Relationship between MRI brain-age heterogeneity, cognition, genetics and Alzheimer’s disease neuropathology

Summary

Background

Brain ageing is highly heterogeneous, as it is driven by a variety of normal and neuropathological processes. These processes may differentially affect structural and functional brain ageing across individuals, with more pronounced ageing (older brain age) during midlife being indicative of later development of dementia. Here, we examined whether brain-ageing heterogeneity in unimpaired older adults related to neurodegeneration, different cognitive trajectories, genetic and amyloid-beta (Aβ) profiles, and to predicted progression to Alzheimer’s disease (AD).

Methods

Functional and structural brain age measures were obtained for resting-state functional MRI and structural MRI, respectively, in 3460 cognitively normal individuals across an age range spanning 42–85 years. Participants were categorised into four groups based on the difference between their chronological and predicted age in each modality: advanced age in both (n = 291), resilient in both (n = 260) or advanced in one/resilient in the other (n = 163/153). With the resilient group as the reference, brain-age groups were compared across neuroimaging features of neuropathology (white matter hyperintensity volume, neuronal loss measured with Neurite Orientation Dispersion and Density Imaging, AD-specific atrophy patterns measured with the Spatial Patterns of Abnormality for Recognition of Early Alzheimer’s Disease index, amyloid burden using amyloid positron emission tomography (PET), progression to mild cognitive impairment and baseline and longitudinal cognitive measures (trail making task, mini mental state examination, digit symbol substitution task).

Findings

Individuals with advanced structural and functional brain-ages had more features indicative of neurodegeneration and they had poor cognition. Individuals with a resilient brain-age in both modalities had a genetic variant that has been shown to be associated with age of onset of AD. Mixed brain-age was associated with selective cognitive deficits.

Interpretation

The advanced group displayed evidence of increased atrophy across all neuroimaging features that was not found in either of the mixed groups. This is in line with biomarkers of preclinical AD and cerebrovascular disease. These findings suggest that the variation in structural and functional brain ageing across individuals reflects the degree of underlying neuropathological processes and may indicate the propensity to develop dementia in later life.

Funding

The National Institute on Aging, the National Institutes of Health, the Swiss National Science Foundation, the Kaiser Foundation Research Institute and the National Heart, Lung, and Blood Institute.

Research in context.

Evidence before this study

We searched PubMed for articles in English published from database inception until October 3, 2023, examining the relation between brain age measurements and dementia using structural and functional MRI. Search terms included “Alzheimer disease”, “Alzheimer’s disease”, “dementia”, “brain age”, “brain aging”, “heterogeneity brain aging”, “brain age prediction”, “structural MRI” and “functional MRI”. Previous studies have used imaging features from a single imaging modality (e.g., cortical thickness or functional connectivity) to obtain measurements of brain age. A couple studies have examined the advantage of combinations of multimodal features on brain age calculations. Brain age models are typically trained on cognitively normal individuals and are then applied to various patient populations including self-reported cognitive impairment, mild cognitive impairment, preclinical Alzheimer’s disease, sporadic early onset Alzheimer’s disease and Alzheimer’s disease with reports of increased brain ages relative to controls across all groups. However, studies have not examined how advanced brain ageing in cognitively unimpaired individuals relate to neuropathologic markers of, and predicted progression to mild cognitive impairment.

Added value of this study

The goal of this study is to understand the markers of Alzheimer’s disease neuropathology and the cognitive and genetic profiles that characterise cognitively unimpaired individuals with varying degrees of structural and functional brain age. This study will categorise each individual based on a dimensional system consisting of two separate brain age measurements (one brain age derived from structural MRI and one derived from functional MRI). We compare brain age groups on a wide range of variables including amyloid burden, Apolipoprotein E4 status, Alzheimer’s disease-like atrophy patterns (e.g., white matter hyperintensities, neuronal loss), multiple cross-sectional and longitudinal cognitive tasks, predicted progression to mild cognitive impairment and genetics.

Implications of all the available evidence

Building on previous evidence that brain age is increased in symptomatic individuals, our findings suggest that it may be possible to identify people who have a higher risk of developing dementia in the future based on brain age measurements even before the onset of symptoms. Considering that Alzheimer’s disease pathology starts decades before the onset of symptoms, longitudinal monitoring of these individuals and targeted therapies may mitigate some of the risk of future dementia.

Introduction

The concept of brain-age is increasingly used to capture individual deviations from normal ageing trajectories.¹ The biological age of the brain may be estimated using population neuroimaging data and individual differences between the predicted biological age and the individual’s chronological age (the delta) serve as non-specific measures of brain health. This provides a measure of whether an individual’s brain appears to have aged more or less (henceforth, advanced and resilient brain-age respectively) than the age-matched population. Typically, one or a combination of imaging modalities are used to obtain a single overall measure of brain-age. However, since brain ageing comprises dynamic and ongoing neurobiological processes,² aspects of the brain’s structure and function may age differently across individuals. For this reason, having two separate measures of brain-age from two modalities may capture variability more accurately and may show distinct patterns of association with cognition and disease.² While some structural³ and functional⁴ changes are expected with normal ageing, individuals with more advanced brain-age in midlife are more likely to develop dementia in the future.⁵

Neurodegenerative diseases cause deviations from normal ageing trajectories with brain changes characteristic of Alzheimer’s Disease and vascular dementia beginning almost three decades before the onset of cognitive decline.⁶ However, there is a lack of studies examining whether cognitively unimpaired individuals with advanced brain-age have brain features common to preclinical or syndromic Alzheimer’s Disease. Preclinical neuroimaging biomarkers of Alzheimer’s disease include structural gray matter loss, functional network abnormalities, white matter pathology and abnormal protein aggregation (amyloid and tau).⁷ Therefore, unimpaired older adults with advanced brain-age may exhibit higher levels of some or all of these preclinical neuroimaging biomarkers compared to people with normal or younger brain-ages.

Healthy participants with higher brain-ages show lower cognitive functioning in both adulthood and in childhood.⁸ However, it is unknown if advanced brain ageing precedes evidence of cognitive decline.

Our aim was to understand the differences in imaging and cognitive profiles of unimpaired individuals with advanced brain-age compared to individuals with resilient brain-age. We obtained separate measures of structural and functional brain-ages for each cognitively normal individual in our sample and used these brain-ages to group individuals according to the degree of advanced/delayed ageing per modality. Individuals who deviated from typical brain ageing, as defined by the structural and functional brain-age indices, fit into one of the following groups: resilient (younger predicted age in both modalities), advanced (older predicted age in both modalities) or mixed (older predicted age in one modality only). We hypothesised that the advanced group would have the most imaging markers of neurodegeneration and the most impaired cognitive profile, the resilient group would show the best profiles, and the two mixed groups would be in-between.

Methods

The data used in this study should be conceptualised in two sets. Initially, we used a larger multisite sample with a wide age range comprising 3460 individuals in order to build a more robust brain age model. However, in order to examine individuals in the upper and lower percentiles of brain age values, the middle 75% of the sample was removed and only 25% was used in further analyses, leaving 867 individuals comprising a second set. The second set was occasionally further reduced because group analyses were performed on subsets of the 867 individuals that had additional available variables, including neuroimaging markers of neurodegeneration, cognitive, amyloid or genetic data. The entire framework is summarised in Fig. 1; and Fig. 2 explains which variables are available across individuals, at baseline and longitudinally.

Fig. 1.

Schematic diagram showing a summary of the brain-age prediction framework used to obtain brain-age groups. The top left hand corner depicts the age distribution for each study. Features are extracted from structural and resting-state functional MRI data (Box 1) which are then harmonised (Box 2). The vectorised, harmonised structural and functional features become the input features to two support vector regressions to obtain separate structural and functional brain-age predictions for each individual (Box 3 and 4). The two predicted brain-ages are plotted and used to identify the 4 brain-age sub-groups (Box 5).

Fig. 2.

UpSet plot describing the overlap of baseline and longitudinal cognitive and neuroimaging variables across individuals in Set 2. The vertical bar chart shows the number of individuals who have the combination (or intersection) of variables described by the connected line and dots below each corresponding bar. Filled-in dots show which variable is part of an intersection, and the remainder are unfilled. For example, 229 individuals have data available for SPARE-AD, APOE4 allele, WMH and NODDI. The horizontal bars represent the set size. Cognitive tests: DSST: baseline Digit Symbol Substitution Task; DSST_L: longitudinal Digit Symbol Substitution Task; MMSE: baseline Mini Mental State Examination; MMSE_L: longitudinal Mini Mental State Examination; TMT-A/B: baseline Trail Making Task A and B; TMT-A/B_L: longitudinal Trail Making Task A and B. Neuroimaging data: NODDI: Neurite Orientation Dispersion And Density Imaging measures; Amyloid PET: cortical centiloid values; SPARE-AD: Spatial Patterns of Abnormality for Recognition of Early Alzheimer’s Disease index; WMH: White Matter Hyperintensity volumes. Genetic data: APOE4 Status: APOE4 allele status.

Set 1: brain-age prediction framework

Participants included in brain age model

We used a subset of 3460 unimpaired participants between the ages of 42–85 years from the multimodal, harmonised iSTAGING⁹ consortium. Structural MRI, resting state functional MRI (rs-fMRI), cognitive and genetic data were available from four different studies with six sites: the UK Biobank (UKBB, N = 1569), the Baltimore Longitudinal Study of Aging (BLSA, N = 646), the Coronary Artery Risk Development in Young Adults (CARDIA, N = 663, 3 Sites: CARDIA-1 (University Minnesota, N = 169), CARDIA-3 (Northwestern University, N = 254), CARDIA-4 (University of California–Berkeley, N = 240) and the Open Access Series of Imaging Studies (OASIS, N = 582). We used a subset of data from the UKBB rather than all available data from this site to prevent over influence from a single study. During image preprocessing, the entire UKBB dataset was randomly split into batches of 10, each containing approximately 2300 subjects. This is because the entire dataset could not be processed in one go due to the size of the data. We then included data from only one batch in this study without specific selection criteria. Supplementary Tables S1 and S2 show the scanner and demographic characteristics of each study.

Image preprocessing for brain age model

Each participant’s T1-weighted MRI scan underwent inhomogeneity correction and multi-atlas skull-stripping. The images were segmented using a multi-atlas parcellation method (MUSE¹⁰) to derive 145 single and 124 composite pre-defined anatomical regions of interest (ROI, total n = 259, Supplementary Table S3).

Rs-fMRI data were pre-processed using the UK Biobank preprocessing pipeline¹¹ (https://biobank.ctsu.ox.ac.uk/crystal/crystal/docs/brain_mri.pdf). The preprocessing steps included motion correction using MCFLIRT,¹² mean intensity normalisation, high pass temporal filtering (>0.01 Hz), removal of structured artifacts using Independent Component Analysis (ICA) + FIX processing (ICA followed by FMRIB’s ICA-based X-noiseifier^13,14). The training dataset for FIX consisted of 20 rs-fMRI datasets from Whitehall II study (WhII_Standarad.RData, TR = 3s, Resolution: 3 × 3x3mm, no spatial smoothing).

The final preprocessed rs-fMRI data were co-registered to the T1 image using FLIRT (FMRIB's Linear Image Registration Tool) with boundary-based registration cost function¹² and then aligned to standard MNI (Montreal Neurological Institute) space using a non-linear transformation (FMRIB's Linear Image Registration Tool – Non-linear: FNIRT¹⁵). Then, the mean time courses for each of the 264 ROIs in the Power ROI atlas¹⁶ were extracted from clean functional data in MNI space. A functional connectivity matrix was calculated for each individual for all ROIs using pairwise Pearson’s correlation of each region’s time course. This step resulted in a 264 × 264 Power atlas-based connectivity matrix for each individual which was then used for subsequent analysis (Supplementary Table S4).

Quality check steps for image preprocessing

A systematic quality check (QC) procedure was performed for the entire data at consortium level for both structural and functional MRI data. Overall, there were data from six different sites whose demographic and scanner information are provided in Supplementary Table S1. We had two different metrics for rs-fMRI QC: mean relative displacement and temporal signal to noise ratio (tSNR). Subjects were excluded according to the following criteria: i) if the mean relative displacement was higher than 0.2 mm or if 60% of time frames had more than 0.3 mm displacement and ii) if tSNR was less than 100. The tSNR QC criterion was not applied to data from the UK Biobank because they used a multiband acceleration factor protocol, which leads to low signal-to-noise ratio compared to other sites.

Out of 4396 rs-fMRI scans from 6 different sites, 713 subjects were excluded as they did not pass the QC criteria and a further 128 subjects were excluded because the field of view was cut (BLSA: 2, CARDIA-1: 28, CARDIA-3: 60, CARDIA-4: 35, OASIS: 3). After QC, another 95 subjects were excluded due to lack of demographic information or failed harmonisation, thus leaving 3460 subjects that were used for further analysis.

T1-weighted images were examined for any image artifacts, defacing issues, motion or reduced field of view. Following preprocessing of T1-weighted images, post-processing QC involved checking for pipeline failures such as: inadequate skull-stripping, segmentation or registration failures or inconsistent image header or poor image resolution. The was also an automated QC step which flagged any subjects that had deviations in ROI volumes based on the Mahalanobis distance to the mean volume of each ROI. These flagged subjects were visually inspected for failures.¹⁷

Harmonisation of structural and resting-state fMRI features

Of the many methods available in the literature for reducing site effects and to reduce potential bias and non-biological variability introduced by site,18, 19, 20, 21 we opted for the recently proposed CovBat method (Correcting for Covariance Batch effects) for harmonisation of rs-fMRI features.²² Functional Connectivity CovBat (which utilises ComBat as the harmonisation step) was applied to 3 Sites + 1 Site (which has 3 sub-sites) ∗264∗264∗N subjects. The method first performs ComBat to remove mean and variance shifts across different sites in rs-fMRI connectivity measures. The model included age, sex and site as covariates. It included an additional step that residualises and estimates the shift and then scales the parameters to bring the mean and variance of each site back to the pooled mean and variance.²³ Pre- and post-harmonisation connectivity matrices are shown in Supplementary Fig. S1.

Volumes for 145 ROIs derived using MUSE¹⁰ were used as the final features for the structural dataset. Site effects were corrected using a recently developed ComBat + GAM model¹⁷ in which ROI volumes are modeled by a non-linear function of age and sex using control subjects and adjusted for site shift and scale parameters. Since all subjects in our dataset are controls, the harmonisation model was trained using the full dataset. Pre- and post-harmonisation figures of structural data can be found in Pomponio et al. (2020).¹⁷

Brain-age estimation model

We implemented a quantitative index of brain changes as a function of age, known as SPARE-BA (Spatial Pattern of Atrophy for Recognition of Brain Aging).²⁴ SPARE-BA uses high dimensional support vector regression (SVR) to learn the spatial brain pattern associated with ageing. Once trained, the model is applied to a new dataset where each individual receives a score indicating how much they express the ageing pattern. Higher SPARE-BA values indicate lower brain atrophy (from the perspective of the spatial pattern associated with ageing) and therefore have a younger brain age than their chronological age and vice versa for lower SPARE-BA values.

Separate models were trained to obtain SPARE-BA scores based on structural and functional MRI data, respectively. The structural SPARE-BA model was previously trained using 145 single harmonised MUSE ROI data from the iSTAGING⁹ consortium from 4 different studies with over 4000 individuals and is detailed in Hwang et al.²⁵ We applied this structural SPARE-BA model to the sample of individuals described here (n = 3460) to obtain structural SPARE-BA scores for each individual.

The functional SPARE-BA model was trained using the sample described here (n = 3460), using the vectorised harmonised upper triangle of the connectivity matrices as input to the linear kernel SVR. Parameter optimisation was done through grid search [range: 10⁻⁴ to 10², excluding zero] in a nested cross-validation procedure. Nested cross-validation includes an inner and an outer loop. The inner loop is used for grid search and parameter optimisation (the optimised parameter is then be passed to the outer loop) while the outer loop is for model evaluation and reporting the MAE by averaging the scores from each cross-validation split.

Brain age estimates are biased in the sense that younger individuals tend to have higher predicted ages whilst older ages are underestimated. A linear age bias correction was applied to both structural and functional SPARE-BA measures to remove the known systematic bias inherent in brain age calculations.²⁶ The residuals from a linear regression of SPARE-BA scores (delta) against chronological age are the corrected (bias-free) SPARE-BA scores. Cross-validated model fits (prior to age bias correction) were assessed using mean absolute error (MAE), R² and Pearson’s correlation between predicted and chronological age. The brain-age prediction framework is depicted in Fig. 1.

Defining brain-age groups

SPARE-BA scores were plotted on a graph with structural SPARE-BA on the X-axis and functional SPARE-BA on the Y-axis (see Box 5 of Fig. 1 and Supplementary Fig. S4). Participants with SPARE-BA scores above the 75th percentile were categorised as ‘advanced’ agers and those below the 25th percentile were categorised as ‘resilient’ agers (Supplementary Fig. S4). Participants with SPARE-BA scores between the 25-75th percentiles were classified as ‘Typical’ for structural or functional brain age and removed from further analyses. This left a total of 867 individuals (age range; 43–85 years) divided across 4 groups with different levels of advanced and resilient brain ages (BA): advanced (advanced BA in both modalities, N = 291), resilient (low BA in both modalities, N = 260), RFAS (resilient functional BA and advanced structural BA, N = 163) and AFRS (advanced functional BA and resilient structural BA, N = 153).

Set 2: comparison of imaging, cognitive and genetic variables across brain-age groups

Image preprocessing for group comparisons

We compared the brain-age groups on four different types of neuroimaging data: 1) amyloid burden (whole cortex centiloid values, n = 87), 2) white matter hyperintensity (WMH, n = 730) volume, 3) intra-cellular volume fraction (ICVF), isotropic volume fraction (ISOVF) and the orientation dispersion index measures from NODDI (Neurite Orientation Dispersion and Density Imaging,^27,28 n = 362), and 4) SPARE-AD²⁹ index (Spatial Patterns of Abnormality for Recognition of Early Alzheimer’s Disease, n = 867). Positive SPARE-AD indices imply a more Alzheimer’s disease-like pattern of brain atrophy whereas a more negative SPARE-AD index implies a more typical pattern of brain morphology.

NODDI measures were available for a total of 362 individuals (advanced = 119, resilient = 100, RFAS = 69, AFRS = 74); WMH volume was available for 730 individuals (advanced = 246, resilient = 219, RFAS = 131, AFRS = 134); gray matter amyloid PET mean centiloid values were available for 87 individuals (advanced = 24, resilient = 31, RFAS = 18, AFRS = 14) and SPARE-AD was available for 867 individuals (advanced = 291, resilient = 260, RFAS = 153, AFRS = 163).

Amyloid beta and APOE status

PET amyloid data was only available from the OASIS-3 study.³⁰ Out of 582 participants from OASIS who had imaging data, 190 participants had centiloid values.³¹ Three different Siemens scanners (Siemens Biograph 40 PET/CT) were used to scan participants with two amyloid radiopharmaceuticals, [11C] PiB (Pittsburgh compound) and [18 F] AV45 Florbetapir. PiB was acquired with a 60-min dynamic PET frame and AV-45 was acquired with a 50–70 min window. All PET image analysis was performed using the PET Unified Pipeline (PUP, https://github.com/ysu001/PUP). The simple processing steps included smoothing, common resolution to 8 mm, inter-frame motion correction using image registration techniques, accounting for partial volume effects with the cerebellum as the reference region. Regional PET processing was performed using FreeSurfer segmentations. Mean cortical standard uptake values (SUVR) were available for the precuneus, pre-frontal cortex, gyrus rectus, lateral temporal cortex and the whole cortex. Centiloid values (0–100 scale) were calculated for the ROIs based on the equations by Su et al. (2018).³¹ For our analyses, we used centiloid values for the whole cortex and normalised them by gray matter volume to correct for age related atrophy and reduced accumulation of amyloid with old age. APOE4 status was derived based on the number of APOE4 alleles. We divided people into two groups: those individuals with two APOE4 alleles were in one group and people with one or zero alleles were in another group.

White matter lesion volume

White matter lesion (WML) volumes were calculated for all studies at the consortium level using FLAIR and co-registered T1 images corrected for inhomogeneity. We used a deep learning based segmentation method called DeepMRSeg³² which was built on a modified U-Net architecture combined with Inception ResNet.³³ A separate training set with ground truth segmentations were used to train the model. WML volumes were log-transformed and used in subsequent analyses.

Diffusion MRI processing and tractography

We used standard diffusion tensor imaging (DTI) measures made available by the UKBB Study. The protocol for DTI can be found in the brain imaging documentation on the UKBB website (http://biobank.ctsu.ox.ac.uk/crystal/crystal/docs/brain_mri.pdf). The data were acquired using a multi shell approach with two b-values (1000 s/mm2 and 2000 s/mm2) and 50 directions were acquired for each shell. After gradient distortion and EDDY current correction, we used NODDI modelling^27,28 to compute microstructural parameters. These included voxel-wise microstructural measures like intra-cellular volume fraction (ICVF), isotropic or free water volume fraction (ISOVF) and Orientation Dispersion Index (OD). These measures were obtained for the right tapetum, right posterior thalamic radiation and the right longitudinal fasciculus ROIs identified from the iCBM³⁴ white matter atlas (Supplementary Table S6).

Neuropsychological tests for group comparisons

Cognitive performance was available at baseline and at a follow-up time point for the following tasks: 1) the Trail Making task: duration to complete numeric path Trails A (TMT-A, n = 541) and alphanumeric path Trails B (TMT-B, n = 529); 2) the digit symbol substitution test (DSST, total number of correct answers, n = 532) and 3) the Mini Mental State Examination (MMSE, total score, n = 505). DSST and MMSE raw scores were standardised to z-scores. Follow-up cognitive tasks were completed on average 4.1 years (MMSE), 1.8 years (DSST) and 2.4 years (TMT) post baseline.

Genetic data for group comparisons

A candidate-gene approach genetic association analysis was performed. The genotype array and the imputed data were directly downloaded from UK Biobank.³⁵ The imputed single nucleotide polymorphisms (SNPs) were preprocessed with a standard quality check (QC) protocol (https://www.cbica.upenn.edu/bridgeport/data/pdf/BIGS_genetic_protocol.pdf), which is detailed elsewhere.^36,37 In summary, we first excluded related individuals (up to 2nd-degree) from the complete UKBB sample using the KING software for family relationship inference. We then removed duplicated variants from all 22 autosomal chromosomes. Individuals whose genetically identified sex did not match their self-acknowledged sex were removed. Other exclusion criteria were: i) individuals with more than 3% of missing genotypes; ii) variants with minor allele frequency (MAF) of less than 1%; iii) variants with larger than 3% missing genotyping rate; iv) variants that failed the Hardy–Weinberg test at 1 × 10-10. In total, 736 UK Biobank participants from the current study with genetic data that passed the QC were used for genetic association analysis to associate the phenotype of interest (i.e. advanced versus resilient groups) with AD-related SNPs.

To choose the candidate SNPs, we manually queried the phenotype “Alzheimer's disease” in the GWAS Catalog³⁸ on January 2023, resulting in 1142 SNPs previously associated with AD in this platform. We then overlapped them with our QC'ed genetic data. This finally included 788 SNPs for 736 UKBB participants. We present the detailed information for the 788 SNPs in Supplementary File S2.

To run the candidate-gene association analysis, a logistic regression model was fit to each valid SNP, controlling for age, sex, intracranial volume, and the first five genetic principal components. The Bonferroni method was used to adjust for multiple comparisons based on the number of SNPs tested (−log10 [p-value] > 4.20).

Statistics

Statistical analysis of set 2 data

All available data were included in this study. Power analyses are reported in Supplementary Table S5. Both neuroimaging data and baseline performance on all cognitive measures (MMSE, DSST, TMT-A, TMT-B) were compared using ANCOVAs with age, sex and site as covariates. Post-hoc Tukey’s HSD test was used to identify pairwise differences whilst simultaneously adjusting the P value to account for multiple comparisons.³⁹ Group comparisons for demographic and clinical variables were examined across groups using the Kruskal–Wallis test. For longitudinal cognitive data, we used a linear mixed effects model with a random slope and intercept to estimate the rate of change (the gradient) in longitudinal scores and used ANOVA and Kruskal–Wallis to test whether the gradients significantly differed across groups (post hoc differences were identified using Tukey’s HSD). A chi-squared test was used to compare APOE4 status across groups (n = 735, advanced = 252, resilient = 223, RFAS = 135, AFRS = 125). The Kaplan–Meier curve was derived for each group to estimate the probability of conversion from cognitively normal to MCI. We used a log-rank test to examine whether the risk of conversion differed across brain-age groups. All analyses were conducted in R version 4.0.0. P = 0.05 was selected as the threshold for significance for all analyses.

Ethics

Data used in this study were consolidated and harmonised as part of the Imaging-Based Coordinate System for Aging and Neurodegenerative Diseases (iSTAGING) study. The current study consists of 4 studies (OASIS, BLSA, CARDIA, UKBIOBANK) and 6 sites. All studies were approved by the corresponding Institutional Review Boards. Each participant consented to provide imaging and cognitive data. The UKBIOBANK study has ethical approval, and the ethics committee is detailed here: https://www.ukbiobank.ac.uk/learn-more-about-uk-biobank/governance/ethics-advisory-committe.

Role of funders

The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. All authors had full access to all the data in the study and final decision to submit for publication.

Results

Model performance in set 1

The brain age prediction models (SPARE-BA) predicted structural and functional brain-age with R² = 0.46, MAE = 5.6 years, Pearson’s r = 0.80 and R² = 0.31, MAE = 6.4 years, Pearson’s r = 0.71, respectively. Mean structural and functional SPARE-BA values are summarised in Table 1. The Pearson’s correlation between structural and functional SPARE-BA scores was r = 0.7, P < 0.001.

Table 1.

Participant demographic variables according to brain-age group.

	Advanced (N = 291)	ASRF (N = 163)	RSAF (N = 153)	Resilient (N = 260)	Test statistic (P value)
Age (Mean ± Std, Years)	62.7 ± 10.7	62.3 ± 10.7	63.4 ± 10.1	63.5 ± 9.2	Χ2 = 1.5 (P = 0.5)
Sex (Female) N (%)	168 (57.7)	99 (60.7)	88 (57.5)	140 (53.8)	Χ2 = 2.1 (P = 0.6)
Sex (Male) N (%)	123 (42.3)	64 (39.3)	65 (42.5)	120 (46.2)	Χ2 = 2.1 (P = 0.6)
Education (Mean ± Std, Years)	16.1 ± 2.7	16.4 ± 2.8	16.4 ± 3.1	16.5 ± 2.7	Χ2 = 2.3 (P = 0.5)
Number with one or more APOE ε4 alleles (yes/no)	77/175	39/96	35/90	66/157	Χ2 = −0.54 (P = 0.96)
Functional SPARE-BA (Mean ± Std, years)	10.1 ± 5.1	−8.7 ± 3.7	9.2 ± 3.9	−9.9 ± 4.4	Χ2 = 651.6 (P < 0.001)
Structural SPARE-BA (Mean ± Std, years)	9.2 ± 4.8	7.7 ± 3.5	−9.2 ± 3.4	−9.7 ± 3.7	Χ2 = 344.3 (P < 0.001)

Brain age group comparisons in set 2

Participant characteristics of each brain-age group are summarised in Table 1. There were no group differences in demographic variables. Imaging and cognitive differences across the brain-age groups are summarised in Table 2.

Table 2.

Differences in cognitive and imaging variables across brain-age groups.

	Advanced (N = 291)	ASRF (N = 163)	RSAF (N = 153)	Resilient (N = 260)	ANCOVA F statistic, P value, effect size (eta squared)
Baseline TMT-A (Mean ± Std, seconds)	27.9 ± 13.6	24.9 ± 8.9	24.4 ± 8.1	25.1 ± 7.6	F = 1.47, P = 0.03, η2 = 0.01
Baseline TMT-B (Mean ± Std, seconds)	70.9 ± 40.3	61.1 ± 23.9	60.1 ± 21.9	58.7 ± 17.7	F = 6.56, P = 0.001, η2 = 0.04
Baseline DSST z score (Mean ± Std)	−0.23 ± 1.0	0.21 ± 1.0	−0.03 ± 0.9	−0.06 ± 0.1	F = 2.56, P = 0.07, η2 = 0.02
Baseline MMSE z score (Mean ± Std)	−0.29 ± 1.1	0.09 ± 0.9	0.1 ± 0.7	0.09 ± 0.1	F = 3.60, P = 0.009, η2 = 0.03
Centiloid values (adjusted for GM) (Mean ± Std)	14.3 ± 5.3	14.2 ± 4.6	12.3 ± 6.5	10.1 ± 5.7	F = 6.01, P = 0.0002, η2 = 0.08
SPARE-AD index (Mean ± Std)	−1.09 ± 0.85	−1.31 ± 0.73	−1.75 ± 0.76	−1.83 ± 0.74	F = 62.8, P < 0.001, η2 = 0.15
WMH volume (Mean ± Std)	3140.29 ± 4913.23	2316.5 ± 6158.1	1506.5 ± 2000.1	1418.2 ± 2168.5	F = 6.15, P = 0.0004, η2 = 0.02
ICVF
Tapetum (Mean ± Std)	0.46 ± 0.09	0.484 ± 0.05	0.492 ± 0.06	0495 ± 0.06	F = 5.06, P = 0.002, η2 = 0.04
Sup Longitudinal Fasciculus (Mean ± Std)	0.62 ± 0.04	0.64 ± 0.03	0.643 ± 0.02	0.65 ± 0.03	F = 6.52, P = 0.0003, η2 = 0.06
Posterior Thalamic Radiation (Mean ± Std)	0.532 ± 0.04	0.545 ± 0.04	0.547 ± 0.03	0.546 ± 0.03	F = 2.54, P = 0.06, η2 = 0.02
ISOVF
Tapetum (Mean ± Std)	0.227 ± 0.086	0.166 ± 0.068	0.1625 ± 0.0549	0.158 ± 0.058	F = 20.24, P < 0.0001, η2 = 0.16
Sup Longitudinal Fasciculus (Mean ± Std)	0.082 ± 0.011	0.0789 ± 0.012	0.081 ± 0.011	0.079 ± 0.015	F = 1.13, P = 0.03, η2 = 0.01
Posterior Thalamic Radiation (Mean ± Std)	0.089 ± 0.016	0.0877 ± 0.016	0.0889 ± 0.014	0.0879 ± 0.016	F = 0.12, P = 0.95, η2 = 0.001
OD
Tapetum (Mean ± Std)	0.067 ± 0.017	0.068 ± 0.014	0.072 ± 0.015	0.07 ± 0.01	F = 1.41, P = 0.24, η2 = 0.01
Sup Longitudinal Fasciculus (Mean ± Std)	0.137 ± 0.015	0.14 ± 0.01	0.140 ± 0.01	0.138 ± 0.013	F = 0.90, P = 0.44, η2 = 0.01
Posterior Thalamic Radiation (Mean ± Std)	0.084 ± 0.012	0.081 ± 0.011	0.074 ± 0.008	0.077 ± 0.008	F = 13.2, P = 0.06, η2 = 0.11

Std: standard deviation; TMT-A/B: trail making task A/B; DSST: digit symbol substitution task; MMSE: mini mental state examination; GM: gray matter; SPARE-AD: Spatial Patterns of Abnormality for Recognition of Early Alzheimer’s Disease; WMH: white matter hyperintensity; ICVF: intra-cellular volume fraction; ISOVF: isotropic or free water volume fraction; OD: orientation dispersion index; Sup: superior.

Advanced versus resilient group

Imaging profile

Even though the proportion of individuals with two APOE4 alleles was no different across the groups (χ² (4) = 0.41, P = 0.98), mean whole cortex centiloid values were higher in the advanced group (P = 0.0004) compared to the resilient group suggesting higher amyloid burden in these individuals (Fig. 5a). Whole brain WMH volume (P = 0.0007, Fig. 5b) and right tapetum ISOVF (P = 1.3 × 10⁻¹⁰, Supplementary Fig. S5a) were elevated in the advanced versus resilient group, whilst ICVF was reduced (right tapetum: P = 0.007 Supplementary Fig. S5b, right superior longitudinal fasciculus: P = 9.5 × 10⁻⁵ Supplementary Fig. S5d) indicating increased atrophy and lower neurite density. The orientation dispersion index was increased in the right posterior thalamic radiation in the advanced group (P = 2.1 × 10⁻⁵, Supplementary Fig. S5i) indicating reduced fiber coherence in this tract. Lastly, advanced agers had significantly higher SPARE-AD index compared to resilient agers (P = 9.0 × 10⁻¹⁴, Fig. 5c) indicating increased Alzheimer’s disease-like patterns of atrophy.

Fig. 5.

Comparison of neuroimaging features across brain-age groups. (a) Whole cortex centiloid values. (b) White matter lesion volume. (c). SPARE-AD index. The brackets indicate the significant (P < 0.05) pairwise P values between the brain-age groups following Tukey’s HSD post-hoc test.

Cognitive profile

Compared to the resilient group, the advanced group took significantly longer to complete both parts of the TMT at baseline (part A: P = 0.03, Fig. 3a; part B: P = 0.0003, Fig. 3b). This TMT result worsened on follow-up testing as indicated by the steeper slope of longitudinal change in the advanced group (part A: P = 1.4 × 10⁻⁵, Fig. 4a; part B: P = 2.6 × 10⁻⁹, Fig. 4b). Mean MMSE score was significantly lower than that of their resilient counterparts at baseline (P = 0.03, Fig. 3d). Performance on the DSST was no different at baseline (Fig. 3c) but longitudinally, DSST score decreased significantly faster in the advanced compared to the resilient group (P = 0.0001, Fig. 4d).

Fig. 3.

Baseline cognitive performance across brain-age groups. (a) Mean time (seconds) to complete trail making test part A (n = 493, Advanced = 165, Resilient = 147, AFRS = 91, RFAS = 90). (b) Mean time (seconds) to complete trail making test part B (n = 484, Advanced = 162, Resilient = 147, AFRS = 89, RFAS = 96). (c) Z score standardised total DSST score (n = 368, Advanced = 124, Resilient = 109, AFRS = 65, RFAS = 70). (d) Z score standardised total MMSE score (n = 342, Advanced = 118, Resilient = 108, AFRS = 55, RFAS = 61). The brackets indicate the significant (P < 0.05) pairwise P values between the brain-age groups following Tukey’s HSD post-hoc test.

Fig. 4.

Longitudinal changes in cognitive performance across brain-age groups. The Y axis on all plots represents the value of the gradient (the rate of change) between the mean baseline score of each different task and the follow-up score for that task. (a) The gradient measures the difference between the mean baseline time to complete part A of the trail making test and the follow-up time to complete the test (n = 1160, Advanced = 324, Resilient = 371, AFRS = 228, RFAS = 237). A positive gradient value represents an increase in the time taken to complete the task at follow-up, which means a worsening over time. (b) The same as panel (A) but for the trail making test part B (n = 1109, Advanced = 309, Resilient = 365, AFRS = 216, RFAS = 219). (c, d) A negative gradient value for the DSST (n = 664, Advanced = 201, Resilient = 212, AFRS = 108, RFAS = 143) and MMSE (n = 1149, Advanced = 315, Resilient = 387, AFRS = 207, RFAS = 240), respectively, represents worsening performance over time. The brackets indicate the significant (P < 0.05) pairwise P values between the brain-age groups following Tukey’s HSD post-hoc test.

Genetic profile

One genetic variant (rs58920042; position: 3:71981089 (GRCh37); minor/effect allele: T; mapped gene: LINC00877) was significantly associated with advanced (N = 106) versus resilient (N = 89) agers (odds ratio = 4.35; SE = 0.47; P = 6.9 × 10⁻⁶, Supplementary Fig. S6). This genetic variant has been associated with protection against early onset AD.⁴⁰

RFAS and AFRS groups

Imaging profile

Mean whole cortex centiloid values in both groups were comparable to that of the resilient group (P > 0.05, Fig. 5c). However, in order to check whether the relation between centiloid and baseline age differed across the groups, we compared the slopes of centiloid with age and found that the RFAS group had a significantly different intercept and slope compared to the resilient group (β = −0.34, P = 0.04, Supplementary Fig. S7). This finding showed that the RFAS group had higher amyloid levels at younger ages and less increase in amyloid burden with increasing age, indicating an earlier onset of amyloid accumulation for the RFAS group. Only the RFAS group had significantly higher SPARE-AD scores than the resilient group (P = 1.2 × 10⁻¹², Fig. 5a). There were no other significant differences in imaging modalities in these groups.

Cognitive profile

There were no group differences in baseline cognitive measures in either group (Fig. 3a–d). However, the AFRS group had the steepest decline in MMSE scores over time (P = 0.0002, Fig. 4d). AFRS performance on the TMT-B also worsened over time (P = 0.01, Fig. 4b) whilst performance on the DSST improved over time in this group (P < 0.00001, Fig. 4c). The RFAS group is characterised by a different set of cognitive changes over time, whereby the slight negative slope for the TMT-A indicates that performance on this task improved slightly compared to the resilient group (P = 7.1 × 10⁻¹¹, Fig. 4a) but performance in the DSST worsened the most (P < 0.00001, Fig. 4c). Contrary to the AFRS group, performance on the TMT-B (Fig. 4b) and MMSE (Fig. 4d) stayed consistent with the resilient group.

Conversion to MCI

The Kaplan–Meier curves (Supplementary Fig. S8) show that the number of total events was low in each group with zero events in the resilient group. Individuals in the RSAF group had a significantly higher risk of converting to MCI (P = 0.015) relative to the resilient group.

Discussion

The current study expands upon previous work that has examined variations in brain-age in cognitively healthy populations by calculating brain-age in 3460 individuals from a diverse and harmonised cohort of multi-site studies using structural and functional MRI data. We investigated the factors that differentiate advanced agers from people with more youthful brains. Our primary finding was that individuals with advanced brain-age had more features indicative of neurodegeneration and poorer cognition. Additional major findings include: 1) whilst the RFAS and AFRS groups had signs of advanced ageing in one modality, they only differed from resilient agers in showing greater declines in cognitive functions; 2) the resilient group has a genetic variant that has been associated with protection against early onset AD.⁴⁰

Our results suggest that the presence of both structural and functional deficits may be most predictive of poor long term outcomes. This echoes studies that have obtained higher accuracies in predicting conversion from MCI to AD using multimodal data rather than single modality studies.⁴¹

The advanced brain-age group displayed evidence of increased atrophy across multiple measurements: higher SPARE-AD index which reflects AD-like atrophy patterns, increased ISOVF which measures the free diffusion of extra-cellular water, increased white matter lesion volume, reduced neurite density and increased fiber coherence which has been associated with ageing.⁴² Common structural changes in MCI include a yearly loss of 0.46% of whole brain volume⁴³ which becomes more restricted to regions of the limbic system as the disease progresses to AD.⁴⁴ WMH elevate the risk for AD by 25% and begin to accelerate in volume around 10 years before MCI onset.^45,46 Therefore, the presence of atrophy and WMH in the advanced group is in line with biomarkers of preclinical AD.

Our results are also in line with previous findings of decreased neurite density and increased isotropic volume fraction in MCI and AD dementia.⁴⁷ More specifically, in both AD dementia and unimpaired individuals, low neurite density was observed in the presence of amyloid and tau pathology.⁴⁸ We only found reduced neurite density in the advanced group who also had higher amyloid levels. Therefore, although neurite density decreases with age,⁴⁹ our findings may highlight a group of people that are at higher risk of developing AD due to the co-occurrence of higher amyloid burden. A previous study reported that the rate of preclinical AD among cognitively normal elderly individuals is around 31%.⁵⁰ Similarly, although amyloid positivity increases with age in persons without dementia, it is consistently higher in cognitively normal individuals and MCI who are APOE-ε4 carriers.⁵¹ Therefore, our results indicate that the advanced group not only shows structural and functional ageing beyond that expected due to normal ageing, but also has other pathologies associated with AD.

Furthermore, our results are in line with our previous work showing an association between the presence of WMH, advanced brain ageing, AD-like atrophy and poor cognition.⁹ Our sample of advanced agers had poor baseline cognition across all tasks and showed declines in performance over time. Preclinical AD may be associated with declines in multiple cognitive domains. The most affected cognitive domains are episodic memory, executive functioning and perceptual speed.⁵²

A positive association was found between the SNP in chromosome 3 (rs58920042) and age of onset of AD in a previous GWAS⁴⁰ suggesting that this SNP (minor/effect allele: T; beta = 4.931; P = 8 × 10⁻⁷) is protective against brain ageing and AD. The resilient group had a higher proportion of the T alleles compared to the advanced group. This corroborates a close relationship between ageing and AD – advanced brain ageing is associated with substantial functional and structural changes that can be linked to AD.⁵³ It has been hypothesised that discovering protective genetic variants, instead of risk variants, is likely to uncover novel biological mechanisms of AD and to provide future gene-guided therapeutics. The identified SNP and the mapped gene deepen our understanding of brain resilience against ageing and neurodegeneration like AD; future analyses at biological pathway or protein levels (e.g. the REST protein) are required to fully elucidate the underlying mechanisms.

The AFRS group was no different from the resilient group for baseline cognition but had the worst decline in MMSE scores compared to the other groups. On the other hand, MMSE scores in the RFAS group were normal at baseline and remained stable longitudinally. In sum, the group with evidence of structural abnormalities remained more cognitively stable over time whilst the group with functional deterioration showed a steeper cognitive decline. According to a recent study, the annual decline in MMSE scores is higher in MCI progressors compared to stable patients with MCI and can predict progression to dementia with 82% accuracy.⁵⁴ Our results are in agreement with this pattern of findings because the AFRS group has the steepest decline in MMSE scores and the highest risk of conversion to MCI.

Furthermore, extensive literature suggests that disrupted functional connectivity patterns appear early in the course of AD development and in normal ageing,⁵⁵ which may explain why RFAS individuals are less likely to convert to MCI or preclinical AD. It remains unclear whether the resilient-like functional connectivity within the RFAS group reflects compensatory response for or was present before the structural decline. However, older adults with superior memory abilities have preserved cortical structural integrity in core regions of the DMN and salience networks that are indistinguishable from that of young adults,⁵⁶ indicating that the abilities of the RFAS group are more likely to be compensatory but it is unclear how long the beneficial effect will last.

Limitations

A strength of our study is that we calculated separate brain-ages for each modality, which has the benefit of being able to identify advanced ageing in each modality separately instead of relying on the algorithm to pick out the most informative features from each modality and then combine them. Our MAE values are comparable to studies using multimodal approaches.⁵⁷ Since our models were trained on data from a large sample of cognitively normal participants with a wide age-range, we can be confident that the predicted age and delta actually reflect deviations from normal brain ageing. However, we were limited by a select few cognitive domains. Pooling data from multiple studies allowed us to have a larger initial sample size and to aggregate a variety of variables. However, a limitation of this is that not all variables are collected in each study which results in unequal and sometimes small subsets in downstream analyses. However, the power calculations suggest that we had sufficiently large samples to detect the effects of interest. While we note that some of the relationships we identified have small effect sizes or were identified in single sites (e.g., genetic analyses were restricted to the UK Biobank) and that these require caution regarding expectations of clinical applications, small effects can still provide critical clues to underlying pathophysiology. Another limitation of our study is the absence of longitudinal outcome data. This prevents us from drawing any firm conclusions about whether conversion to AD is more likely in one brain-age group over another. Continued acquisition of longitudinal neuroimaging data in many of these studies will allow future investigation of rates of change across the measures of atrophy and will provide insight into how the groups differ over time.

Conclusion

In a sample of healthy individuals, we showed that different combinations of advanced and resilient structural and functional brain-ages are differentially characterised by AD-like brain features, genetics and cognitive outcomes.

Contributors

All authors read and approved the final version of the manuscript. M.A. wrote the manuscript and helped create the tables. D.S. processed data, performed all analyses, and created the figures and tables. J.W. designed all genetic analyses and supervised J.C. who performed them. G.E. provided ideas with modelling/processing and supervised. A.A. provided ideas for brain age model and integrated code for functional processing. E.Z. and R.M. aggregated the data and performed harmonisation for structural data. G.H. helped with structural brain age calculation.

S.T.G. provided ideas with refinement. A.A.C. scripted harmonisation of functional data. Z.Z. provided ideas for functional harmonisation. Z.Y. involved in structural brain age modelling. R.P. wrote the code for neuroHarmonize. S.S. provided ideas for the analysis of cognitive variables. M.B. processed amyloid variables for OASIS data. Manuscript revision and submission approval: P.L, Y.A, A.S, T.B, L.B.H, D.S.M, K.Y, L.J.L, J.C.M, D. T, L.F, N.R.B, S.M.R.

A.S. integrated code for functional processing. M.H. supervised the project. D.W. supervised and provided ideas. Y.F. supervised and provided ideas for functional MRI processing. H.S and I.M.N. supervised statistical and other analyses and provided critical review. C.D. provided conceptualisation, funding, supervision, and writing (review and editing).

Data sharing statement

Data presented in this manuscript is available upon reasonable request. Data from the UKBB are available upon request from the UKBB website (https://www.ukbiobank.ac.uk/). Data from the BLSA study are available upon request at https://www.blsa.nih.gov/how-apply. Data from the OASIS study are available upon request at https://www.oasis-brains.org/.

Declaration of interests

TB has received investigator-initiated research awards from the NIH, the Alzheimer’s Association, the Foundation at Barnes-Jewish Hospital, Siemens Healthineers, Hyperfine and Avid Radiopharmaceuticals (a wholly-owned subsidiary of Eli Lilly and Company). She participates as a site investigator in clinical trials sponsored by Eli Lilly and Company, Biogen, Eisai, Jaansen, and Roche. She has served as a paid and unpaid consultant to Eisai, Siemens, Biogen, Janssen, Hyperfine, Merck Lilly, and Bristol-Myers Squibb.

JCM has served as a paid consultant to the Barcelona Brain Research Center and the Native Alzheimer Disease-related Resource Center in Minority Aging Research. He also received payments for presentations at the AAIM meeting, Longer Life Foundation and the International Brain Health Symposium. JCM has received travel support to attend meetings including: AAIM, DIAN, AD/PD, ATRI/ADNI, ADRC, ADC, the International conference on Health Aging & Biomarkers and the International Brain Health Symposium. He has served on the advisory board for the Cure Alzheimer’s Fund and LEADS at Indiana University.

IMN has received payments from Premier, Inc for participating in an advisory board, from Peerview for an educational talk, and from Subtle Medical, Inc for consulting.

DW has served as a paid consultant to Qynapse, Beckman Coulter and Eli Lilly. He also received grants from the NIH and Biogen paid to his institution and received travel support from the Alzheimer's Association.

SR is an NIA IRP employee and has served on the advisory board of Dementia Platforms, UK, the Canadian Consortium on Neurodegeneration in Aging and the Adult Aging Brain Connectome. She has received travel support from the McKnight Foundation to attend an annual meeting.