Connectome-based predictive models of cognitive control based on volumetric versus grayordinate data

919 

Abstract Submission 

Shijie Qu¹, Kwangsun Yoo², Marvin Chun¹

¹Yale University, New Haven, CT, ²Sungkyunkwan University, Seoul, Korea, Republic of

Shijie Qu  
Yale University
New Haven, CT

Kwangsun Yoo  
Sungkyunkwan University
Seoul, Korea, Republic of

Marvin Chun  
Yale University
New Haven, CT

Connectome-based predictive modeling (CPM, Finn et al., 2015; Shen et al., 2017) can predict individual cognitive differences based on variations in brain connectomes. Here, we develop resting-state connectome models to predict different aspects of cognitive control (Miyake et al., 2000). In doing so, we also compared models based on volumetric vs. grayordinate data, and we found grayordinate models to be more predictive of behavior and more generalizable across tasks.

For this study, we used neuroimaging data from the latest release of the WU-Minn Human Connectome Project (HCP) dataset (Van Essen et al., 2013), which contains data from approximately 1200 healthy adults, including neuroimaging and behavioral measurements. We focused on resting-state fMRI data and cognitive control-related behavioral measurements. The resting-state fMRI data comes from the first session, consisting of two 15-minute runs. The volumetric data were minimally preprocessed using the default pipeline, as illustrated in (Glasser et al., 2013). We further regressed out the motion parameters, white matter, cerebrospinal fluid signals, global signals, as well as linear trends. The grayordinate data, provided by HCP, preserves only gray matter signals, and represents cortical areas using 2-D vertices instead of 3-D voxels. The data were preprocessed using the pipelines described in Smith et al. (2013), which additionally use the ICA-FIX method to denoise the data. Since ICA-FIX already handled motion regression and linear detrending, here we only further regressed out white matter signals, cerebrospinal fluid signals, and global signals for grayordinate data. When generating the functional connectivity matrices, we used the Shen268 (Shen et al., 2013) atlas for volumetric data and the Cole-Anticevic atlas (Ji et al., 2019) for grayordinate data. For behavioral data, we included corresponding measurements for out-of-scanner list sorting, card sorting, and flanker tasks-three established tasks that respectively tap into three cognitive-control-related components: working memory update, set switching, and response inhibition (Miyake et al., 2000). Between each combination of data format and behavioral measurement, we trained a CPM model with 10-fold cross-validation and repeated the procedure 1000 times to obtain reliable estimates of CPM performance. Pearson correlation coefficients are calculated to represent model predictive performance.

Connectome-based predictive models based on volumetric data or grayordinate data significantly predicted individual differences in the tasks they were trained on (all p's < .05, all r's > .12). Grayordinate-data models significantly (p < .05) outperformed volumetric-data models for list sort and card sort within-task testing (tasks used for model training with 10-fold cross-validation) and for all cross-task testing (tasks not used for model training), even after controlling for the number of edges used for model testing. Lastly, anatomical analyses of the predictive edges revealed that the models trained on grayordinate data captured a denser group within and between frontal-parietal, cingulo-opercular, and default networks. Conversely, models trained on volumetric data covered a broader range of edges, including additional motor and visual networks.

Overall, we identified several differences in the application of grayordinate versus volumetric fMRI data in CPM analysis of cognitive control. Further analyses are needed to understand the reasons for these differences and their implications for dissociating components of cognitive control and their underlying neural networks.

Executive Function, Cognitive Control and Decision Making ¹

Connectivity (eg. functional, effective, structural)

fMRI Connectivity and Network Modeling ²

ADULTS

Cognition

Data analysis

FUNCTIONAL MRI

Machine Learning

NORMAL HUMAN