Multimodal brain age prediction using 5-HT2A receptor binding and gray matter volume
Poster No:
1133
Submission Type:
Abstract Submission
Authors:
Ruben Dörfel1,2, Joan Arenas-Gomez2, Claus Svarer2, Gitte Knudsen2,3, Jonas Svensson1,2, Pontus Plavén-Sigray1,2
Institutions:
1Centre for Psychiatry Research, Department of Clinical Neuroscience, Karolinska Institutet, Stockholm, Sweden, 2Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet, Copenhagen, Denmark, 3Department of Clinical Medicine, University of Copenhagen, Copenhagen, Denmark
First Author:
Ruben Dörfel, MSc
Centre for Psychiatry Research, Department of Clinical Neuroscience, Karolinska Institutet|Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Stockholm, Sweden|Copenhagen, Denmark
Centre for Psychiatry Research, Department of Clinical Neuroscience, Karolinska Institutet|Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Stockholm, Sweden|Copenhagen, Denmark
Co-Author(s):
Joan Arenas-Gomez, MSc
Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Copenhagen, Denmark
Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Copenhagen, Denmark
Claus Svarer, PhD
Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Copenhagen, Denmark
Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Copenhagen, Denmark
Gitte Knudsen, MD, DMSc
Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet|Department of Clinical Medicine, University of Copenhagen
Copenhagen, Denmark|Copenhagen, Denmark
Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet|Department of Clinical Medicine, University of Copenhagen
Copenhagen, Denmark|Copenhagen, Denmark
Jonas Svensson, PhD
Centre for Psychiatry Research, Department of Clinical Neuroscience, Karolinska Institutet|Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Stockholm, Sweden|Copenhagen, Denmark
Centre for Psychiatry Research, Department of Clinical Neuroscience, Karolinska Institutet|Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Stockholm, Sweden|Copenhagen, Denmark
Pontus Plavén-Sigray, PhD
Centre for Psychiatry Research, Department of Clinical Neuroscience, Karolinska Institutet|Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Stockholm, Sweden|Copenhagen, Denmark
Centre for Psychiatry Research, Department of Clinical Neuroscience, Karolinska Institutet|Neurobiology Research Unit, Copenhagen University Hospitalet Rigshospitalet
Stockholm, Sweden|Copenhagen, Denmark
Introduction:
Developing biomarkers that capture age-related biological changes in the human brain is crucial for understanding the causal role of biological aging in the genesis of neurodegenerative disorders (Mattson and Arumugam 2018). Brain serotonin 2A (5-HT2A) receptor binding, measured using positron emission tomography (PET), has been shown to decline with age and has been related to various degenerative disorders (Karrer et al. 2019; Hasselbalch et al. 2008). Similarly, gray matter (GM) volume in the brain, measured using T1 weighted magnetic resonance imaging (MRI), has been shown to decrease during adulthood (Bethlehem et al. 2022), and neuronal atrophy is a hallmark of e.g., dementia.
In this study, we investigated the decline in 5-HT2A receptor binding using the brain age paradigm (Franke et al. 2010) to evaluate its usefulness as a biomarker for biological aging. Specifically, we aimed to 1) predict brain age using PET data, i.e., 5-HT2A receptor binding outcomes; 2) compare to predictions based on MRI data, i.e., gray matter volume; and 3) investigate whether a multimodal approach combining PET and MRI derived data yields improved predictions over unimodal approaches.
In this study, we investigated the decline in 5-HT2A receptor binding using the brain age paradigm (Franke et al. 2010) to evaluate its usefulness as a biomarker for biological aging. Specifically, we aimed to 1) predict brain age using PET data, i.e., 5-HT2A receptor binding outcomes; 2) compare to predictions based on MRI data, i.e., gray matter volume; and 3) investigate whether a multimodal approach combining PET and MRI derived data yields improved predictions over unimodal approaches.
Methods:
We investigated 209 healthy subjects between 18 and 82 years (mean=36.6, std=16.8). The data was derived from the CIMBI database (Knudsen et al. 2016). All subjects had imaging of the 5-HT2A receptor using PET and structural imaging of the brain using MRI. Binding potentials and GM volumes were quantified in 14 cortical and subcortical regions with 5-HT2A receptors (Varnäs, Halldin, and Hall 2004).
Various machine learning algorithms were implemented to predict age based on imaging-derived feature sets: only PET, only MRI, and PET + MRI (Figure 1). We selected commonly used algorithms for brain age prediction (Baecker et al. 2021). Multimodal models using PET and MRI data were implemented using a stacking approach. A dummy regressor and a pre-trained structural MRI-based brain age prediction software (Leonardsen et al. 2022) were also implemented as a reference.
All models were trained and evaluated using a 20-times repeated 5-fold CV setup. We reported the average mean absolute error (MAE) and the average correlation between predicted and chronological age over all hold-out folds.
Various machine learning algorithms were implemented to predict age based on imaging-derived feature sets: only PET, only MRI, and PET + MRI (Figure 1). We selected commonly used algorithms for brain age prediction (Baecker et al. 2021). Multimodal models using PET and MRI data were implemented using a stacking approach. A dummy regressor and a pre-trained structural MRI-based brain age prediction software (Leonardsen et al. 2022) were also implemented as a reference.
All models were trained and evaluated using a 20-times repeated 5-fold CV setup. We reported the average mean absolute error (MAE) and the average correlation between predicted and chronological age over all hold-out folds.
·Figure 1: Overview of the various prediction pipelines.
Results:
Overall, all models using PET, MRI, and PET+MRI-derived features performed better than the dummy regressor and worse than the state-of-the-art software pyment. The results for all trained models were visualized in Figure 2. Following, we reported the results for the best model in each feature set, which was Bayesian Ridge Regression (Bridge) for PET and the Gaussian process regressor using an RBF kernel (rbfGPR) for MRI and PET+MRI.
We found that cerebral 5-HT2A receptor binding predicted chronological age accurately (average MAE=6.63 years, r=0.87) and outperformed gray matter volume-based predictions (average MAE=7.76 years, r=0.79). The difference between PET and MRI-based predictions was statistically significant (p=0.04).
We further found that the accuracy increased after combining MRI- and PET-derived regional measures (average MAE=5.93 years, r=0.88). However, the difference to the PET-based model was not statistically significant (p=0.14).
We found that cerebral 5-HT2A receptor binding predicted chronological age accurately (average MAE=6.63 years, r=0.87) and outperformed gray matter volume-based predictions (average MAE=7.76 years, r=0.79). The difference between PET and MRI-based predictions was statistically significant (p=0.04).
We further found that the accuracy increased after combining MRI- and PET-derived regional measures (average MAE=5.93 years, r=0.88). However, the difference to the PET-based model was not statistically significant (p=0.14).
·Figure 2: MAE for all the trained models.
Conclusions:
We showed that 5-HT2A receptor binding could be used to predict chronological age accurately. Those age predictions were more accurate than predictions based solely on volumetric MRI data. Combining PET and MRI data into one model increased the accuracy, suggesting that both contribute unique information when predicting age. This indicated that both measures were affected by distinct aging-related biological mechanisms. Therefore, 5-HT2A receptor binding might be suitable as a putative biomarker for aging-related changes in the human brain. However, further studies assessing if 5-HT2A-based age predictions are predictive of poor health outcomes are necessary to establish its use as a biomarker.
Lifespan Development:
Aging 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Keywords:
Aging
Data analysis
Machine Learning
MRI
Positron Emission Tomography (PET)
Seretonin
Other - Gray Matter
1|2Indicates the priority used for review
Provide references using author date format
Bethlehem, R.A.I. et al. (2022) ‘Brain charts for the human lifespan’, Nature. Available at: https://doi.org/10.1038/s41586-022-04554-y.
Franke, K. et al. (2010) ‘Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: Exploring the influence of various parameters’, NeuroImage, 50(3), pp. 883–892. Available at: https://doi.org/10.1016/j.neuroimage.2010.01.005.
Hasselbalch, S.G. et al. (2008) ‘Reduced 5-HT2A receptor binding in patients with mild cognitive impairment’, Neurobiology of Aging, 29(12), pp. 1830–1838. Available at: https://doi.org/10.1016/j.neurobiolaging.2007.04.011.
Karrer, T.M. et al. (2019) ‘Reduced serotonin receptors and transporters in normal aging adults: a meta-analysis of PET and SPECT imaging studies’, Neurobiology of Aging, 80, pp. 1–10. Available at: https://doi.org/10.1016/j.neurobiolaging.2019.03.021.
Knudsen, G.M. et al. (2016) ‘The Center for Integrated Molecular Brain Imaging (Cimbi) database’, NeuroImage, 124, pp. 1213–1219. Available at: https://doi.org/10.1016/j.neuroimage.2015.04.025.
Leonardsen, E.H. et al. (2022) ‘Deep neural networks learn general and clinically relevant representations of the ageing brain’, NeuroImage, 256(December 2021), p. 119210. Available at: https://doi.org/10.1016/j.neuroimage.2022.119210.
Mattson, M.P. and Arumugam, T.V. (2018) ‘Hallmarks of Brain Aging: Adaptive and Pathological Modification by Metabolic States’, Cell Metabolism, 27(6), pp. 1176–1199. Available at: https://doi.org/10.1016/j.cmet.2018.05.011.
Properzi, M.J. et al. (2019) ‘Nonlinear Distributional Mapping (NoDiM) for harmonization across amyloid-PET radiotracers’, NeuroImage, 186, pp. 446–454. Available at: https://doi.org/10.1016/j.neuroimage.2018.11.019.
Varnäs, K., Halldin, C. and Hall, H. (2004) ‘Autoradiographic distribution of serotonin transporters and receptor subtypes in human brain’, Human Brain Mapping, 22(3), pp. 246–260. Available at: https://doi.org/10.1002/hbm.20035.