The magnitude of prediction error for behavior relates to sociodemographic and scan factors
Poster No:
1379
Submission Type:
Abstract Submission
Authors:
Jingwei Li1,2, Jianxiao Wu1,2, Simon Eickhoff2,1, B. T. Thomas Yeo3,4, Sarah Genon1,2
Institutions:
1Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich, Jülich, Germany, 2Institute of Systems Neuroscience, Heinrich Heine University Düsseldorf, Düsseldorf, Germany, 3CSC, TMR, ECE, WisDM, ISEP, National University of Singapore, Singapore, Singapore, 4Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Charlestown, MA
First Author:
Jingwei Li
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute of Systems Neuroscience, Heinrich Heine University Düsseldorf
Jülich, Germany|Düsseldorf, Germany
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute of Systems Neuroscience, Heinrich Heine University Düsseldorf
Jülich, Germany|Düsseldorf, Germany
Co-Author(s):
Jianxiao Wu
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute of Systems Neuroscience, Heinrich Heine University Düsseldorf
Jülich, Germany|Düsseldorf, Germany
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute of Systems Neuroscience, Heinrich Heine University Düsseldorf
Jülich, Germany|Düsseldorf, Germany
Simon Eickhoff
Institute of Systems Neuroscience, Heinrich Heine University Düsseldorf|Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich
Düsseldorf, Germany|Jülich, Germany
Institute of Systems Neuroscience, Heinrich Heine University Düsseldorf|Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich
Düsseldorf, Germany|Jülich, Germany
B. T. Thomas Yeo
CSC, TMR, ECE, WisDM, ISEP, National University of Singapore|Martinos Center for Biomedical Imaging, Massachusetts General Hospital
Singapore, Singapore|Charlestown, MA
CSC, TMR, ECE, WisDM, ISEP, National University of Singapore|Martinos Center for Biomedical Imaging, Massachusetts General Hospital
Singapore, Singapore|Charlestown, MA
Sarah Genon
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute of Systems Neuroscience, Heinrich Heine University Düsseldorf
Jülich, Germany|Düsseldorf, Germany
Institute of Neuroscience and Medicine (INM-7: Brain and Behaviour), Research Centre Jülich|Institute of Systems Neuroscience, Heinrich Heine University Düsseldorf
Jülich, Germany|Düsseldorf, Germany
Introduction:
MRI-based prediction of behavioral phenotypes is a promising approach for individualized diagnosis and treatment of mental health. However, this field lacks generalizability: some (groups of) individuals are generally misclassified or show relatively large prediction errors[1-3]. Although it has been suggested that misclassified people tended to deviate from the stereotypical patterns observed in correctly classified participants, the range of sociodemographic and neuroimaging-specific factors relating to prediction errors is unknown. In this study, we investigated the associations between the prediction error magnitude of multiple behavioral domains and a broad range of variables in developmental and young adult cohorts.
Methods:
We focused on three scan-related and five sociodemographic variables of interest in three open datasets: the Adolescent Brain Cognitive Development cohort (ABCD[4]; N = 5351; 9-11y; 36 behavioral measures), the Human Connectome Project -Young Adults (HCP-YA[5]; N = 948; 22-37y; 51 behavioral measures), and the Human Connectome Project – Development (HCP-D[6]; N = 455; 8-22y; 22 behavioral measures).
The preprocessing of resting-state fMRI followed our previous work (ABCD & HCP-YA: [1]; HCP-D: [7]), resulting in functional connectivity (FC) matrices across 400 cortical areas[8] and 19[9] (for ABCD & HCP-YA) / 54[10] (for HCP-D) subcortical areas. Similarly, kernel ridge regression (for ABCD & HCP-YA) and support vector regression (for HCP-D) were used to predict the behavioral measures from FC. Age, gender, education (parental education for ABCD), intracranial volume, and head movement (and family income for HCP-YA) were regressed from behavioral scores and FC.
For simplicity, behavioral measures were clustered within each dataset based on their similarities in prediction errors (Fig. 1A). Prediction errors were averaged within each behavioral cluster. The associations between the averaged prediction errors and each continuous covariate were examined by Pearson's correlation. The associations of categorical covariates with prediction errors were quantified by a two-sample t-test (gender) and one-way ANOVA (other covariates).
To control the effect of dataset size, we randomly subsampled 100 times in ABCD and HCP-YA to match HCP-D.
The preprocessing of resting-state fMRI followed our previous work (ABCD & HCP-YA: [1]; HCP-D: [7]), resulting in functional connectivity (FC) matrices across 400 cortical areas[8] and 19[9] (for ABCD & HCP-YA) / 54[10] (for HCP-D) subcortical areas. Similarly, kernel ridge regression (for ABCD & HCP-YA) and support vector regression (for HCP-D) were used to predict the behavioral measures from FC. Age, gender, education (parental education for ABCD), intracranial volume, and head movement (and family income for HCP-YA) were regressed from behavioral scores and FC.
For simplicity, behavioral measures were clustered within each dataset based on their similarities in prediction errors (Fig. 1A). Prediction errors were averaged within each behavioral cluster. The associations between the averaged prediction errors and each continuous covariate were examined by Pearson's correlation. The associations of categorical covariates with prediction errors were quantified by a two-sample t-test (gender) and one-way ANOVA (other covariates).
To control the effect of dataset size, we randomly subsampled 100 times in ABCD and HCP-YA to match HCP-D.
Results:
In scan-related factors, head movement was widely correlated with multiple behavioral domains across datasets (Fig. 1B). In developing cohorts (ABCD, HCP-D), head size, captured by intracranial volume, was significantly negatively correlated with prediction errors of Child Behavior Checklist (CBCL), prodromal psychosis, and emotion recognition. In sociodemographic factors, consistent with our previous work[1], ethnicity was strongly associated with prediction errors for most of the behavioral domains in ABCD and HCP-YA. In addition, (parental) education and family income were even more broadly associated with prediction errors of almost all behavioral domains across all datasets. Age was correlated with prediction errors in both developing cohorts. The associations observed in the full samples still hold in subsamples. Examples of the observed associations are illustrated in Fig. 2.
Conclusions:
FC-based behavioral prediction errors were broadly associated with many scan-related and sociodemographic factors in young populations. More associations were observed for the behavioral domains that were more difficult to predict, such as CBCL and prodromal psychosis in ABCD and emotion recognition in the two HCP datasets. It suggests that the predictive models might learn more information from non-neurobiological factors to predict such behavioral measures. More associations were observed in ABCD compared to the other two datasets, possibly related to the higher diversity of this dataset. In sum, this study importantly contributes to a better understanding of the insufficient prediction power and low generalizability in the field.
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Task-Independent and Resting-State Analysis 2
Keywords:
Data analysis
FUNCTIONAL MRI
Machine Learning
Other - human behavior prediction, socioeconomic status, individual variation, generalizability
1|2Indicates the priority used for review
·Figure 1
·Figure 2
Provide references using author date format
[2] Dhamala, E. et al. (2023). One Size Does Not Fit All: Methodological Considerations for Brain-Based Predictive Modeling in Psychiatry. Biological Psychiatry. 93(8). 717-728.
[3] Greene, A.S. et al. (2022). Brain–phenotype models fail for individuals who defy sample stereotypes. Nature. 609, 109–118.
[4] Casey, B.J. et al. (2018). The Adolescent Brain Cognitive Development (ABCD) study: Imaging acquisition across 21 sites. Dev. Cogn. Neurosci. 32. 43-54.
[5] Van Essen, D.C. et al. (2013). The WU-Minn human connectome project: An overview. NeuroImage. 80, 62–79.
[6] Somerville, L.H. et al. (2018). The Lifespan Human Connectome Project in Development: A large-scale study of brain connectivity development in 5-21 year olds. Neuroimage. 183. 456-468.
[7] Wu, J. et al. (2021). A Connectivity-Based Psychometric Prediction Framework for Brain-Behavior Relationship Studies. Cereb. Cortex. 31(8). 3732-375
[8] Schaefer, A. et al. (2018). Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI. Cereb. Cortex 28, 3095–3114.
[9] Fischl, B. et al. (2002). Whole brain segmentation: Automated labeling of neuroanatomical structures in the human brain. Neuron 33, 341–355.
[10] Tian, Y. et al. (2020). Topographic organization of the human subcortex unveiled with functional connectivity gradients. Nat Neurosci. 23(11):1421-1432.