Shen et al. 2017) has become a mainstream method for predicting individual differences in cognition. CPM has previously shown exceptional capability in predicting individual cognitive performance such as sustained attention (Rosenberg et al. 2016; Yoo, Rosenberg, Kwon, et al. 2022), fluid intelligence (Finn et al. 2015; Yoo et al. 2019), creativity (Beaty et al. 2018), and working memory (Avery et al. 2020). Here, we extend the application of CPM to predict executive function in novel individuals. For our analyses, while both resting‐state and 2‐back fMRI data were used to build CPMs, we focused our analyses on using 2‐back task‐based fMRI data over resting‐state data for building CPMs due to its higher data quality (Huijbers et al. 2017), greater test–retest reliability (Kristo et al. 2014; Rosazza et al. 2014; Wang et al. 2017), and superior predictive accuracy (Greene et al. 2018, 2020; Yoo et al. 2018).

The behavioral scores and neuroimaging data used in this project were obtained from the open‐access Human Connectome Project (HCP) dataset (Van Essen et al. 2012). We used the Flanker task as an approximate measure for Inhibition, the Card Sort task for Shifting, and the 2‐back task for Updating. These EF component assignments to tasks are rough proxies, as we assume task impurity, that is, any given task involves multiple EFs and other cognitive processes (Miyake et al. 2000). Indeed, our goal is to use CPM as a novel approach to tease out shared versus distinct EF components of these behavioral tasks.

Using the CPM trained on 2‐back fMRI data, we explored the unity and diversity of EF at the connectome level from two distinct perspectives. First, we developed three CPM task models based on the three raw EF behavioral measures and achieved significant predictive performance. For each task model, we further identified its underlying canonical functional networks. We also evaluated each task model's cross‐prediction performance to the other two EF task measures, as this allowed us to explore the separability and interdependence of the underlying EF components. That is, when a given model can also predict individual differences in task scores different from the one used for model training, then we can infer a general (shared) executive function component. Similar logic was used to identify a general attention component that explained performance across diverse tasks involving attention (Rosenberg et al. 2016; Yoo, Rosenberg, Kwon, Scheinost, et al. 2022).

Secondly, to further address the limitation that individual behavioral measures do not purely reflect the intended cognitive constructs, we developed measures of general and specific EF. General EF, corresponding to unity, can be estimated by extracting out the common variance of the three EF components (Friedman and Miyake 2017; Miyake et al. 2000). Specific EF, corresponding to the diversity of EF, can be estimated by the task‐specific residual variance. CPM analysis on these new measures confirms the existence of general and specific EF components and their connectome basis. Also, going beyond past work, our models are the first that can predict general and specific EF performance in novel participants not used in training.

2. Materials and Methods

2.1. FMRI Data

The data for this project comes from the WU‐Minn Human Connectome Project (HCP) (Van Essen et al. 2012) S1200 Release of February 2017. Among the various neuroimaging modalities available in the dataset, we specifically used the 3 T resting‐state fMRI scans and 2‐back fMRI data. Detailed scanning parameters for all fMRI sessions are available in Van Essen et al. (2012).

Our initial filtering process retained subjects who had fully completed the resting‐state and working memory (n‐back task) fMRI sessions, as well as the full set of out‐of‐scanner NIH Toolbox tasks (NIH Toolbox; n = 445 removed from this step). We excluded participants whose motion parameter file is missing or who exhibited excessive head motion, defined as exceeding any of the following thresholds: 3 mm translation, 3° rotation, or 0.15 mm mean framewise displacement (n = 116 removed). Furthermore, we excluded subjects with fMRI data flagged for known defects by the HCP team (n = 10 removed). Consequently, our final sample comprised n = 635 (female: 355) subjects. Our sample size is comparable or larger than that of previous CPM studies (e.g., Avery et al. 2020; Yoo, Rosenberg, Kwon, Lin, et al. 2022) and provides a power close to 1 even with minimal effect size and significance level (effect size = 0.1, $\alpha$ = 0.0001).

For each included participant, two resting‐state sessions were available, collected across two separate days. Each session consists of two runs (about 15 min each) with different phase encoding directions: left‐to‐right (LR) and right‐to‐left (RL). The n‐back fMRI data was collected over two sessions, each lasting approximately 5 min and using 2 different phase encoding directions, the same as in the resting sessions. The n‐back fMRI session included both 2‐back and 0‐back blocks. Importantly, we only used the fMRI timeseries associated with the 2‐back trials, since the 0‐back trials do not significantly engage working memory update, but primarily involves passive target detection (See Miller et al. 2009 for further discussion). To reflect this change, we will refer to the fMRI data as “2‐back fMRI data”.

To compare the robustness of fMRI data representations, we tried both the volumetric (NIFTI) and grayordinate (CIFTI) fMRI data for our analysis. Both types of resting‐state fMRI data were processed using the HCP minimal preprocessing pipeline (Glasser et al. 2013). The volumetric data was registered to 2 mm MNI space, while the grayordinate data was additionally transformed to the standard CIFTI grayordinate space. Further preprocessing of the volumetric data involved customized Python code to remove 12 motion parameters, white matter and cerebral spinal fluid (CSF) signals, global signals, and linear trends in the timeseries data. For the resting‐state grayordinate fMRI data, we used a version further denoised by an additional ICA‐FIX procedure (Salimi‐Khorshidi et al. 2014), which enhances the signal‐to‐noise ratio by isolating and removing independent components linked to motion and other artifacts. Additionally, white matter and CSF signals, global signals, and linear trends were regressed out for consistency with the volumetric preprocessing.

The 2‐back volumetric fMRI data underwent the same preprocessing procedures as the resting‐state data. However, for the 2‐Back grayordinate data, ICA‐FIX was not applied due to insufficient data to train the denoising classifier. Instead, we regressed out the 12 motion parameters as done in the volumetric case, followed by the same nuisance variable regression for white matter and CSF, global signal, and linear trend.

2.2. Behavioral Data

The behavioral data for this project were derived from the performance measures for the N‐back task, the Dimensional Change Card Sort Test (DCCS), and the Flanker task from the HCP dataset, which serve as task proxies for Inhibition, Shifting, and Updating components of EF (Diamond 2013; Friedman and Miyake 2017). Note that the out‐of‐scanner List Sorting scores were excluded as a measure of Updating, as the task does not consistently require participants to replace old information with newly acquired ones, which is a core aspect of the definition of Updating (Friedman and Miyake 2017). Additionally, the 2‐list condition in the list sorting task may introduce Shifting demands, as participants alternate between object and animal categories, further making the task less suitable as a measure of Updating.

Every subject included in our neuroimaging sample had all 3 behavioral scores available. All measurements were normalized to have a mean of 100 and a standard deviation of 15, following the standard normalization procedure outlined in the NIH Toolbox manual. The n‐back task was conducted inside the scanner during the working memory session, whereas the DCCS and Flanker tasks were completed outside during a “NIH Toolbox Behavioral Tests” session (see the NIH Toolbox Scoring and Interpretation Guide, n.d.). The in‐scanner working memory task was conducted over two runs, each lasting about 5 min (~10 min total). As noted above, we focused exclusively on the 2‐back trials, so the effective duration of the data used was approximately 5 min. In contrast, the out‐of‐scanner DCCS and Flanker tasks were each completed once, with the DCCS taking approximately 4 min and the Flanker around 3 min. Both accuracy and response time were recorded for all 3 tasks and integrated into a normalized score for each subject. The scores for DCCS and Flanker were provided by the HCP team and preprocessed in accordance with the procedures detailed in the NIH Toolbox Scoring and Interpretation Guide found in the reference section. For the 2‐back task, we processed the 2‐back accuracy and response time for each individual following the same procedure using our customized Python code. The only modification we made was to lower the accuracy threshold from 4 to 2 when combining accuracy and threshold, to ensure data normality (Figure S14). Notably, in an additional analysis, we confirmed that our main findings are robust to this thresholding choice: using the same threshold as in the Flanker and DCCS yielded comparable CPM prediction performance to the version with the adjusted threshold (Table S11).

2.3. Connectome‐Based Predictive Modelling

Connectome‐based Predictive Modelling (CPM) (Finn et al. 2015; Shen et al. 2017) links individual differences in brain functional connectivity and behavior measures.

To construct the functional connectivity matrices, we used two well‐established atlases tailored to the two fMRI data formats in our study: Shen268 (Shen et al. 2013) whole‐brain atlas for parcellating the volumetric data, and Schaefer300 atlas (Schaefer et al. 2018) for the parcellation of the cortical part of the grayordinate data. We selected the Shen268 atlas for its extensive use in previous CPM analysis on volumetric fMRI data (Avery et al. 2020; Beaty et al. 2018; Finn et al. 2015; Rosenberg et al. 2016; Yoo et al. 2018) and the Schaefer300 atlas for its comparable parcellation size in the cortical surface space. Subcortical parcels in the grayordinate data were adopted from the original CIFTI labels. The average timeseries within each ROI was computed to represent the activity at that node, and pairwise Pearson's correlation of all nodes was used to generate the functional connectivity matrix for each subject. Each Pearson r value underwent Fisher transformation to obtain a z value. For ease of calculation, we vectorized each individual's connectivity matrix and concatenated them to form the connectivity matrix for the entire sample. It is important to note that each scanning session (e.g., REST1, REST2, WM) yields two connectivity matrices corresponding to the two different phase encoding runs (L‐R and R‐L). We later averaged these two matrices to produce a single connectivity matrix for that session, an approach supported by past work on HCP data (Cao et al. 2023; Smith et al. 2013). In addition, for resting‐state functional connectivity data, we averaged the connectivity matrices from both sessions to create a single resting‐state functional connectivity matrix per individual, following prior work on brain‐based prediction using HCP data (Greene et al. 2018; Yoo, Rosenberg, Kwon, Scheinost, et al. 2022).

2.4. CPM Prediction of Executive Functions

We constructed CPMs for the three EF components—Inhibition, Shifting, and Updating—by using behavioral measures from the Flanker task, the Dimensional Change Card Sort (abbreviated as “Card Sort” below) task, and the 2‐back task, respectively.

The CPM approach involves two principal stages: feature selection and model fitting. During feature selection, we used Pearson's correlation to associate the connectivity edges with the behavior measure, identifying the correlation score between each edge and the selected behavior. Only edges surpassing the significance threshold (p < 0.01, two‐sided) were kept for model fitting. This process distinguishes two types of edges for selection: positively predictive edges (referred to as “positive edges” hereafter) and negatively predictive edges (“negative edges”), based on the sign of their correlation with the behavior scores. To control for variations in parcellation atlas sizes and further constrain the model to the most predictive edges, we retained only the top 100 significant positive and negative edges for subsequent analysis. Next, to fit the model, we summed together the selected positive and negative edge weights for each participant to generate an aggregate positive score and an aggregate negative score, respectively. Three linear regression models are then developed: one using the positive score only (“positive model”), one using the negative score only (“negative model”), and one using both scores (“both model”).

To evaluate predictive performance without overfitting, we employed a 10‐fold cross‐validation method. Specifically, we shuffled and divided the data into 10 equal parts, training the CPM on 9 of them and testing it on the remaining one. This procedure is iterated 10 times, ensuring each fold is used for testing exactly once. The training set is utilized for feature selection and model fitting, whereas the testing set is for assessing the model's performance. Pearson's correlation between the predicted and actual behavioral scores is computed to gauge the model's prediction accuracy. The average correlation across all 10 folds serves as the final measure of model fit. This process was repeated 1000 times, with a different shuffle each time, to confirm the reliability and reproducibility of our findings. The mean model fit score across these 1000 iterations was reported as the model's final score. Additionally, a permutation test follows the same steps but with the subject's behavioral scores randomly shuffled before being split into folds, ensuring a rigorous evaluation of model performance.

We assessed the within‐task prediction accuracy of each CPM—evaluating the prediction for the same behavioral task measure on which the model was trained. This approach allows us to directly ascertain whether the individual differences in functional connectivity identified by the CPM model are robust enough to accurately predict a new subject's performance.

In addition, we explored cross‐prediction across every pair of EF tasks to evaluate the generalizability of each of the three EF CPM models to predict individual differences in a different EF task measure. Cross‐predictions also allow us to infer the separability of different EF components, namely what's common and what's specific across tasks. For cross‐validation, during each iteration, we maintained the exact same train‐test split and applied the previously trained CPM model to predict a different task measure within the testing cohort.

2.5. CPM Canonical Network Analysis

We also examined the localization of the features selected in each CPM to evaluate the contributions of specific functional networks on each EF measure. Due to the variability in edge selection across iterations of CPM training resulting from different training to testing splits, we included an edge only if it was selected in more than half of the total iterations (over 500 out of 1000 iterations). This approach helps preserve only core edges that reflect meaningful individual differences in behavior. To confirm that the choice of threshold did not qualitatively change the results, we varied the thresholds between 40% and 60%, but did not see substantial changes in the edges within or between canonical networks (Figure S15).

To define canonical networks on volumetric data, we used the 10‐network version of the Shen 268 atlas, where each parcel is categorized into one of the medial frontal, frontoparietal, default mode, motor, visual A, visual B, visual association, salience, subcortical, and cerebellum networks. For grayordinate data, we employed the 7‐network version of the Schaefer 300 atlas for cortical regions. For subcortical structures, we divided them into two separate networks based on their original label given by the CIFTI file: a subcortical network and a cerebellum/brainstem network. Furthermore, since the subcortical network includes regions such as the hippocampus and thalamus, which are traditionally classified within the limbic network, we combined these regions with the limbic network as defined in the Schaefer 300 atlas and retained the designation “subcortical network” for consistency.

2.6. CPM Computational Lesion Analysis

To further assess a canonical network's contribution and importance to prediction performance, we performed a computational lesion analysis (also known as ablation analysis). In this approach, we removed all connectivity patterns associated with a specific network and evaluated how this affected the predictive performance of CPM. Specifically, for each target network under investigation, we retrieved the CPM model trained using the standard procedure, removed all within and cross‐connectivity edges associated with the target network, and then predicted outcomes on the hold‐out testing set using the same cross‐validation procedure.

2.7. General and Specific Executive Functions Analysis

We conducted additional analyses to address the following two questions. First, is there an overarching functional connectome for general EF (Unity; Friedman and Miyake 2017)? Second, given that the original measures may not purely reflect the underlying EF constructs (as indicated by their high correlation), can we identify the functional connectome unique to each construct (Diversity; Friedman and Miyake 2017)?

To address these questions, we devised two types of measures: a general EF score and three component‐specific scores of Inhibition, Shifting, and Updating. The general EF score was derived from the mean of the original z‐scored task measures, where we hope to retain only the core information shared by different EF components. On the other hand, the component‐specific scores were defined as the residuals after regressing out the other two measures from each target measure, with the aim to preserve only the variance that cannot be explained by other measures. We created CPMs for the general EF score and the three component‐specific scores using the same training and testing procedures as above.

2.8. Cross‐Prediction Based on the General Attention CPM

Successful executive functioning involves constant and adaptive attentional control (Engle 2002; Engle and Kane 2003; Kane and Engle 2002). To investigate whether executive function (EF) and attention share a common connectome basis, we applied a CPM developed by Yoo, Rosenberg, Kwon, et al. (2022) to predict a general attention measure. Yoo, Rosenberg, Kwon, Scheinost, et al. (2022) general attention CPM predicts performance in a variety of attention tasks, including gradual continuous performance task, multiple object tracking, and visual change detection. It further generalized to predict attention performance across four independent datasets (N = 495) not used for model training. Because Yoo et al.'s CPM was based on volumetric rather than grayordinate data, for consistency, we made our comparison here using only the resting‐state volumetric fMRI data.

For the cross‐prediction analysis, we applied the binary functional connectome mask identified in Yoo et al.'s study and computed Pearson correlation between the summed connectivity strength within this mask and each EF behavioral measure. This allowed us to assess how well the attention‐based CPM could predict EF performance.

3. Results

3.1. The Raw Executive Function Measures Are Correlated

To investigate the EF components of response inhibition, set switching, and working memory updates, we used behavioral measures from the Flanker task, the Card Sorting task, and the 2‐back task, respectively.

Descriptive statistics revealed that every pair of measurements was correlated with each other (with p < 0.001), as indicated in Table 1. The highest correlation was found between the Card Sort and Flanker measures. More statistics for each measurement can be found in Table S1.

TABLE 1.

Pearson correlation between behavior measures.

	Flanker	Card sort	2‐back
Flanker	—	—	—
Card sort	0.4915***	—	—
2‐back	0.2955***	0.3253***	—

^*** p < 0.001, uncorrected.

3.2. CPMs Predict Individual Differences for 3 Executive Function Tasks

Table 2 shows the results of CPMs trained and tested on Flanker, Card Sort, and 2‐back behavioral measures using connectomes constructed based on 2‐back (left column) or resting‐state fMRI data (right column). Most predictions were statistically significant (permutation test p < 0.05, corrected for FWE). The three rows of tables correspond to CPMs based on positive, negative, or both types of edges. For instance, the top left entry (0.13) in the right sub‐table corresponds to the predictive performance of a positive CPM model when trained and tested both on Flanker measures using resting‐state functional connectivity.

TABLE 2.

CPM prediction accuracy of raw EF measures, grayordinate.

		Working memory (2‐back)			Rest
		Flanker	Card sort	2‐back	Flanker	Card Sort	2‐back
Positive	Flanker	0.21***	0.18*,a	0.29***	0.13†,b	0.11	0.16*,c
	Card sort	0.20***	0.25***	0.34***	0.10	0.15*,d	0.20***
	2‐back	0.23***	0.23***	0.42***	0.14†,e	0.17***	0.21***
Negative	Flanker	0.29***	0.21***	0.31***	0.14*,f	0.09	0.16*,g
	Card sort	0.24***	0.25***	0.36***	0.08	0.14*,h	0.18*,i
	2‐back	0.27***	0.27***	0.44***	0.13	0.20***	0.23***
Both	Flanker	0.29***	0.22***	0.33***	0.15*,j	0.11	0.18***
	Card sort	0.25***	0.27***	0.38***	0.10	0.15*,k	0.20***
	2‐back	0.28***	0.28***	0.48***	0.15*,l	0.19***	0.23***

Note: Each sub‐table (bounded by bold lines) represents a set of CPM prediction performance scores for a specific fMRI task state (2‐back working memory vs. rest) using one type of edges (positive, negative, or both). For example, the top‐left sub‐table shows the CPM performance using working memory task fMRI data and only positive‐associated edges. In each 3 × 3 sub‐table, each row represents the training behavior, and each column represents the testing behavior. The maximums of each row and column in the sub‐tables are colored in blue and yellow, respectively, with the overlap colored in green. ***p < 0.001, **p < 0.01; *p < 0.05; ^† p < 0.1, all corrected for family‐wise error (FWE) after permutation testing. Exact p‐values: ^a: p = 0.036; ^b: p = 0.084; ^c: p = 0.012; ^d: p = 0.030; ^e: p = 0.060; ^f: p = 0.042; ^g: p = 0.012; ^h: p = 0.024; ⁱ: p = 0.012; ^j: p = 0.036; ^k: p = 0.012; ^l: p = 0.036. All other non‐significant p‐values > 0.120.

Within‐task predictions, as represented by the diagonal of each table, were significant in 17 out of 18 cases (permutation test p < 0.05, corrected for FWE). Notably, 2‐back performance was predicted significantly more accurately than Flanker and Card Sort scores (p < 0.05, corrected for FWE), regardless of brain state. More interestingly, cross‐task prediction performance (i.e., train the CPM on one behavioral measure and test on another) was also significant in most cases (permutation test p < 0.05, corrected for FWE). 2‐back performance was most strongly cross‐predicted when trained on any of the EF task measures (p < 0.001 in all cases, corrected for FWE), and the model trained on 2‐back behavior scores also showed high cross‐prediction accuracy on both Card Sort and Flanker scores. On the other hand, cross‐prediction performance for Card Sort and Flanker (both ways) was numerically lower and even non‐significant when using resting‐state fMRI data.

Comparing the left and right columns shows that CPM performance was significantly higher from 2‐back task scans than from resting‐state scans (p's < 0.001 in all cases, corrected for FWE). In other words, the 2‐back task fMRI functional connectivity data enabled higher prediction $r$ (permutation test p < 0.05, corrected for FWE) across all model types and behavioral combinations.

We also examined the impact of fMRI data format on CPM performance by comparing models trained and tested on grayordinate data versus volumetric data. Grayordinate‐based data showed superior CPM predictive performance over traditional volumetric‐based data in most cases (p < 0.001 in all cases using 2‐back fMRI data; p < 0.001 in 25 out of 27 cases using resting‐state fMRI data, both corrected for FWE). Detailed CPM performance on volumetric fMRI data is presented in Tables S4 and S5. In an additional analysis, we showed that the performance improvement is not solely due to the choice of parcellation atlas, as the gain persisted when applying the Schaefer 300 atlas to volumetric data (Table S13).

Therefore, due to the greater variance captured by using grayordinate‐based task fMRI data, we will primarily present the results based on 2‐back grayordinate fMRI data in the following sections. The patterns of results from resting‐state or volumetric data were similar, as shown in Tables S4 and S5.

3.3. The FPN, DMN, and DAN Engage Across all Executive Function Tasks

We next investigated the connectome anatomy for each of the EF task measures. The resulting connectome profile for each EF task is depicted in Figure 1.

FIGURE 1.

Predictive edges identified by CPM on three EF measures. The top row displays the edges identified by CPMs trained on each of the three types of EF measures. Each heatmap is divided into left and right halves, where the right half represents negative edges, and the left half represents positive edges. Each cell within the heatmap indicates the percentage of reliable edges (see the Methods section for detailed inclusion criteria) picked up by CPM between the corresponding pair of canonical networks. Higher percentages are represented by more intense pink or blue hues. The bottom row illustrates the union (sum) and intersection (minimum) of the three heatmaps above. FP, frontoparietal; DM, default mode; DA, dorsal attention; SM, somatomotor; VI, visual; SA, salience; SC, subcortical; BS/CB, brain stem/cerebellum.

In summary, positive edges for the Flanker performance were mainly within the frontoparietal network and with the dorsal attention network and cerebellum. Positive edges for the Card Sort performance involved mainly the frontoparietal, default mode, and dorsal attention networks. Lastly, positive edges for 2‐back performance were mostly within the frontoparietal network, as well as between the frontoparietal network and both the default mode and dorsal attention networks. Overall, the edges that positively predicted the three EF measures span a wide range of canonical networks and are relatively distinct from one another. The core cluster of positive edges common to all three EF CPMs, illustrated in Figure 1 (bottom right), was located within the frontoparietal and between the frontoparietal, dorsal attention, and default mode networks.

On the other hand, CPM also picked up a set of negative edges where their strengths predicted lower performance in each EF task. For Flanker, negative edges were found within and between the frontoparietal, default mode, and dorsal attention networks, as well as some edges within the visual network. The CPM for Card Sort included negative edges within the default mode, visual, and salience networks, as well as between the salience network and the dorsal attention and frontoparietal networks. The negative CPM model for 2‐back performance included mostly edges associated with the frontoparietal network, along with those between the visual network and brainstem/cerebellum. The intersection plot (Figure 1, bottom right) shows that most of the negative edges common to all three CPMs were within and between frontoparietal and default mode networks, as well as between salience and dorsal attention networks.

3.4. Lesioning the FPN and DMN Led to Largest Performance Drop

To further study the contribution of each network in predicting individual differences, we performed computational lesion analysis on each CPM by removing the connectivity of each network one at a time. As shown in Figure 2, in the positive models, lesioning frontoparietal connectivity resulted in the largest prediction performance drop in most cases, where lesioning the default mode network showed the largest prediction performance drop in the other cases (FPN Cohen's d: −2.665, DMN Cohen's d: −2.019). In the negative models, most impairments were again driven by lesioning the default mode network and the frontoparietal network (DMN Cohen's d: −3.685, FPN Cohen's d: −2.709). To a lesser extent, lesioning positive or negative edges from the salience, visual, and dorsal attention networks also resulted in significant drops in prediction performance. The complete table of Cohen's d values for each network is provided in Table S10a (For results based on volumetric data and resting‐state scans, see Table S10b–d).

FIGURE 2.

CPM lesion analysis on grayordinate, 2‐back task‐fMRI data. The figure is divided into two main sections: The left half displays CPM performance after lesioning positive edges of each network, while the right half shows the results after lesioning negative edges. Each section is further divided into subplots based on different train/test task combinations. In these subplots, each row corresponds to a training task, and each column represents a testing task. For example, the top‐left subplot in the left half illustrates CPM prediction when trained and tested both on the Flanker scores after lesioning positive edges of each canonical network. The horizontal gray line attached to each bar denotes the 95% (±2 standard deviation) confidence interval. The vertical gray dashed line represents the baseline CPM performance (without lesioning) for each scenario. To assess the impact of lesioning, paired t‐tests were conducted comparing the lesioned performance to the regular performance. The results are depicted with colored bars: Blue bars indicate significant reductions in performance (p < 0.05, after FWE correction), orange bars represent significant increases, and white bars denote no significant difference from regular performance. FP, frontoparietal; DM, default mode; DA, dorsal attention; SM, somatomotor VI, visual; SA, salience; SC, subcortical; BS/CB, brain stem/cerebellum.

3.5. General Executive Function Is Predicted Better Than Specific Executive Function Components

To study the shared connectome underlying general EF, as well as the idiosyncratic connectomes for specific EF components, we repeated the CPM procedure on four newly devised measures: one general EF score and three component‐specific scores. The general EF score was derived from the mean of the three original z‐scored EF measures, and the component‐specific scores were defined as the residuals after regressing out the other two measures from each target measure. More statistics for each measurement can be found in Tables S2 and S3.

As illustrated in Table 3, when tested on the same measure as the CPM was trained on, the general EF measure was predicted better than the component‐specific measures using 2‐back fMRI data (p < 0.001 in all 9 cases, corrected for FWE). On the other hand, the cross‐task prediction performance between different component‐specific measures was dampened compared to that based on the original EF measures (p < 0.001 in all 18 cases, corrected for FWE).

TABLE 3.

CPM prediction accuracy of general and specific EF measures, grayordinate.

		General EF	Flanker specific	Card sort specific	2‐back specific
Positive	General EF	0.38***	0.10	0.08	0.29***
	Flanker specific	0.17†,a	0.15*,b	−0.05	0.10
	Card sort specific	0.14	−0.02	0.10	0.09
	2‐back specific	0.32***	0.02	0.03	0.37***
Negative	General EF	0.44***	0.14	0.07	0.32***
	Flanker specific	0.14	0.11	−0.01	0.07
	Card sort specific	0.14	−0.04	0.12	0.08
	2‐back specific	0.34***	0.06	0.04	0.33***
Both	General EF	0.44***	0.14	0.08	0.32***
	Flanker specific	0.18*,c	0.15*,d	−0.04	0.10
	Card sort specific	0.16†,e	−0.04	0.13*,f	0.10
	2‐back specific	0.38***	0.04	0.04	0.41***

Note: Each sub‐table (bounded by horizontal lines) corresponds to the CPM prediction performance scores using 2‐back task‐fMRI and one type of edges (positive, negative, or both). In every 4 × 4 sub‐table, each row represents the training behavior, and each column represents the testing behavior. The maximums of each row and column in the sub‐tables are colored in blue and yellow, respectively, with the overlap colored in green. ***p < 0.001, **p < 0.01; *p < 0.05; ^† p < 0.1, all corrected for family‐wise error (FWE) after permutation testing. Exact p‐values: ^a: p = 0.096; ^b: p = 0.040; ^c: p = 0.024; ^d: p = 0.040; ^e: p = 0.072; ^f: p = 0.040. All other non‐significant p‐values > 0.120.

Aligned with our earlier analysis, using 2‐back fMRI data enhanced predictive performance over the resting‐state data. Therefore, the following sections will emphasize results obtained from task‐based fMRI data over resting‐state fMRI data. More detailed results for the other conditions are provided in the Tables S6–S8 and Figures S7–S12.

3.6. The Component‐Specific Measures Bear Little Network Connectivity Overlap

Examining the canonical functional networks for each component‐specific positive CPM model reveals little overlap (Figure 2, bottom right), suggesting that the connectome for each EF component became more unique when its shared variance with others was removed.

While the connectome for each component looks roughly similar to the previous ones, one can spot some differences relative to the models without shared variance removed (Figure 1). For positive networks, the Flanker‐specific component model now consists of fewer frontoparietal‐related edges, both in terms of within frontoparietal edges and between frontoparietal and other networks such as dorsal attention network and brainstem/cerebellum (all p < 0.001, FWE corrected). Within‐salience network connectivity, on the other hand, was more relevant to predict individual differences in the Flanker‐specific component (p < 0.001, FWE corrected). Similar patterns were observed for the Card‐Sort‐specific component model, where the frontoparietal network showed fewer connections within itself and with the dorsal attention network, while the salience network's connections with DMN, DAN, and SM all increased (p < 0.001, FWE corrected). Lastly, the 2‐back‐specific component model showed fewer positive edges between dorsal attention and frontoparietal networks, but more edges between within dorsal attention and between dorsal attention and default mode networks (p < 0.001, FWE corrected).

Negative CPM models in Figure 3 also showed a decrease in overlapping edges relative to the models without shared variance removed (Figure 1). Again, subtle changes can be observed when closely examining each component CPM. For the Flanker‐specific component model, the reduction in within‐frontoparietal edges was accompanied by an increase in the interconnection between the salience and default mode networks (p < 0.001, FWE corrected). The Card‐Sort‐specific component model showed more within‐frontoparietal and between visual‐somatomotor edges (p < 0.001, FWE corrected). Lastly, the 2‐back‐specific component model showed an increase in visual‐frontoparietal edges (p < 0.001, FWE corrected). The exact statistical comparisons of network profile differences between the CPM models based on the original and derived EF measures are provided in Figure S13.

FIGURE 3.

Predictive edges identified by CPM on the three component‐specific measures. The top row displays the edges identified by CPMs trained on each of the three component‐specific measures. The bottom row illustrates the union (sum) and intersection (minimum) of the three heatmaps above. FP, frontoparietal; DM, default mode; DA, dorsal attention; SM, somatomotor; VI, visual; SA, salience; SC, subcortical; BS/CB, brain stem/cerebellum.

3.7. General Executive Function Involves the Interplay of the FPN, DMN, and DAN

Lastly, we examined the connectome for the general EF measure, defined as the shared variance across the Flanker, Card Sort, and 2‐back task measures. As shown in Figure 4, the positive edges were densely located around the frontoparietal networks, including within‐network connections and their interconnections with the dorsal attention and default mode networks. Other edges, though less densely populated, mostly bridged the default mode network with other networks. In contrast, the negative edges were most numerous within the frontoparietal network, between the frontoparietal and default mode networks, and between the brainstem/cerebellum and visual networks. Additionally, the dorsal attention and visual networks comprised a substantial number of inter‐network connections that correlate negatively to individual EF performance. Notably, we found that the connectome profile of the general EF component (Figure 4) closely resembles the intersection connectome derived from the three original EF measures (Figure 1, bottom right). This similarity suggests that, across two different analytical approaches—one using the original EF measures and the other using the derived general EF component—a consistent pattern of connectivity emerges, highlighting a robust neural signature of shared executive function.

FIGURE 4.

Predictive edges identified by CPM on the general EF measure. FP, frontoparietal; DM, default mode; DA, dorsal attention; SM, somatomotor; VI, visual; SA, salience; SC, subcortical; BS/CB, brain stem/cerebellum.

3.8. General EF and General Attention Differ in Their Connectome Profile

The discovery of a stable connectome underlying general executive function (EF) led us to ask whether it might share predictive power and a common neural basis with another core cognitive construct—attention. Prior work by Yoo, Rosenberg, Kwon, et al. (2022) discovered a robust connectome profile predictive of general attention in a variety of attention tasks. Hence, we ask whether the connectome‐based predictive model (CPM) developed for general attention could also generalize to predict EF‐related behaviors.

As shown in Table S9a,b, the general attention CPM did not generalize well to predict any of the individual EF behavioral measures, nor the composite general EF measure. In contrast, the CPM trained on general EF successfully predicted both the original EF tasks and the general EF construct significantly.

4. Discussion

In this study, we identified both shared and specific executive function components and their brain connectomes. Utilizing Connectome‐based Predictive Modeling (CPM), we revealed the distributed functional connectivity patterns that robustly predict individual differences and neural connectome profiles underlying the Flanker task, the Dimensional Change Card Sort task, and the 2‐back task performance. These tasks were chosen from the HCP dataset to reflect the EF components of Inhibition, Shifting, and Updating (Friedman and Miyake 2017), although we do not assume that each task purely measures its corresponding EF component. Accordingly, we identified both shared and specific EF connectomes through the use of cross‐task prediction analysis and derived EF measures.

4.1. Cross‐Prediction Performance Suggests That the 2‐Back Task and Updating Are More Central to General Executive Function at the Connectome Level

CPM cross‐prediction patterns on the three original task measures revealed that the model trained on 2‐back behavioral scores generalized the best to predict performance in the other two EF tasks. The high cross‐prediction performance suggests that the connectome basis of the 2‐back behavior also tracks individual performance differences in Flanker and Card Sort tasks. On the other hand, the Flanker and Card Sort tasks appear to be more independent from the connectome standpoint, evident by their lower cross‐prediction accuracy. It is worth noting that the higher cross‐prediction accuracy of the 2‐back model is not merely due to using the same fMRI and behavioral task, since the cross‐prediction gain of CPMs based on 2‐back scores persists with resting‐state fMRI data.

Our finding that Updating is central to EF from a neural connectome perspective aligns with its crucial role in executive function—a notion supported by various prior studies. In Lemire‐Rodger et al. (2019), multivariate analysis on fMRI data alluded to Updating as a common factor supporting the other EF processes. In another meta‐analysis, Rodríguez‐Nieto et al. (2022) reported that the Updating network highly overlaps with the Shifting and Inhibition networks, while the latter two exhibited minimal overlap. At the behavioral level, Updating stood out as the sole process among the three EF mechanisms showing a substantial correlation with general intelligence (Friedman et al. 2006), suggesting it is domain‐general. Updating, commonly operationalized as working memory, involves controlling attention to resist interference, a key component that scaffolds many higher‐order executive processes (Engle 2002; Engle et al. 1999; Kane and Engle 2003; Unsworth et al. 2004).

Note that our finding is less consistent with some previous studies that posited Shifting (Dajani and Uddin 2015) and Inhibition (Miyake and Friedman 2012) as more central components for general EF. Also, our CPM prediction results appear at odds with the behavioral correlations among the three EF measures, where Flanker and DCCS show the strongest association. We speculate that this discrepancy may stem from differences in the level of analysis—behavioral versus neural connectome—as well as the limited variance in behavior that functional connectivity can capture. We encourage future research to directly examine EF components across different levels of measurement to better understand these divergences.

On the other hand, the Flanker task CPM (Inhibition) and the Card Sort task CPM (Shifting) exhibited lower generalizability when applied to predict other tasks, suggesting they reflect distinct components of EF. This proposition aligns with several previous studies (Lemire‐Rodger et al. 2019; Miyake et al. 2000; Rodríguez‐Nieto et al. 2022). Upon closer examination of our results using 2‐back fMRI data, we found that models trained on the Card Sort task and tested on the Flanker task performed better than those trained and tested in the opposite direction. This observation may be related to the idea that task switching is facilitated by inhibition, as smoothly transitioning to a new task set requires effectively suppressing the previous set of rules (Davidson et al. 2006; Diamond 2013; Koch et al. 2010).

Note here, we followed the standard CPM training and testing procedure (Shen et al. 2017), in which feature selection is performed on the training set, with the testing set held out. Although it is possible to use separate datasets for feature selection and model training, this approach can lead to a numerical reduction in model performance (Table S12). While one might attribute this reduction to overfitting in the standard approach, we believe it is more likely driven by the smaller sample size used for feature selection, which increases the likelihood of including less robust features for prediction (Figure S16). The strong dependence of brain‐wide association studies on sample size has been noted in several recent studies (Cui et al. 2020; Ooi et al. 2025; Schulz et al. 2024), and we encourage interested readers to explore this issue further, particularly in the context of CPM analysis.

4.2. Derived Measures Reveal General and Specific Executive Function Connectomes

The high within‐task prediction accuracy of our CPM trained on the general EF measure, derived from the mean of the three tasks, indicates the presence of a general EF factor. The significant within‐task prediction accuracy of the 2‐back‐specific component model, along with its high cross‐predictive performance on the general EF factor, suggests a stronger role for the EF component underlying the 2‐back task, putatively Updating. Conversely, the cross‐prediction accuracy between Flanker‐specific and Card Sort‐specific component measures was numerically lower, suggesting that they became more distinctive after removing the shared variance with each other. Overall, the identification of a general EF factor aligns with Miyake's theory (Friedman and Miyake 2017; Miyake et al. 2000) that posited a construct that unifies different types of executive processes.

Examining the functional networks supporting general EF, we observed that many positive edges (i.e., edges that positively correlate with performance) reside in and between the frontoparietal (FPN), dorsal attention (DAN), default mode (DMN), and salience networks (SN). All of these canonical functional networks have previously been implicated in various aspects of EF (Friedman and Robbins 2022; Menon and D'Esposito 2022). The FPN is associated with the initialization and adjustment of control, executive task performance, and interactions between attention and other cognitive processes (Dosenbach et al. 2008; Marek and Dosenbach 2018; Seeley et al. 2007). The DAN directs top‐down attention and assists successful spatial attention (Corbetta and Shulman 2002, 2011; He et al. 2007). The SN plays a role in perceiving event saliency, monitoring conflicts, and initiating access to working memory and attention (Carter and van Veen 2007; Menon and Uddin 2010). Finally, although the DMN tends to quench its activity during tasks, it has a putative role in switching between internal and external attention modes (Leech et al. 2011). Additionally, we also found a number of positive edges in the visual network (VN). Although this may not be directly related to EF, the strength of connectivity within the VN is consistent with the fact that all three tasks involved visual perception (Baldassarre et al. 2012).

While each of these networks serves their distinct roles in EF, they communicate with each other to subserve more complicated EF processes, as reported by various prior studies. For instance, researchers have found that the FPN shows differential connectivity patterns with the DMN and the DAN. The former strengthens during cue‐independent introspective tasks, while the latter relates more to perceptual attention (Dixon et al. 2018). Another study revealed that the connectivity between the DMN and other task‐related networks (e.g., the SN, FPN) was strengthened during a battery of tasks and is correlated with task performances (Elton and Gao 2015). Aligned with these results, our general EF CPM picked up these connectivity features and thus further consolidates the notion that these functional networks may act as hubs for general EF.

In addition to the edges that have established roles in prior studies, CPM also detected numerous edges that are less reported previously. For instance, our results revealed an inverse relationship between general EF performance and the cerebellum/brainstem's co‐activation with the VN. However, we did not find reports in the literature about this negative correlation with cognitive performance. Therefore, confirming or disproving this relationship warrants deliberate investigation in the future. Overall, we believe that CPM can play a valuable role in expanding our current knowledge base by generating new hypotheses for future testing.

On the other hand, the connectome profiles for each component‐specific measure were more distinct from one another. For example, in the positive models, the Flanker task was characterized by a high density of edges within the FPN, DMN, and SN. In contrast, the Card Sort task showed more edges connecting the DMN with the SN and DAN. The 2‐back task's edges were primarily concentrated within and between the FPN, DMN, and DAN. These network motifs might hint at unique signatures for each specific EF component.

4.3. Positively and Negatively Predictive Edges Show High Similarity and Subtle Distinctions

Using 2‐back grayordinate fMRI data, we found that models based solely on positive or negative edges can both achieve significant within‐task and cross‐task prediction accuracy (Table 2 left), which implies that both types of edges are informative for individual differences in EF. Including the two types of edges together into a single CPM improves accuracy (p < 0.001 in all cases, FWE corrected), but the gain is not numerically large, indicating that the positive and negative edges capture overlapping information.

The overlap of variance explained by positive and negative edges is also reflected in the canonical network profiles. As shown in Figure 4, some connectivity features that positively or negatively correlated with general EF performance were from the same set of canonical networks. For instance, a high proportion of edges within the FPN, between the FPN and DMN, and between the FPN and DAN appeared in both the positive and negative models. One possible interpretation is that the canonical networks such as FPN encompass a large number of brain regions that may be heterogeneous in their contributions to EF (Dixon et al. 2018).

Despite the overall similarity in their network profiles, positive and negative edges show subtle differences in their regional density. Consistent with past literature (Anticevic et al. 2012; Ellwood‐Lowe et al. 2021; Wen et al. 2018), we observed significantly more positive edges between the DAN and FPN (p < 0.001), whereas a greater number of negative edges were found between the DMN and FPN (p < 0.001). This dissociation aligns with the idea that the DAN supports externally directed attention, while DMN activity reflects internally focused processes (Buckner et al. 2008; Fox et al. 2005). As such, greater engagement of external attention may facilitate better executive functioning, whereas reliance on internally directed attention—potentially indicative of mind‐wandering—may impair performance. The FPN, in this case, may serve as a regulatory hub, arbitrating the balance between task‐relevant, externally focused attention and task‐unrelated, internally focused mental activity (Spreng et al. 2013).

4.4. Grayordinate Data Representation Enhances Prediction Accuracy

Throughout our analysis, we also studied the impact of fMRI data representation and brain state on CPM performance. We found that using grayordinate (CIFTI) data provides a significant boost in CPM predictive accuracy over traditionally used volumetric (NIFTI) data. This advantage can be attributed to the inherent benefit of using grayordinate data, which registers the cortical areas into a flat surface while maintaining 3D structures of subcortical areas. By doing so, it provides a more compact representation with higher inter‐subject spatial correspondence (Glasser et al. 2013), better signal‐to‐noise ratio (Smith et al. 2013), and reduced signal contamination that inflates functional connectivity (Brodoehl et al. 2020).

It is important to note that this improvement is not solely driven by differences in the parcellation atlas used for the two fMRI representations. As shown in Table S13, even when holding the parcellation atlas constant (Schaefer 300), grayordinate data still yielded superior performance compared to volumetric data.

Admittedly, this is not an exhaustive comparison between the two data representations, but it suggests that a well‐chosen fMRI data format can enhance CPM analysis performance. At the same time, one should be mindful of potential artifacts introduced by surface‐based analysis (Jeganathan et al. 2025). We encourage future studies to systematically compare data formats to better determine their impact on predictive performance.

4.5. Task‐Based Connectome Supports Better Predictions Over Resting‐State

While past research has demonstrated the satisfactory performance of brain‐based prediction using resting‐state functional connectivity (e.g., Dubois et al. 2018; Finn et al. 2015; Kong et al. 2019), our findings suggest that task‐based connectivity features yielded improved predictive performance. Specifically, when comparing resting‐state fMRI data with 2‐back task‐based fMRI data, we noticed a significant improvement associated with the use of 2‐back data across both volumetric and gradyordinate data representation formats. This enhancement is consistent with previous findings that task‐based fMRI data generally aids in more accurate prediction of traits and behaviors, even if the task data differs from the behavior being predicted (Chen et al. 2022; Elliott et al. 2019; Finn and Bandettini 2021; Greene et al. 2018; Jiang et al. 2020; Yoo et al. 2018; Yoo, Rosenberg, Kwon, et al. 2022). In combination with other reported benefits of task fMRI data, such as its greater information gain for inferring hidden parameters (Tuominen et al. 2023), less head motion during acquisition (Huijbers et al. 2017), and better test–retest reliability for network identification (Kristo et al. 2014; Rosazza et al. 2014; Wang et al. 2017), when possible, we support the use of task‐based fMRI data over resting‐state data (or a hybrid use of resting and task, see Finn and Bandettini 2021 for further perspectives) to achieve better predictive accuracy.

4.6. Future Directions

While our study provides a fresh perspective on the connectome profiles of various EF processes, we also consider several limitations of our current approach as well as potential future directions. First off, the so‐called task impurity problem (Burgess 1997; Phillips 1997) poses a challenge to accurately measuring the psychological processes targeted by EF tasks. Because each EF task typically involves a combination of cognitive, perceptual, and motor processes, one cannot assert that the variance captured by the model reflects only the intended process. Given the inherent difficulty in creating a “pure” EF task, our project takes the valuable approach of distinguishing between the common and specific components that each EF score measures. Future work could build on these analyses by extending it to other cognitive and non‐cognitive measures, allowing us to evaluate the extent to which each task relies on general versus specific EF components, as well as the generalizability of our EF connectome.

Secondly, due to the constrained nature of public neuroimaging datasets, we were only able to build our CPMs on the 2‐back and resting‐state data. While the 2‐back data already demonstrates good predictive performance across all three behavior measures, it would be valuable to replicate these analyses using other types of EF‐related task fMRI data. A similar concern is that the connectome profile captured using 2‐back fMRI data differs from that obtained using resting‐state data. Interpreting this discrepancy is essential to developing a connectome profile of EF that is agnostic to brain state.

Third, it will be informative to examine the relation between general EF and other cognitive constructs. For example, EF is thought to involve attention and both heavily depend on the prefrontal cortex (Diamond 2013; Kane and Engle 2002). Accordingly, as an exploratory analysis, we compared the general EF connectome identified in this study with a general attention model previously developed by Yoo, Rosenberg, Kwon, Scheinost, et al. (2022). Our findings showed that the attention‐based model did not generalize well in predicting individual differences in EF performance, despite its strong generalizability to predict sustained attention, multiple object tracking, change detection, and the attention network task (e.g., Posner and Petersen 1990; Rosenberg et al. 2016). This dissociation suggests that, from a functional connectivity perspective, general attention and general EF may rely on distinct neural underpinnings. It is worth noting, however, that our comparison is preliminary. First, it is a null result. Second, to maintain consistency with Yoo et al.'s original analysis, we restricted our evaluation to CPMs trained on resting‐state volumetric fMRI data. Third, more rigorous research is needed to directly compare a broader set of in‐scanner tasks targeting both constructs. More broadly, we advocate for further investigation into how the general EF connectome interfaces with networks associated with other cognitive domains, such as memory and decision‐making.

Finally, while this study specifically examines the EF components in adults, it will be useful to explore the connectome basis of EF from a developmental perspective. Extensive studies have shown that EF and prefrontal cortex undergo prolonged developmental trajectories before stabilizing in adulthood (Anderson 2002; Davidson et al. 2006; Diamond 2002; Kolb et al. 2012; Luna 2009) and the relevant functional networks exhibit significant changes to facilitate enhanced EF performance (Keller et al. 2023; Nomi et al. 2017; Anderson 2002; Davidson et al. 2006; Diamond 2002; Kolb et al. 2012; Luna 2009), and the relevant functional networks exhibit significant changes to facilitate enhanced EF performance (Keller et al. 2023; Nomi et al. 2017). From the whole‐brain connectivity standpoint, there are indeed shared and unique connectome features that stably predict children's cognitive performance (Chen et al. 2022). Thus, it would be beneficial to generalize our CPM analysis pipeline to developmental cohorts, leveraging datasets like the ABCD Study to investigate the neurodevelopmental trajectories of each facet of EF. A thorough understanding of how the connectome of EF components evolves across different ages could illuminate the origins of EF and provide insights into the neural signatures of atypical EF development.

5. Conclusions

In this study, we performed connectome‐based predictive modeling analysis on large‐scale task fMRI data to reveal both shared and unique aspects of executive functions from a whole‐brain connectivity perspective. Our analyses suggest the centrality of the Updating component in executive function and highlight the roles of the frontoparietal, default‐mode, and dorsal attention networks in supporting general executive functioning. We further demonstrated the predictive advantages of using task‐based grayordinate fMRI data over resting‐state or volumetric data. Future work could build on this pipeline to generalize findings across additional brain states and behaviors that fall under the umbrella of executive control.

Supporting information

Table S1. Statistics of raw behavior measures.

Table S2. Pearson Correlation between the 4 new measures.

Table S3. Statistics of the new behavior measures.

Table S4. CPM prediction accuracy of raw EF measures, volumetric.

Table S5. Paired t‐test for grayordinate vs. volumetric, rest.

Table S6. CPM prediction accuracy of general and specific EF measures, grayordinate, rest.

Table S7. CPM prediction accuracy of general and specific EF measures, volumetric, 2‐back.

Table S8. CPM prediction accuracy of general and specific EF measures, volumetric, rest.

Table S9a. Use general attention CPM to predict the general and specific EF measures, volumetric, rest.

Table S10a. Cohen's d for network lesion analysis (2back, grayordinate).

Table S10b. Cohen's d for network lesion analysis (2back, volumetric fMRI).

Table S10c. Cohen's d for network lesion analysis (rest, grayordinate fMRI).

Table S10d. Cohen's d for network lesion analysis (rest, volumetric fMRI).

Table S11. CPM Predictive Performance on 2‐back measures with dierent thresholds.

Table S12. CPM prediction accuracy of raw EF measures, grayordinate, 2‐Back fMRI.

Table S13. CPM prediction accuracy using Schaefer 300 atlas on volumetric data versus.

Figure S1. CPM Canonical network analysis, using 2‐Back, volumetric fMRI.

Figure S2. CPM Canonical network analysis, using resting, grayordinate.

Figure S3. CPM Canonical network analysis, using resting, volumetric fMRI.

Figure S4. CPM lesion analysis on 2‐Back, volumetric fMRI data.

Figure S5. CPM lesion analysis on resting, grayordinate fMRI data.

Figure S6. CPM lesion analysis on resting, volumetric fMRI data.

Figure S7. CPM canonical network analysis of component‐specific.

Figure S8. CPM canonical network analysis of general EF using resting.

Figure S9. CPM canonical network analysis of component‐specific.

Figure S10. CPM canonical network analysis of general EF using 2‐back.

Figure S11. CPM canonical network analysis of component‐specific.

Figure S12. CPM canonical network analysis of general EF using resting.

Figure S13. Differences in canonical network profiles between the original.

Figure S14. Visualization of the original EF behavioral data distributions.

Figure S15. CPM canonical network analyses for the original Flanker, DCCS.

Figure S16. CPM canonical network features showing the intersection of.

Qu, S. , Qu Y. L., Yoo K., and Chun M. M.. 2025. “Connectome‐Based Predictive Models of General and Specific Executive Functions.” Human Brain Mapping 46, no. 14: e70358. 10.1002/hbm.70358.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.