Print  Close

Predicting individual differences from task-evoked effective connectivity

Poster No:

1425 

Submission Type:

Abstract Submission 

Authors:

Shufei Zhang1,2, Kyesam Jung1,2, Robert Langner1,2, Esther Florin3, Simon Eickhoff1,2, Oleksandr Popovych1,2

Institutions:

1Institute of Neuroscience and Medicine, Brain and Behaviour (INM-7), Research Centre Jülich, Germany, Jülich, Nordrhein-Westfalen, 2Institute of Systems Neuroscience, Medical Faculty and University Hospital, Heinrich Heine University, Düsseldorf, Nordrhein-Westfalen, 3Institute of Clinical Neuroscience and Medical Psychology, Medical Faculty and University Hospital, Heinrich Heine University Düsseldorf, Germany

First Author:

Shufei Zhang  
Institute of Neuroscience and Medicine, Brain and Behaviour (INM-7), Research Centre Jülich, Germany|Institute of Systems Neuroscience, Medical Faculty and University Hospital
Jülich, Nordrhein-Westfalen|Heinrich Heine University, Düsseldorf, Nordrhein-Westfalen

Co-Author(s):

Kyesam Jung  
Institute of Neuroscience and Medicine, Brain and Behaviour (INM-7), Research Centre Jülich, Germany|Institute of Systems Neuroscience, Medical Faculty and University Hospital
Jülich, Nordrhein-Westfalen|Heinrich Heine University, Düsseldorf, Nordrhein-Westfalen
Robert Langner  
Institute of Neuroscience and Medicine, Brain and Behaviour (INM-7), Research Centre Jülich, Germany|Institute of Systems Neuroscience, Medical Faculty and University Hospital
Jülich, Nordrhein-Westfalen|Heinrich Heine University, Düsseldorf, Nordrhein-Westfalen
Esther Florin  
Institute of Clinical Neuroscience and Medical Psychology, Medical Faculty and University Hospital
Heinrich Heine University Düsseldorf, Germany
Simon Eickhoff  
Institute of Neuroscience and Medicine, Brain and Behaviour (INM-7), Research Centre Jülich, Germany|Institute of Systems Neuroscience, Medical Faculty and University Hospital
Jülich, Nordrhein-Westfalen|Heinrich Heine University, Düsseldorf, Nordrhein-Westfalen
Oleksandr Popovych  
Institute of Neuroscience and Medicine, Brain and Behaviour (INM-7), Research Centre Jülich, Germany|Institute of Systems Neuroscience, Medical Faculty and University Hospital
Jülich, Nordrhein-Westfalen|Heinrich Heine University, Düsseldorf, Nordrhein-Westfalen

Introduction:

In the research field of predicting individual behavior and phenotypes from functional magnetic resonance imaging (fMRI) data, recent studies [1, 2] have demonstrated that task-evoked functional connectivity (FC) may perform better at predicting individual traits as compared to resting-state FC. However, it remains largely unknown whether task-modulated and intrinsic effective connectivity (EC) that allows for causal inferences on the brain possess distinct properties in predicting individual behavior, which we explored in this study. We investigated this issue for a variety of data-processing conditions involving different designs of the general linear model (GLM), applications of Bayesian model reduction (BMR) and self-connectivity (SC) [3, 4] as well as for a few cross-validation (CV) schemes, whose impact on the EC-based prediction of behavioral scores has been underexplored in the literature yet. In this study we therefore considered task-evoked EC as calculated by dynamic causal modeling (DCM) [5] for the machine-learning prediction of individual reaction time (RT) in the stimulus-response compatibility (SRC) task [6] and age.

Methods:

Task-evoked EC during performance of the SRC task was calculated by DCM from the task fMRI data for 271 subjects (123 females, 18-85 years old, mean age: 52.6 ± 16.5 years) recruited from the subject pool of the 1000BRAINS project [7]. The parameters of intrinsic EC (I-EC, matrix A of DCM) and task-modulated EC (M-EC, matrix B of DCM) of individual subjects were used in a least absolute shrinkage and selection operator (LASSO) regression to predict RT and age in training sets and test sets of unseen subjects (Fig. 1). The features were selected based on the results of the group-level and behavior-related Parametric Empirical Bayes (PEB) analyses for the training sets. We considered several conditions that were assumed to influence predictive performance: cross-validation schemes of 5-fold, 10-fold, or leave-one-out CV (LOOCV), event-related vs. block-based GLM designs, application of BMR and self-connectivity (SC). In addition, we compared prediction results with those of the full task-evoked and task-residual FC patterns.
Supporting Image: workflow.jpg
   ·Fig. 1 The workflow for the prediction of individual behavioral characteristics based on task-evoked EC and PEB results. (a) The subjects and their corresponding behavioral characteristics (RT and age
 

Results:

We observed differences in predictive performance (correlation between empirical and predicted behavioral scores) between I-EC and M-EC, as well as between event-related (trial-based) and block-based GLM designs (Fig. 2). There were few differences in prediction accuracy across different CV schemes (but see LOOCV in Fig. 2), and with respect to the application of BMR and SC. For the 5-fold CV, we found statistically significant prediction results for the event-related GLM design, but not for the block-based GLM. For the event-related GLM, M-EC outperformed I-EC in RT prediction (r = 0.26 vs r = 0.09, Cohen's d = 3.17), but was somewhat less effective in age prediction (r = 0.22 vs r = 0.28, Cohen's d = 1.09). The considered task-evoked and task-residual FC patterns showed higher prediction accuracy for age than EC (r = 0.34/0.37 vs. 0.28) but were behind task-evoked EC in predicting RT (r = 0.13/0.22 vs. 0.26).
Supporting Image: updated_prediction_all.jpg
   ·Fig. 2 Overview of the mean prediction accuracy of all considered conditions in predicting individual reaction time (RT) and age under event-related (Event) and block-based (Block) GLM designs with fe
 

Conclusions:

Our results showed that task-modulated and intrinsic EC may capture different behavioral attributes, where M-EC showed higher predictive accuracy for individual RT than I-EC, but I-EC was better predictive of individual age than M-EC. The predictive performance was notably affected by the choice of GLM design for task-fMRI data modeling, and using the event-related GLM design may improve the predictive accuracy for individual RT and age. The selection of EC types for predicting individual differences and the choice of the optimal data processing during DCM estimation of EC should thus be made with care, where our results may guide further research on employing task-evoked EC for the prediction of individual behavior.

Learning and Memory:

Working Memory

Lifespan Development:

Aging

Modeling and Analysis Methods:

Classification and Predictive Modeling 1

Novel Imaging Acquisition Methods:

BOLD fMRI 2

Keywords:

Aging
Data analysis

1|2Indicates the priority used for review

Provide references using author date format

1. Greene, A.S., et al., Task-induced brain state manipulation improves prediction of individual traits. Nature Communications, 2018. 9(1): p. 2807.
2. Zhao, W., et al., Task fMRI paradigms may capture more behaviorally relevant information than resting-state functional connectivity. NeuroImage, 2023. 270: p. 119946.
3. Friston, K.J., et al., Bayesian model reduction and empirical Bayes for group (DCM) studies. NeuroImage, 2016. 128: p. 413-431.
4. Snyder, A.D., et al., Dynamic Causal Modeling Self-Connectivity Findings in the Functional Magnetic Resonance Imaging Neuropsychiatric Literature. Frontiers in Neuroscience, 2021. 15.
5. Friston, K.J., L. Harrison, and W. Penny, Dynamic causal modelling. NeuroImage, 2003. 19(4): p. 1273-1302.
6. Fitts, P.M. and R.L. Deininger, S-R compatibility: correspondence among paired elements within stimulus and response codes. J Exp Psychol, 1954. 48(6): p. 483-92.
7. Caspers, S., et al., Studying variability in human brain aging in a population-based German cohort—rationale and design of 1000BRAINS. Frontiers in Aging Neuroscience, 2014. 6: p. 149-162.