Predicting brain age with deep learning from raw imaging data results in a reliable and heritable biomarker

Highlights

• Chronological age can be accurately predicted using convolutional neural networks.

• Age predicted is accurate even using raw structural neuroimaging data.

• Brain-predicted age can be generated in a clinically applicable timeframe.

• Brain-predicted age is significantly heritable.

• Brain-predicted age is highly reliable, both within and between scanners.

Abstract

Machine learning analysis of neuroimaging data can accurately predict chronological age in healthy people. Deviations from healthy brain ageing have been associated with cognitive impairment and disease. Here we sought to further establish the credentials of ‘brain-predicted age’ as a biomarker of individual differences in the brain ageing process, using a predictive modelling approach based on deep learning, and specifically convolutional neural networks (CNN), and applied to both pre-processed and raw T1-weighted MRI data.

Firstly, we aimed to demonstrate the accuracy of CNN brain-predicted age using a large dataset of healthy adults (N = 2001). Next, we sought to establish the heritability of brain-predicted age using a sample of monozygotic and dizygotic female twins (N = 62). Thirdly, we examined the test-retest and multi-centre reliability of brain-predicted age using two samples (within-scanner N = 20; between-scanner N = 11). CNN brain-predicted ages were generated and compared to a Gaussian Process Regression (GPR) approach, on all datasets. Input data were grey matter (GM) or white matter (WM) volumetric maps generated by Statistical Parametric Mapping (SPM) or raw data.

CNN accurately predicted chronological age using GM (correlation between brain-predicted age and chronological age r = 0.96, mean absolute error [MAE] = 4.16 years) and raw (r = 0.94, MAE = 4.65 years) data. This was comparable to GPR brain-predicted age using GM data (r = 0.95, MAE = 4.66 years). Brain-predicted age was a heritable phenotype for all models and input data (h² ≥ 0.5). Brain-predicted age showed high test-retest reliability (intraclass correlation coefficient [ICC] = 0.90–0.99). Multi-centre reliability was more variable within high ICCs for GM (0.83–0.96) and poor-moderate levels for WM and raw data (0.51–0.77).

Brain-predicted age represents an accurate, highly reliable and genetically-influenced phenotype, that has potential to be used as a biomarker of brain ageing. Moreover, age predictions can be accurately generated on raw T1-MRI data, substantially reducing computation time for novel data, bringing the process closer to giving real-time information on brain health in clinical settings.

Introduction

The human brain changes across the adult lifespan. This process of brain ageing occurs in accord with a general decline in cognitive performance, cognitive ageing. Although the changes associated with brain ageing are not explicitly pathological, with increasing age comes increasing risk of neurodegenerative disease and dementia (Abbott, 2011). However, the wide range of onset ages for age-associated brain diseases indicates that the effects of ageing on the brain vary greatly between individuals. Thus, advancing our understanding of brain ageing and identifying biomarkers of the process are vital to help improve detection of early-stage neurodegeneration and predict age-related cognitive decline.

One promising approach to identifying individual differences in brain ageing derives from the research showing that neuroimaging data can be used to accurately predict chronological age in healthy individuals, using machine learning (Dosenbach et al., 2010, Franke et al., 2010). By ‘learning’ the correspondence between patterns in structural or functional neuroimaging data and an age ‘label’, machine-learning algorithms can formulate massively high-dimensional regression models, fitting large neuroimaging datasets as independent variables to predict chronological age as the dependent variable. The resulting brain-based age predictions are generally highly accurate, particularly when algorithms learn from large training datasets and are applied to novel or ‘left-out’ data (i.e., test datasets).

Neuroimaging-derived age predictions have been explored in the context of different brain diseases. By training models on healthy individuals, brain-based predictions of age can then be made in independent clinical samples. If ‘brain-predicted age’ is greater than an individual's chronological age, this is thought to reflect some aberrant accumulation of age-related changes to the brain. The degree of this ‘added’ brain ageing can be simply quantified by subtracting chronological age from brain-predicted age. This approach is being used more frequently and has demonstrated increased brain-predicted age in adults with mild cognitive impairment who progress to Alzheimer's (Franke and Gaser, 2012, Gaser et al., 2013), after traumatic brain injury (Cole et al., 2015), in schizophrenia (Koutsouleris et al., 2013, Schnack et al., 2016), HIV (Cole et al., 2017c), epilepsy (Pardoe et al., 2017), Down's syndrome (Cole et al., 2017a) and diabetes (Franke et al., 2013). At the same time, brain-predicted age has been used to demonstrate protective influences on brain ageing, including meditation (Luders et al., 2016) and increased levels of education and physical exercise (Steffener et al., 2016). Evidently, the extent to which one's brain resembles the typical structure or function appropriate for one's age can be affected by both positive and negative influences. By conceptualising brain ageing in this manner, highly-complex multivariate datasets and statistical procedures can be reformulated into an intuitively straightforward and widely-applicable biomarker. However, the practicality of using such a marker clinically, its reliability and relevance for normal variation in brain ageing need to be further demonstrated.

One hindrance to clinical applications for neuroimaging generally is the time needed for image ‘post-processing’ after acquisition (referred to as ‘pre-processing’ by neuroimagers), which can take hours or days, while clinical decisions often need to occur in minutes or less. Regardless of learning algorithm, previous brain-predicted age studies have required several pre-processing stages. Such steps are typically a sequence of data transformations that produce a representation of the original images that is sufficiently structured, compact and informative to support machine learning. These include the removal of non-brain tissue (i.e., skull stripping or brain extraction), affine or non-linear image registration, interpolation and smoothing. While pre-processing may reduce noise and permit voxelwise inter-individual statistical comparisons, there are numerous additional assumptions required for any pre-processing pipeline. These assumptions are often not met, particularly when analysing brain images containing gross pathology (Avants et al., 2008, Liu et al., 2015) and can even be an increased source of error. Recently, however, modelling methods that require little or no image pre-processing have become available, including so-called ‘deep learning’.

The resurgence of interest in artificial neural networks for learning data representations, deep learning, offers a new way of approaching statistical modelling in neuroimaging, thanks to improvements in computing infrastructure. When sufficiently large volumes of data are available, no ‘hand-engineering’ (i.e., manually selecting a priori which features should be used as input) is needed as the deep learning algorithm is able to infer a compact representation of the data, starting only with raw images as input, which is optimally tailored for the particular predictive modelling task at hand. In this respect, deep learning offers several practical advantages for high-dimensional prediction tasks, that should enable the learning of both physiologically-relevant representations and latent relationships (Plis et al., 2014). Of particular interest to us is the potential for deep learning techniques, such as convolutional neural networks (CNN), to make predictions from raw, unprocessed neuroimaging data, thus obviating the reliance on time-consuming pre-processing and improving the clinical applicability of models of brain ageing.

Beyond improving clinical applicability, a biomarker of brain ageing needs to relate to naturally occurring variation, such as that caused by genetic factors. Many aspects of brain ageing and susceptibility to age-related brain disease are thought to be under genetic influence (Lee and Sachdev, 2014, Lu et al., 2004, Peters, 2006, Teter and Finch, 2004). Therefore, demonstrating a brain ageing biomarker is sensitive to genetic influences gives some external, genetic, validity to the measure. Furthermore, if a neuroimaging biomarker is heritable, this motivates further research into specific candidate genes, or sets of genes, that may affect this aspect of brain ageing. These candidate genes can then, in turn, provide biological targets for pharmacological interventions which aim to improve brain health in older adults.

Another important facet of any biomarker is reliability. If a biomarker is to be evaluated longitudinally, in clinical trials or research settings, to track change over time, establishing test-retest reliability is vital. Furthermore, as many neuroimaging studies are now international collaborative efforts, data collection often takes place across multiple scanning sites. Therefore, between-scanner reliability, which indicates that a method of obtaining a biomarker is generalizable to data acquired from other sites, is of increasing importance.

In this work, we sought to establish the credentials of CNN-predicted age as a potential biomarker of brain ageing in three different ways: 1) Demonstrate that CNNs can accurately predict age using structural neuroimaging data and compare predictions using pre-processed and ‘raw’ input data; 2) Establish the heritability of brain-predicted age using a sample of monozygotic and dizygotic twins; 3) Assess both the test-retest (i.e., within-scanner) and multi-centre (i.e., between-scanner) reliability of brain-predicted age.