Brain-Age Prediction: Systematic Evaluation of Site Effects, and Sample Age Range and Size
Poster No:
1267
Submission Type:
Abstract Submission
Authors:
Yuetong Yu1, Hao-Qi Cui1, Shalaila Haas2, Faye New2, Nicole Sanford1, Kevin Yu1, Denghuang Zhan1, Guoyuan Yang3, Jia-hong Gao4, Dongtao Wei5, Jiang Qiu5, Boris Bernhardt6, Paul Thompson7, Sophia Frangou1, RUIYANG GE1, ENIGMA consortium8
Institutions:
1University of British Columbia, Vancouver, British Columbia, 2Icahn School of Medicine at Mount Sinai, New York, NY, 3Beijing Institute of Technology, Beijing, China, 4Peking University, Beijing, Beijing, 5Southwest University, Chongqing, Chongqing, 6Montreal Neurological Institute and Hospital, Montreal, Quebec, 7USC, Marina Del Rey, CA, 8Imaging Genetics Center, Mark and Mary Stevens Neuroimaging and Informatics Institute, Marina del Rey, CA, CA
First Author:
Co-Author(s):
ENIGMA consortium
Imaging Genetics Center, Mark and Mary Stevens Neuroimaging and Informatics Institute
Marina del Rey, CA, CA
Imaging Genetics Center, Mark and Mary Stevens Neuroimaging and Informatics Institute
Marina del Rey, CA, CA
Introduction:
Structural neuroimaging data have been utilized to estimate the biological age of the brain (brain-age). This metric has shown associations with other biologically and behaviorally significant indicators of brain development and aging. The continuous interest in brain age research underscores the necessity for robust, publicly accessible pre-trained brain age models, constructed using extensive data from healthy individuals. In response to this demand, we previously introduced a developmental brain age model (5-22 years). In this work, we expand upon our prior efforts to create, empirically validate, and share a pre-trained brain age model that spans across a broader spectrum of the human lifespan.
Methods:
We examined the impact of site harmonization, age range, and sample size on brain age prediction using a discovery sample comprising 35,683 healthy individuals (53.59% female, age range 5-90 years), a replication sample (N=2101, 55.35% female, age range 8-80 years), and a longitudinal consistency sample (N=377, 49.87% female, age range: 9-25 years).
Morphometric features were extracted from FreeSurfer processing, and included Desikan-Killiany atlas measures of cortical thickness (n=68), cortical surface area (n=68), and regional subcortical volumes (n=14) based on the Aseg atlas. Following our previous work [1], support vector regression with radial basis function kernel was adopted to train the sex-specific brain-age models. The primary performance measures for all models were the mean absolute error (MAE), and the correlation coefficient (CORR) between brain-age and chronological age.
The procedures used to generate optimized sex-specific models are illustrated in Figure 1. For all models, hyper-parameter tuning was performed in the discovery sample using a grid search approach in a 10-fold cross-validation scheme across five repetitions. We evaluated 3 site handling strategies in each of 5 scenarios after partitioning the discovery sample into different age bins as follows: (i) a single bin with the full sample age range (5-90 years); (ii) nine bins each covering sequential 10-year intervals; (iii) four bins each covering sequential 20-year intervals; (iv) three bins each covering sequential 30-year intervals; (v) two age bins each covering sequential 40-year intervals. The following 3 site handling strategies were applied to each bin: (i) data residualization with respect to the scanning site using Combat-GAM [2]; (ii) data residualization with respect to the scanning site using a generalized linear model [3], and (iii) no site harmonization. The approach and age partition with the best-performing MAE and CORR values were considered for further evaluation. To estimate the minimum sample size required for a stable model, the discovery sample was randomly partitioned into 30 sex-specific subsets, ranging from 200 to 6,000 participants in increments of 200, without replacement.
The model with the lowest replication MAE and highest replication CORR in the replication sample and in the longitudinal consistency sample was chosen as the preferred model.
Morphometric features were extracted from FreeSurfer processing, and included Desikan-Killiany atlas measures of cortical thickness (n=68), cortical surface area (n=68), and regional subcortical volumes (n=14) based on the Aseg atlas. Following our previous work [1], support vector regression with radial basis function kernel was adopted to train the sex-specific brain-age models. The primary performance measures for all models were the mean absolute error (MAE), and the correlation coefficient (CORR) between brain-age and chronological age.
The procedures used to generate optimized sex-specific models are illustrated in Figure 1. For all models, hyper-parameter tuning was performed in the discovery sample using a grid search approach in a 10-fold cross-validation scheme across five repetitions. We evaluated 3 site handling strategies in each of 5 scenarios after partitioning the discovery sample into different age bins as follows: (i) a single bin with the full sample age range (5-90 years); (ii) nine bins each covering sequential 10-year intervals; (iii) four bins each covering sequential 20-year intervals; (iv) three bins each covering sequential 30-year intervals; (v) two age bins each covering sequential 40-year intervals. The following 3 site handling strategies were applied to each bin: (i) data residualization with respect to the scanning site using Combat-GAM [2]; (ii) data residualization with respect to the scanning site using a generalized linear model [3], and (iii) no site harmonization. The approach and age partition with the best-performing MAE and CORR values were considered for further evaluation. To estimate the minimum sample size required for a stable model, the discovery sample was randomly partitioned into 30 sex-specific subsets, ranging from 200 to 6,000 participants in increments of 200, without replacement.
The model with the lowest replication MAE and highest replication CORR in the replication sample and in the longitudinal consistency sample was chosen as the preferred model.
Results:
(1) The accuracy of age prediction from morphometry data was higher when no site harmonization was applied (Fig 1A); (2) dividing the sample into two age-bins (5-40 and 40-90 years) provided a better balance between model accuracy and explained age variance than any other alternatives (Fig1A, Fig2A), as well as achieved optimal consistency on the longitudinal data (Fig2B); (3) model accuracy for brain-age prediction plateaued at a sample size exceeding 1,600 participants (Fig1B). These findings have been incorporated into an open platform for individualized neuroimaging metrics [4].
Conclusions:
In this work, we present empirically validated models for brain-age that can accommodate studies using data across most of the lifespan. We outline the methodological choices that have led to these models and their performance within and across samples as well as longitudinally.
Lifespan Development:
Lifespan Development Other 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Multivariate Approaches
Keywords:
Aging
Machine Learning
Modeling
Morphometrics
MRI
Multivariate
NORMAL HUMAN
1|2Indicates the priority used for review
·Fig 1
·Fig 2
Provide references using author date format
2. Pomponio, R., et al., Harmonization of large MRI datasets for the analysis of brain imaging patterns throughout the lifespan. NeuroImage, 2020. 208: p. 116450.
3. de Lange, A.M.G., et al., Mind the gap: Performance metric evaluation in brain‐age prediction. Human Brain Mapping, 2022. 43(10): p. 3113-3129.
4. https://centilebrain.org/#/brainAge2