MultiPAD: a multimodal network biomarker for predicting Amyloid positivity in healthy elderly
Poster No:
1853
Submission Type:
Abstract Submission
Authors:
Alle Meije Wink1, Luigi Lorenzini1, Leonard Pieperhoff1, Giuseppe Pontillo2, James Cole3, Frederik Barkhof1
Institutions:
1Amsterdam University Medical Centre, Amsterdam, Noord-Holland, 2Amsterdam University Medical Centre, Amsterdam, Netherlands, 3University College London, London, London
First Author:
Co-Author(s):
Introduction:
Alzheimer's Disease (AD), the main cause of dementia worldwide,is characterised by a progressive loss of connectivity between brain regions.
Different aspects of brain connections are measured with various magnetic resonance (MR) imaging protocols. Diffusion-weighted imaging (DWI) measures the presence of white matter fibres as they guide the diffusion of water, functional MRI (fMRI) measures co-activity between regions as connections, and with anatomical MRI, corresponding morphological properties across the brain are seen as caused by a common factor.
Instead of analysing these network measurements in isolation, we propose a new brain network model that incorporates them, and connects the corresponding nodes between them, forming a multiplex network (see figure 1).
Different aspects of brain connections are measured with various magnetic resonance (MR) imaging protocols. Diffusion-weighted imaging (DWI) measures the presence of white matter fibres as they guide the diffusion of water, functional MRI (fMRI) measures co-activity between regions as connections, and with anatomical MRI, corresponding morphological properties across the brain are seen as caused by a common factor.
Instead of analysing these network measurements in isolation, we propose a new brain network model that incorporates them, and connects the corresponding nodes between them, forming a multiplex network (see figure 1).
Methods:
Participants from the European consortium for Prevention of Alzheimer's Dementia (EPAD, https://ep-ad.org) were healthy elderly with no cognitive decline (CDR < 1), some of whom were tested amyloid-positive based on cerebrospinal fluid (CSF) testing. Selected participants underwent anatomical MR imaging, fMRI and DWI. Connectomes (pairwise connections between brain regions) were connected with MIND for anatomical MRI, fmriprep for fMRI and qsiprep for DWI, using the Schaefer brain atlas with 100 regions.
Connectomes were stored as matrices, stronger connections being represented by higher values, and missing connections by the value zero. Connections between matrices were only possible within the same region. Mathematically, this corresponds to a block-diagonal matrix of the 3 connectomes, together with diagonals in the rest of the blocks (see figure 2). The values on these diagonals represent the 'connectivity', or the 'cost', of building paths in the network that use more than one connectome (the green connections in figure 1).
Keeping the measured connectomes fixed, we computed the optimal weights for predicting the CSF-derived amyloid status as those maximising the difference between the mean group whole-brain centrality of the 10 amyloid-positive subjects with the lowest (most extreme) CSF amyloid levels and the 10 amyloid-negative subjects with the highest CSF amyloid levels, respectively. This optimisation framework was implemented in Python with the lmfit package for nonlinear fitting.
Connectomes were stored as matrices, stronger connections being represented by higher values, and missing connections by the value zero. Connections between matrices were only possible within the same region. Mathematically, this corresponds to a block-diagonal matrix of the 3 connectomes, together with diagonals in the rest of the blocks (see figure 2). The values on these diagonals represent the 'connectivity', or the 'cost', of building paths in the network that use more than one connectome (the green connections in figure 1).
Keeping the measured connectomes fixed, we computed the optimal weights for predicting the CSF-derived amyloid status as those maximising the difference between the mean group whole-brain centrality of the 10 amyloid-positive subjects with the lowest (most extreme) CSF amyloid levels and the 10 amyloid-negative subjects with the highest CSF amyloid levels, respectively. This optimisation framework was implemented in Python with the lmfit package for nonlinear fitting.
·Figure 1: Left: a complete brain network (C) is made by joining connectomes (A and B) within regions. Right: a supra-adjacency matrix made from 3 connectomes.
Results:
Different patterns could be seen in the three sets of between-layer connections, when they were optimised to get the highest group-average whole-brain centrality (see Figure 2).
The amyloid-negative (A-) and amyloid-positive (A+) groups showed similar patterns of connection weights between the structural and functional connectome layers, with higher weights in the somatomotor cortex. A marked difference between the groups was an area of high frontal between-layer weights in the A- group, which was absent in the A+ group.
In the connection channel between the functional and anatomical layers, higher weights were only found in the A- group in the prefrontal, lower visual and occipital lobes. Connections between these layers of similar strength were not found in the A+ group.
The channel between the structural and anatomical layers showed higher weights in the left frontal cortex in the A- group. In the A+ group, no pattern of relatively higher weights were found in specific regions, although overall weights were higher than in the A- group.
The amyloid-negative (A-) and amyloid-positive (A+) groups showed similar patterns of connection weights between the structural and functional connectome layers, with higher weights in the somatomotor cortex. A marked difference between the groups was an area of high frontal between-layer weights in the A- group, which was absent in the A+ group.
In the connection channel between the functional and anatomical layers, higher weights were only found in the A- group in the prefrontal, lower visual and occipital lobes. Connections between these layers of similar strength were not found in the A+ group.
The channel between the structural and anatomical layers showed higher weights in the left frontal cortex in the A- group. In the A+ group, no pattern of relatively higher weights were found in specific regions, although overall weights were higher than in the A- group.
·Figure 2: The inter-layer weights in both groups (amyloid negative and positive); top: structral to functional, middle: functional to anatomical, bottom: structural to anatomical.
Conclusions:
We have developed a multi-layer brain network model that incorporates connectomes measured with different MR imaging modalities. The interlayer weights between the connectomes of EPAD participants were optimised to maximise their group average whole-brain centrality in the multiplex networks.
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s)
Lifespan Development:
Aging
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 2
Methods Development 1
Multivariate Approaches
Keywords:
Aging
Computing
Data analysis
Design and Analysis
FUNCTIONAL MRI
Modeling
MRI
STRUCTURAL MRI
WHITE MATTER IMAGING - DTI, HARDI, DSI, ETC
1|2Indicates the priority used for review
Provide references using author date format
Esteban, O. et al. (2018), "fMRIPrep: a robust preprocessing pipeline for functional MRI", Nature Methods 16: 111–16
Newville, M. et al. (2023), "LMFIT: Non-Linear Least-Square Minimization and Curve-Fitting for Python", https://doi.org/10.5281/zenodo.8145703
Schaefer, A. et al (2018), "Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI", Cerebral Cortex 28: 3095-114
Sebenius, I. et al. (2023), "Robust estimation of cortical similarity networks from brain MRI", Nature Neuroscience 26: 1461–71