Prediction of individual subject's age across the human lifespan using diffusion tensor imaging: A machine learning approach
Introduction
Post-mortem studies have previously shown that the human brain undergoes several changes during maturation and ageing. These changes include, myelin level changes (Benes et al., 1994), synaptic density variations (Huttenlocher, 1979, Huttenlocher and de Courten, 1987) and Schwann cell subunit density changes (Kanda et al., 1991). Recently though, magnetic resonance imaging (MRI) techniques such as diffusion tensor imaging (DTI) have allowed characterization of anatomical tissue in vivo. This has led to increased interest in applying DTI measurements in lifespan cerebral maturation and ageing studies as demonstrated in these detailed reviews (Cascio et al., 2007, Gunning-Dixon et al., 2009, Johansen-Berg and Behrens, 2009, Moseley, 2002, Sullivan and Pfefferbaum, 2006).
Several DTI microstructural measurements have proved to be particularly important in this endeavour and below we briefly explore these measurements. First, fractional anisotropy (FA) is used to quantify tissue water diffusion directionality or anisotropy in vivo (Basser and Pierpaoli, 1996). A higher FA value (maximum = 1, minimum = 0) represents diffusion occurring along one direction but largely restricted in all other directions (Johansen-Berg and Behrens, 2009). Higher FA values are often associated with highly myelinated white-matter tracts and vice versa (Hasan et al., 2008, Klingberg et al., 1999, Kochunov et al., 2012). Numerous studies have used FA as a proxy measure of myelin levels with success (Abe et al., 2002, Barnea-Goraly et al., 2005, Eluvathingal et al., 2007, Grieve et al., 2007, Hasan et al., 2008, Head et al., 2004, Hsu et al., 2010, Kennedy and Raz, 2009, Kochunov et al., 2012, Lebel et al., 2008, Lebel et al., 2012, Ota et al., 2006, Pfefferbaum et al., 2000, Salat et al., 2005a, Schmithorst et al., 2002, Westlye et al., 2010). In a nutshell, these studies point to two consistent themes. First, FA increases consistently during healthy childhood and adolescence in a majority of white matter tracts. Second, FA decreases consistently with ageing. These results have also been confirmed by white matter volumetric studies showing that white matter tissue undergoes accelerated volumetric changes during early and late life with a plateau in early and middle adulthood (Westlye et al., 2010). Specifically, age related FA decline has been reported in anterior cingulum bundles, frontal gyrus, cerebral peduncle and corpus callosum (Brown et al., 2012, Malykhin et al., 2011, Pfefferbaum et al., 2005, Salat et al., 2005a, Sullivan and Pfefferbaum, 2006, Virta et al., 1999). Notably, FA decline is statistically comparable in both male and female and linear from about 20 years onwards (Malykhin et al., 2011, Ota et al., 2006, Pfefferbaum and Sullivan, 2003, Sullivan and Pfefferbaum, 2006, Virta et al., 1999).
Other DTI measurements which offer complementary information not available with FA include radial diffusivity (RD), mean diffusivity (MD) and axial diffusivity (AD). Since tissue water molecular anisotropy and diffusivity are not only specific to white-matter tissue (Q. Wang et al., 2010), we envisaged that these additional DTI-derived metrics may be predictive of individual subject's age as well to help in elucidating the cerebral maturation and ageing process in both grey and white matter tissues. Correspondingly, several studies have recently applied MD, AD and RD to study diffusivity in both white and grey matter tissues with success (Benedetti et al., 2006, Kumar et al., 2012, Ota et al., 2006, Song et al., 2002, Song et al., 2003, Stadlbauer et al., 2008b, Wang et al., 2010a). In particular, studies comparing healthy young and older adults conclude that AD and MD in grey matter tissue increase considerably with ageing (Abe et al., 2008, Bhagat and Beaulieu, 2004, Camara et al., 2007, Hasan, 2010, Hasan and Frye, 2011, Kochunov et al., 2010, Liu et al., 2012, Pfefferbaum et al., 2010, Stadlbauer et al., 2008a).
However, despite gaining significant insights into cerebral maturation and ageing using these DTI measurements, many studies preceding this report have mainly used pre-defined regions-of-interest (ROIs) and to some extent whole-brain data with univariate data analysis techniques (Johansen-Berg and Behrens, 2009, Whitcher et al., 2007). In this study, we adopted a multivariate machine-learning technique with whole-brain FA, MD, AD and RD to predict individual subject's age. There are two notable benefits in adopting this approach. First, the ability to accurately and efficiently decode a continuous variable (e.g. age) from neuroimaging scans (Haynes and Rees, 2006, Johnston et al., 2012, Mwangi et al., 2012a, Mwangi et al., 2012b). This is achieved by separating training and test data using a cross-validation process. Second, the possibility to select anatomical regions relevant to the decoding process, a step commonly known as feature subset selection (FSS) in machine learning. This process takes into account interactions of full spatial anatomical patterns without necessarily making any a-priori assumptions (Ziegler et al., 2012). This is particularly important as it has been reported that cerebral ageing is a ‘global’ process affecting major white-matter tracts simultaneously (Penke et al., 2010).
Multivariate machine learning regression models have recently been used to decode continuous targets (e.g. age, clinical scores) from neuroimaging scans with success (Brown et al., 2012, Cherubini et al., 2009, Dosenbach et al., 2010, Formisano et al., 2008, Franke et al., 2010, Kohannim et al., 2012, Mwangi et al., 2012a, Stonnington et al., 2010, Wang et al., 2010b).
In this study, we applied the relevance vector regression (RVR) algorithm (Tipping, 2001) to predict individual subject's age using whole brain FA, MD, RD and AD volumes. Specific benefits of using RVR over other equivalent algorithms such as support vector regression (SVR) (Smola and Scholkopf, 2004) should be noted. First, RVR selects relatively few features (sparse) resulting to efficiency in making predictions (Tipping, 2001). Second, RVR estimates model ‘nuisance’ or regularisation parameters automatically (Bishop, 2006, Tipping, 2001). In this study, anatomical voxels (features) relevant in making accurate predictions were selected using a multivariate recursive feature elimination method (Guyon and Elisseeff, 2003). A post-hoc analysis of these ‘relevant’ anatomical features was used to evaluate cross-sectional cerebral maturation and ageing trajectories. A major objective was to elucidate whether these maturation and ageing trajectories follow linear or non-linear paths.
Section snippets
Subjects and DTI acquisition protocol
Raw diffusion tensor scans were acquired from the International Neuroimaging Data Sharing Initiative (INDI) online database (http://fcon_1000.projects.nitrc.org/indi/pro/nki.html). Additional variables included subjects' age and a battery of clinical assessment scores. This sample was provided by the Nathan Kline Institute (NKI, NY, USA) as part of the ‘original NKI-Rockland sample’. All necessary approvals and procedures for human subject studies were followed as required by NKI. DTI images
Results
Fig. 4 summarizes RVR age prediction results from the four DTI measurements (FA, MD, AD and RD) in whole-brain white and grey matter. Overall, RD was the best in predicting age (r = 0.899, p < 0.0001). Predictions from all four DTI measurements were significant (p < 0.0001). Younger subjects (less than median age of 30.5 years) were predicted more accurately than older subjects (more than median age) as shown in Table 1.
Interestingly, whilst the model was exposed to whole-brain DTI volumes in MD, RD
Discussion
To our knowledge, this is the first study to report predictions of individual subject's age using whole-brain DTI-derived scalar metrics. All four DTI measurement volumes (FA, MD, RD and AD) successfully predicted age with high accuracy as evaluated using various statistical measures for multivariate regression models. These results demonstrate that although brain maturation and healthy ageing effects may be subtle, DTI measurements contain sufficient information to detect these effects and
Acknowledgments
Data used in this study were obtained from the International Neuroimaging Data Sharing Initiative (INDI) online database (http://fcon_1000.projects.nitrc.org/indi/pro/nki.html) as part of the original Nathan Kline Institute/Rockland sample. We acknowledge the entire INDI team involved in data acquisition and management especially Dr. Maarten Mennes for answering various questions regarding the data. The INDI team was not involved in data analysis or report preparation.
Conflict of interest
This
References (99)
- et al.
Normal aging in the central nervous system: quantitative MR diffusion-tensor analysis
Neurobiol. Aging
(2002) - et al.
Aging in the CNS: comparison of gray/white matter volume and diffusion tensor data
Neurobiol. Aging
(2008) A fast diffeomorphic image registration algorithm
NeuroImage
(2007)- et al.
Neuroanatomical assessment of biological maturity
Curr. Biol.
(2012) - et al.
The posterior parahippocampal gyrus is preferentially affected in age-related memory decline
Neurobiol. Aging
(2011) - et al.
Age-related water diffusion changes in human brain: a voxel-based approach
NeuroImage
(2007) - et al.
Diffusion tensor imaging: application to the study of the developing brain
J. Am. Acad. Child Adolesc. Psychiatry
(2007) - et al.
Aging of subcortical nuclei: microstructural, mineralization and atrophy modifications measured in vivo using MRI
NeuroImage
(2009) - et al.
Multivariate analysis of fMRI time series: classification and regression of brain responses using machine learning
Magn. Reson. Imaging
(2008) - et al.
Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters
NeuroImage
(2010)
Diffusion tensor tractography quantification of the human corpus callosum fiber pathways across the lifespan
Brain Res.
Pediatric diffusion tensor imaging: normal database and observation of the white matter maturation in early childhood
NeuroImage
Gender differences and age-related white matter changes of the human brain: a diffusion tensor imaging study
NeuroImage
Microstructural white matter changes in normal aging: a diffusion tensor imaging study with higher-order polynomial regression models
NeuroImage
Aging white matter and cognition: differential effects of regional variations in diffusion properties on memory, executive functions, and speed
Neuropsychologia
Development of the corpus callosum in childhood, adolescence and early adulthood
Life Sci.
The quantitative evaluation of functional neuroimaging experiments: mutual information learning curves
NeuroImage
Relationship between white matter fractional anisotropy and other indices of cerebral health in normal aging: tract-based spatial statistics study of aging
NeuroImage
Genetics of microstructure of cerebral white matter using diffusion tensor imaging
NeuroImage
Fractional anisotropy of water diffusion in cerebral white matter across the lifespan
Neurobiol. Aging
Support vector machines for temporal classification of block design fMRI data
NeuroImage
Microstructural maturation of the human brain from childhood to adulthood
NeuroImage
Diffusion tensor imaging of white matter tract evolution over the lifespan
NeuroImage
Frontal connections and cognitive changes in normal aging rhesus monkeys: a DTI study
Neurobiol. Aging
Statistical validation
Ecol. Model.
Stereotaxic white matter atlas based on diffusion tensor imaging in an ICBM template
Neuroimage
Age-related degeneration of corpus callosum measured with diffusion tensor imaging
NeuroImage
Frontal circuitry degradation marks healthy adult aging: evidence from diffusion tensor imaging
NeuroImage
Diffusion tensor imaging of deep gray matter brain structures: effects of age and iron concentration
Neurobiol. Aging
Visualization of nonlinear kernel models in neuroimaging by sensitivity maps
NeuroImage
Age-related alterations in white matter microstructure measured by diffusion tensor imaging
Neurobiol. Aging
When does age-related cognitive decline begin?
Neurobiol. Aging
Tract-based spatial statistics: voxelwise analysis of multi-subject diffusion data
NeuroImage
Dysmyelination revealed through MRI as increased radial (but unchanged axial) diffusion of water
NeuroImage
Diffusion tensor imaging detects and differentiates axon and myelin degeneration in mouse optic nerve after retinal ischemia
NeuroImage
Predicting clinical scores from magnetic resonance scans in Alzheimer's disease
NeuroImage
Diffusion tensor imaging and aging
Neurosci. Biobehav. Rev
Visualizing and characterizing white matter fiber structure and architecture in the human pyramidal tract using diffusion tensor MRI
Magn. Reson. Imaging
High-dimensional pattern regression using machine learning: from medical images to continuous clinical variables
NeuroImage
White matter development during childhood and adolescence: a cross-sectional diffusion tensor imaging study
Cereb. Cortex
Microstructural and physiological features of tissues elucidated by quantitative-diffusion-tensor MRI
J. Magn. Reson.
Influence of aging on brain gray and white matter changes assessed by conventional, MT, and DT MRI
Neurology
Myelination of a key relay zone in the hippocampal formation occurs in the human brain during childhood, adolescence, and adulthood
Arch. Gen. Psychiatry
Diffusion anisotropy in subcortical white matter and cortical gray matter: changes with aging and the role of CSF-suppression
J. Magn. Reson. Imaging
Pattern Recognition and Machine Learning
Genetics of brain fiber architecture and intellectual performance
J. Neurosci.
Prediction of individual brain maturity using fMRI
Science
Quantitative diffusion tensor tractography of association and projection fibers in normally developing children and adolescents
Cereb. Cortex
Age-related changes of the insular cortex and lateral ventricles: conventional MRI volumetric measures
J. Neuroimaging
Cited by (89)
A review on brain age prediction models
2024, Brain ResearchPrediction of brain age based on the community structure of functional networks
2023, Biomedical Signal Processing and ControlCitation Excerpt :The key issue of brain age prediction is to extract effective features from raw data [6,7]. Recently, the majority of brain age prediction studies have used structural neuroimaging data as features [7–11]. For example, Lancaster et al. employed the Bayesian optimization method to optimize resampling parameters based on T1-weighted magnetic resonance imaging (MRI), which improve brain age prediction performance [8].
Common and unique neural activities in subclinical depression and major depressive disorder indicate the development of brain impairments in different depressive stages
2022, Journal of Affective DisordersCitation Excerpt :Because of its characteristic flexibility and powerful classification capacity, SVM has been frequently utilized in neuroimaging studies for the classification of mood disorders and healthy subjects (Redlich et al., 2014; Mwangi et al., 2016; Stolicyn et al., 2020). Moreover, classification was validated with leave-one-out cross-validation (LOOCV) (Mwangi et al., 2013) strategy to ensure consistent performance for the modest sample size by maximizing the training subsample. Finally, the performance of the classifiers was quantified by the accuracy, sensitivity, and specificity (Rosa et al., 2015).
Brain age estimation using multi-feature-based networks
2022, Computers in Biology and MedicineWhite matter microstructure associated with anhedonia among individuals with bipolar disorders and high-risk for bipolar disorders
2022, Journal of Affective Disorders