Continuous axes of individual variability in big data: enhanced fMRI biomarkers for traits & health
Poster No:
1753
Submission Type:
Abstract Submission
Authors:
Seyedeh-Rezvan Farahibozorg1, Mark Woolrich1, Stephen Smith1
Institutions:
1University of Oxford, Oxford, UK
First Author:
Co-Author(s):
Introduction:
Functional MRI from thousands of individuals, e.g., as made available by UK Biobank, allows to examine the brain function at population-scale [1]. This holds great promise for addressing fundamental questions in brain health: can differences in brain function be used to predict cognition and brain disorders? To leverage this potential, there is an increasing need for functional connectivity techniques that can extract accurate and clinically relevant brain features [2, 3]. Here we characterised a high-dimensional embedding space built on cross-subject similarities in the brain's resting state networks (RSNs) across 20,000 individuals in UK Biobank. This embedding space captures continuous axes of subject variability, where the most similar individuals will be placed close together along axes of variation. We demonstrate that this simple and flexible modelling of continuous variations across the population substantially improved cross-validated prediction accuracies for a wide range of phenotypes (>1000) in UK Biobank.
Methods:
RSNs were modelled using stochastic Probabilistic Functional Modes (PFMs) (Fig1a), which estimates RSNs in population and individuals simultaneously [4]. PFMs decompose subject fMRI data (D) into a set of functional modes, represented by spatial maps (P), time courses (A), amplitudes (H) and noise (E): D=PHA+E. Subject-specific models are regularised hierarchically using group-level priors [4,5].
P_concat (size N_subject x N_voxel x N_RSN), the subject-specific spatial configuration of RSNs across brain voxels, was used to estimate summary features in two ways:
1- Conventional Feature extraction (Fig1a): P_ concat was dimension-reduced to X1: N_subject x 500 feature space using Dictionary Learning and linked ICA, as described in our previous work [4].
2- Continuous Axes of Individual Variability (Fig1b): Cross-subject similarities were computed based on X1 and PCA was applied, while treating 'subjects' as variables, obtaining X2: N_ subject x 500.
X1 and X2 were used separately to predict phenotypes using ElasticNet regression and repeated 5-fold nested cross-validation. Imaging confounds were regressed out [6].
Phenotypes included: age (1), sex (1), region-wise cortical area (148 phenotypes) and thickness (148), White Matter (WM) tracts and microstructure (453), cognitive (68), alcohol (37), tobacco (19), health metrics related to blood/heart (77), bone (93) and mental health (142).
P_concat (size N_subject x N_voxel x N_RSN), the subject-specific spatial configuration of RSNs across brain voxels, was used to estimate summary features in two ways:
1- Conventional Feature extraction (Fig1a): P_ concat was dimension-reduced to X1: N_subject x 500 feature space using Dictionary Learning and linked ICA, as described in our previous work [4].
2- Continuous Axes of Individual Variability (Fig1b): Cross-subject similarities were computed based on X1 and PCA was applied, while treating 'subjects' as variables, obtaining X2: N_ subject x 500.
X1 and X2 were used separately to predict phenotypes using ElasticNet regression and repeated 5-fold nested cross-validation. Imaging confounds were regressed out [6].
Phenotypes included: age (1), sex (1), region-wise cortical area (148 phenotypes) and thickness (148), White Matter (WM) tracts and microstructure (453), cognitive (68), alcohol (37), tobacco (19), health metrics related to blood/heart (77), bone (93) and mental health (142).
·Figure1
Results:
First, we found that modelling continuous axes of individual variability substantially improved prediction accuracies for all the categories of phenotypes (Fig2a), with the following average accuracies (r): age: 0.61 -> 0.75; sex: 0.72 -> 0.79; cortical area: 0.37 -> 0.45; cortical thickness: 0.22 -> 0.28; WM: 0.30 -> 0.38; Cognition: 0.03 -> 0.05 ; Alcohol: 0.04 -> 0.07 ; Tobacco: 0.06 -> 0.08; Heart health: 0.12 -> 0.16; Bone Health: 0.28 -> 0.34; Mental Health: 0.02 -> 0.03.
Second, we computed the projection of the 500 axes of individual variability (i.e., columns of X2) onto the original voxel space of each RSN, yielding 500 X2_based spatial maps per RSN. We next computed phenotype-informed RSNs: 500 maps were weighted by their contribution in ElasticNet prediction of a specific phenotype (e.g., age) and averaged to obtain phenotype-informed maps, e.g., age-informed Default Mode Network in Fig2b. These can be used to examine age related population variations in each RSN.
Second, we computed the projection of the 500 axes of individual variability (i.e., columns of X2) onto the original voxel space of each RSN, yielding 500 X2_based spatial maps per RSN. We next computed phenotype-informed RSNs: 500 maps were weighted by their contribution in ElasticNet prediction of a specific phenotype (e.g., age) and averaged to obtain phenotype-informed maps, e.g., age-informed Default Mode Network in Fig2b. These can be used to examine age related population variations in each RSN.
·Figure2
Conclusions:
We characterised continuous axes of individual variability in brain function using a flexible data-driven approach for modelling subpopulation continua in large human populations. This new representation yielded clearly enhanced fMRI biomarkers for a wide range of traits related to cognition, health and substance use. Additionally, we used these axes of variability to compute phenotype-informed RSNs, which can be used as new population priors for RSN modelling in future studies.
Modeling and Analysis Methods:
Bayesian Modeling
Classification and Predictive Modeling 2
Connectivity (eg. functional, effective, structural)
fMRI Connectivity and Network Modeling 1
Task-Independent and Resting-State Analysis
Keywords:
FUNCTIONAL MRI
Machine Learning
Modeling
Other - functional connectivity, big data fMRI, subpopulation modelling, PROFUMO, phenotype prediction
1|2Indicates the priority used for review
Provide references using author date format
[2] T. He et al., “Meta-matching: a simple framework to translate phenotypic predictive models from big to small data,” bioRxiv, p. 2020.08.10.245373, 2020.
[3] Vidaurre D, llera A, S. M. Smith, and M. W. Woolrich, “Behavioural relevance of spontaneous, transient brain network interactions in fMRI,” Neuroimage, vol 229, p 117713, 2021.
[4] S.-R. Farahibozorg et al., “Hierarchical modelling of functional brain networks in population and individuals from big fMRI data,” Neuroimage, 2021.
[5] S. J. . Harrison et al., “Modelling subject variability in the spatial and temporal characteristics of functional modes,” Neuroimage, vol. 222, p. 117226, 2020.
[6] F. Alfaro-Almagro et al., “Confound modelling in UK Biobank brain imaging,” Neuroimage, vol. 224, p. 117002, 2020.