Predicting brain-age from multimodal imaging data captures cognitive impairment
Introduction
The brain continues to change throughout adult life. Structural aspects, such as cortical thinning, demonstrate robust patterns of alteration during adulthood (Hogstrom et al., 2013, Storsve et al., 2014). Likewise, age-related differences in brain function, demonstrated through studies of functional connectivity, have also been observed (Damoiseaux et al., 2008, Dennis and Thompson, 2014).
Establishing the trajectories of such changes over the lifespan provides a basis for characterizing clinically relevant deviations (Ziegler et al., 2012, Raz and Rodrigue, 2006). Brain-based age prediction offers a promising approach for providing personalized biomarkers of future cognitive impairments by capturing deviations from typical development of brain structure and function.
Brain-based age prediction aims to estimate a person's age based on brain data acquired using magnetic resonance imaging (MRI, Franke et al., 2010, Franke and Gaser, 2012). In a first step, an age prediction model is trained based on brain imaging data from a large lifespan sample. In a second step, this model can be used to estimate a novel individual's age based solely on their brain-imaging data. By comparing a person's estimated age with their chronological age, conclusions about age-typical and atypical brain development can be drawn.
Brain-based age prediction exemplifies a larger trend in neuroscience (Bzdok, 2016, Gabrieli et al., 2015, Pereira et al., 2009, Varoquaux and Thirion, 2014) and psychology (Yarkoni and Westfall, 2016) to move from correlative to predictive studies, often using tools from machine learning. Individual brain-based prediction and classification may give rise to brain imaging-based biomarkers that could aid clinical diagnostics, for instance, by predicting an individual's risk of developing dementia based on their brain (Bron et al., 2015).
One successful age prediction framework is based on structural brain data analyzed with voxel-based morphometry (VBM, Franke et al., 2010, Franke and Gaser, 2012). Using this approach, accelerated brain aging was found in patients with Alzheimer's disease (Franke et al., 2010, Franke and Gaser, 2012), traumatic brain injuries (Cole et al., 2015), psychiatric disorders (Koutsouleris et al., 2013), and subjects with risks to physical health (Franke et al., 2014). This brain-age metric can also predict the future conversion from mild cognitive impairment to Alzheimer's disease (Gaser et al., 2013). This computational approach is not restricted to showing accelerated brain aging as a negative effect but has also been used to demonstrate the positive effects of education, physical exercise (Steffener et al., 2016), and meditation (Luders et al., 2016) on brain aging. Other work has shown that accelerated brain development is related to accelerated cognitive development in young subjects (Erus et al., 2014).
In addition to brain structure, functional connectivity based on resting-state fMRI data (Craddock et al., 2013) also has the potential to provide clinically relevant biomarkers (Craddock et al., 2009, Castellanos et al., 2013), as the data is easily acquired in a clinical setting (Greicius, 2008, Damoiseaux et al., 2012). Similar to the structural age estimation approach, Dosenbach et al. (2010) demonstrated that this is also feasible with resting-state functional connectivity data from young subjects. As different MRI modalities capture not only shared but also unique information about brain aging (Groves et al., 2012), prediction accuracy may benefit by incorporating these additional sources of information. For instance, Brown et al. (2012) and Erus et al. (2014) have shown that combining information from gray and white matter anatomy increases prediction accuracy in young subjects. The present study investigates how combining data from two even more dissimilar sources, brain anatomy and functional connectivity, influences age prediction in a lifespan sample. This is important as function and structure convey converging as well as diverging information (Damoiseaux and Greicius, 2009).
While machine-learning methods enable predictions on a single-subject level, factors driving these predictions are often difficult to determine. Predictions that appear to be based on brain information may actually be driven by confounds. One major confound in functional and structural MRI is head motion (Satterthwaite et al., 2013, Power et al., 2012, Reuter et al., 2015, Alexander-Bloch et al., 2016). For instance, head motion can make cortex appear thinner (Reuter et al., 2015). An age-related increase in head motion might give rise to a supposedly ‘brain-based’ age predictor that relies heavily on head motion. Furthermore, while machine learning models are trained on one dataset and evaluated on another, in neuroimaging these datasets often come from the same study, i.e., same site and scanner. In such cases, models may overfit one site's subtle idiosyncrasies, rendering poor predictive power for data from another site. Therefore, in the current study we aimed to address these confounds by determining the effect of head motion on brain-based age prediction and predictive performance on data from a novel site.
The present study investigates (i) whether incorporating multiple imaging modalities increases prediction accuracy, (ii) whether cognitive impairments are related to brain aging, and (iii) how robust our predictive models are, specifically regarding head motion and generalizabilty to new datasets. Using data from brain anatomy and functional connectivity, we show that (i) incorporating multiple modalities increases predictive performance, (ii) cognitive impairments are related to advanced brain aging, and (iii) our models are robust as they are not driven by head motion and generalize reasonably to new datasets.
Section snippets
Materials: lifespan data and preprocessing
Two independent samples were investigated in this study: the LIFE (Loeffler et al., 2015) and the Enhanced Nathan Kline Institute – Rockland sample (NKI, Nooner et al., 2012). Since the majority of analyses is performed on the LIFE dataset, the NKI set is described in detail in Appendix A.
Age prediction
Models were trained to predict age based on a variety of input data, i.e., functional connectomes of two different spatial resolutions and measures of cortical anatomy (cortical thickness, cortical surface area, subcortical volumes). A schematic overview of the age prediction analysis is shown in Fig. 1.
Results
Our results demonstrate that (i) incorporating multiple brain imaging modalities increases age prediction performance (Fig. 2, Fig. 3); (ii) subjects with objective cognitive impairment show advanced brain aging compared to subjects without objective cognitive impairment (Fig. 4); (iii) our prediction models are robust against confounds (Fig. 5), i.e., not driven by head motion and generalize to new datasets. For the comparison of modalities, data for models of all modalities will be presented.
Discussion
The aim of the current study was to establish a novel multimodal brain-based age prediction framework that makes use of information from anatomy and functional connectivity. We found that (i) including multimodal information increases prediction accuracy, (ii) objective cognitive impairment is associated with increased brain aging, and (iii) our framework is robust against confounds, most importantly, against head motion, and generalizes to new datasets, especially if the training set is
Conclusions
In the present study, we demonstrated that including information from multiple MR modalities, i.e., anatomy and functional connectivity, increased accuracy of brain-based age prediction. Brain-age measured with this multimodal framework was accelerated in subjects with cognitive impairment. Importantly, head motion does not drive brain-based age prediction and predictive models generalize to new datasets, especially if those are trained on large and heterogeneous datasets. Given these findings,
Acknowledgments
The first author thanks all colleagues inside and outside the Max Planck Institute for Human Cognitive and Brain Sciences that provided valuable feedback for this project, especially the members of the Neuroanatomy and Connectivity Group.
We would like to thank the Enhanced Nathan Kline Institute-Rockland Sample initiative for sharing their data.
Franziskus Liem is supported by the Swiss National Science Foundation (SNSF), grant number P2ZHP1_155200. Gaël Varoquaux and Mehdi Rahim are supported
References (73)
- et al.
A reproducible evaluation of ANTs similarity metric performance in brain image registration
NeuroImage
(2011) - et al.
A component based noise correction method (CompCor) for BOLD and perfusion based fMRI
NeuroImage
(2007) - et al.
Multi-level bootstrap analysis of stable clusters in resting-state fMRI
NeuroImage
(2010) - et al.
Standardized evaluation of algorithms for computer-aided diagnosis of dementia based on structural MRIthe CADDementia challenge
NeuroImage
(2015) - et al.
Neuroanatomical assessment of biological maturity
Curr. Biol.
(2012) - et al.
Clinical applications of the functional connectome
NeuroImage
(2013) - et al.
A calibration method for quantitative BOLD fMRI based on hyperoxia
NeuroImage
(2007) - et al.
Building a Science of Individual Differences from fMRI
Trends Cogn. Sci.
(2016) FreeSurfer
NeuroImage
(2012)- et al.
“Mini-mental state”. A practical method for grading the cognitive state of patients for the clinician
J. Psychiatr. Res.
(1975)