https://orcid.org/0000-0003-0396-1707
Abstract
The nondemented old–old over the age of 80 comprise a rapidly increasing population group; they can be regarded as exemplars of successful aging. However, our current understanding of successful aging in advanced age and its neural underpinnings is limited. In this study, we measured the microstructural and network-based topological properties of brain white matter using diffusion-weighted imaging scans of 419 community-dwelling nondemented older participants. The participants were further divided into 230 young–old (between 72 and 79, mean = 76.25 ± 2.00) and 219 old–old (between 80 and 92, mean = 83.98 ± 2.97). Results showed that white matter connectivity in microstructure and brain networks significantly declined with increased age and that the declined rates were faster in the old–old compared with young–old. Mediation models indicated that cognitive decline was in part through the age effect on the white matter connectivity in the old–old but not in the young–old. Machine learning predictive models further supported the crucial role of declines in white matter connectivity as a neural substrate of cognitive aging in the nondemented older population. Our findings shed new light on white matter connectivity in the nondemented aging brains and may contribute to uncovering the neural substrates of successful brain aging.
Introduction
Aging has a deleterious effect on human brain structure and cognitive functioning (Koen and Rugg 2019). Lifespan trajectory studies indicate that white matter undergoes a complex age-related deterioration in old age, where estimated declines accelerate after the relatively stable adult period (Westlye et al. 2010; Fjell et al. 2021). For example, white matter volumes peak and then plateau in the late adulthood, followed by an accelerated decline (Westlye et al. 2010). However, previous studies on white matter tractography in the aging brain have mainly focused on the young–old, which raised question that whether the pattern of white matter connectivity due to aging observed in the young–old can be applicable to the nondemented old–old over the age of 80, even in the much less-studied oldest–old and centenarian.
The nondemented old–old comprise a rapidly increasing population group. Variously defined as aged 80, 85, or 90 years and over (Mengel-From et al. 2013; Spaniolas et al. 2014; Silverman and Schmeidler 2018), the nondemented old–old over the age of 80 can be regarded as exemplars of successful aging, having avoided neurodegenerative disease and maintained normal cognitive functioning (Brodaty et al. 2016; Neltner et al. 2016). There have been several studies on the complex cerebral aging processes in the much less-studied nondemented old–old and oldest–old. Yang et al. (2016) directly compared the over 90 years old group with the young–old and reported a differential pattern of global and regional atrophy with increased age, with the greatest age effects seen in the medial temporal lobe and parietal and occipital cortices. Tang et al. (2021) recently further validated that most of the sulci significantly widened with increased age and that the rates of sulcal widening were lower in the old–old compared with young–old. These studies focused on the alterations in the cerebral cortex in the old–old and oldest–old and seemed to convey a perspective that successful cognitive aging in this population may stem from a slower age-related deterioration in the cerebral cortex. However, there still remains one important gap that should be filled: whether and how the age-related decline in white matter connectivity contributes to cognitive aging in nondemented old–old. Bennett et al. (2017) reported age-related differences in the white matter integrity of the old–old and oldest–old, but their study did not focus on the changes between the old–old and young–old samples, nor revealed the delicate relationship between cognitive aging and white matter degeneration in the nondemented old–old. Therefore, investigating this will have important significance for uncovering the neural underpinning of successful cognitive aging in this population.
During the aging process, white matter undergoes a series of neuropathological changes at both macroscopic and microstructural levels, including atrophy, demyelination, neuronal degeneration, expansion of perivascular spaces, and proliferation of glial cells (Westlye et al. 2010). At the microstructural level, numerous diffusion weighted imaging (DWI) based studies have reported an age-related decline of microstructural integrity using metrics such as fractional anisotropy (FA) and mean diffusivity (MD) (Jolly et al. 2017; Li et al. 2020). Histopathologic and morphological evidence has highlighted demyelinated fibers, glial swelling, bulk-volume loss of white matter, gross gray matter atrophy, and ventricular dilation in brain aging (Qin et al. 2013; Zerbi et al. 2013), which can result in the microstructural abnormalities in the orientation of white matter fibers (Zhao et al. 2021a; Zhao et al. 2021b). Quantitatively measuring the microstructural changes including integrity and orientation in nondemented old–old may provide potential evidence for neuropathological changes in brain aging that have not been explored.
From the network point of view, structural connectivity of white matter shapes the essential connectivity between cerebral regions (Straathof et al. 2018). The age-related deterioration of white matter connectivity can contribute to the disconnection of neural networks and subsequently lead to the decline and malfunction in some aspects of cognitive abilities observed in aging (Hirsiger et al. 2016). However, little is known about the age-related changes in structural connectivity of white matter in the nondemented old–old. Moreover, one important unresolved issue is the underlying relationship between these age-related deteriorations of structural connectivity and cognition in this population. Most identified aspects of network–cognition relationship were detected by weak correlational analysis rather than the more robust and unbiased data-driven machine-learning predictive approach. The machine-learning predictive model has been an effective and robust tool for uncovering which biomarkers are most related to the aging process (Fabris et al. 2017).
The aim of this study was to verify the hypothesis that nondemented old–old had a differential pattern of white matter connectivity from the young–old and to investigate the underlying neural substrate of cognitive aging. We recruited 419 community-dwelling nondemented older participants who were then divided into 2 age groups—the young–old, aged 72–79 years (n = 230, mean = 76.25 ± 2.00, 127 males), and the old–old, aged 80–92 years (n = 219, mean = 83.98 ± 2.97, 108 males), and completed a longitudinal follow-up cognition in 4 years. At the microstructural level, we applied a novel mathematical framework called Director Field Analysis to quantify the orientational properties and integrity of white matter fibers (Cheng and Basser 2018). At the network level, we focused on the topological properties of structural brain networks (SCNs) including the network-based measures such as small world properties, modular organization, and degree.
Methods
Participants
Participants were drawn from the Sydney Memory and Ageing Study, a longitudinal study of community-dwelling nondemented older adults aged 70–90 years at baseline, recruited randomly via the electoral roll from 2 areas of Sydney, Australia (Sachdev et al. 2010). Participants were excluded at baseline if they were diagnosed with dementia or any other major brain disease, scored < 24/30 on the Mini–Mental State Examination, or were unable to complete neuropsychological tests because of inadequate English skills. A total of 419 nondemented participants who had completed magnetic resonance imaging (MRI) scans at wave 2 were included in this study. We separated these participants into 2 age groups—the young–old, aged 72–79 years (n = 230, mean = 76.25 ± 2.00), and the old–old, aged 80–92 years (n = 219, mean = 83.98 ± 2.97). There were no significant differences between the 2 age groups in sex or years of education. All participants gave written informed consent. Ethics approval was obtained from the Human Research Ethics Committees of the University of New South Wales and the South Eastern Sydney and Illawarra Area Health Service.
White matter microstructure
Diffusion tensor imaging (DTI) image preprocessing was conducted using FSL (www.fmrib.ox.ac.uk/fsl), ANTs (http://stnava.github.io/ANTs/), and DMRITool box (http://diffusionmritool.github.io). Briefly, we used FSL for Eddy-current correction and head movement. BET was then used to remove nonbrain tissue, and the extracted brain images were visually inspected. Corrected b0 images were linearly aligned to skull-stripped T1 images using FSL linear registration with boundary-based registration and then nonlinearly registered to the T1-weighted images with ANTs to correct for echo planar imaging (EPI) induced susceptibility artifacts. Finally, the resultant 3D deformation fields were applied to the remaining diffusion data. We used DMRITool to reconstruct diffusion tensors and calculate the DWI indices reflecting the integrity (i.e. FA and MD) and orientation (i.e. splay, bend, twist, and total distortion, Fig. 1a and b) of white matter microstructure for each subject. For the orientational indices, detailed interpretations have been described by Cheng and Basser (Cheng and Basser 2018) and Supplementary Text S3.
Tract-based spatial statistics (TBSS) was firstly performed. Mean FA maps were created and thresholded to obtain a projection of all participants’ FA data onto a mean FA skeleton that represented the centers of all tracts common to the group. Briefly, all participants’ FA data were nonlinearly aligned to a standard template space (FMRIB58_FA) with FNIRT. Then, the mean FA image was created and thresholded to create the mean FA skeleton. Next, each participant’s FA data were projected onto the thresholded mean FA skeleton. TBSS was also performed for the other 5 indices (i.e. MD, splay, bend, twist, and total distortion). The displacement fields of the nonlinear registration estimated from the FA images were applied to warp the other index maps before applying voxel-wise statistics.
Tract of interest analysis was further performed. The obtained white matter skeletons were categorized into following white matter fiber tracts considered as essential underpinnings of cognitive functioning; these tracts were identified with Johns Hopkins University’s white matter atlas available in FSL and included in this analysis: commissural bundles of the body of the corpus callosum (CCb), forceps minor, and forceps major; projection fiber tracts of bilateral anterior thalamic radiations (ATR) and corticospinal tracts (CST); association tracts of bilateral uncinate fasciculus ((UF), bilateral inferior longitudinal fasciculus (ILF), bilateral superior longitudinal fasciculus (SLF), bilateral anterior cingulum (CG), bilateral inferior fronto-occipital fasciculus (IFOF) (Fig. 1c). Finally, the signals within each fiber tracts were averaged and extracted into the further statistical analysis.
Structural connectivity network
The SCN construction and white matter fiber tracking were conducted using DSI Studio (http://dsi-studio.labsolver.org/). Specifically, the diffusion data were reconstructed using q-space diffeomorphic reconstruction (QSDR) methods that calculated the orientational distribution of the density of diffusing water in MNI space. QSDR first reconstructed diffusion-weighted images in native space and computed the quantitative anisotropy (QA) in each voxel. These QA values were used to warp the brain to a template QA volume in MNI space using a nonlinear registration algorithm implemented in the statistical parametric mapping software (Luppi and Stamatakis 2021). A diffusion sampling length ratio of 1.25 was used, and the output resolution was 1 mm.
Fiber tracking was performed on whole-brain with 10,000 seeding, using a QA-assisted deterministic fiber tracking algorithm (Yeh et al. 2013) with these parameters: angular threshold of 45 degrees, anisotropy threshold automatically determined by DSI Studio, step size of 0.5 voxels, and track length from 15 to 150 mm. Finally, brain networks were constructed at the macroscale with 90 nodes as defined by the automated anatomical labeling atlas (AAL-90, Fig. 1d and Table S1) and the mean FA-weighted, undirected connections between the nodes.
To systematically analyze the topological properties in SCN, we calculated the global network attributes of SCN, including the global efficiency (Eg), local efficiency (Eloc), small-world properties such as clustering coefficient Cp, characteristic path length Lp, normalized clustering coefficient Gamma, normalized characteristic path length Lambda, and small-world coefficient Sigma. We also calculated subnetwork attributes (i.e. modular connectivity) and nodal properties (i.e. degree centrality and betweenness centrality). In addition, to quantitatively measure the modular connectivity within the SCN, we identified 5 brain networks including somatosensory–motor network (SMN), visual network (VN), salience network (SAN), executive control network (ECN), and default mode network (DMN) based on previous atlas/studies (Power et al. 2011; Yeo et al. 2011). Based on the modular structures, the intramodule connection was calculated as the sum of total edges within a module, and the intra-module strength was calculated as the average strength of all connections within a module. The intermodule connection was calculated as the sum of total edges between any pair of 2 modules, and the intermodule strength was calculated as the average strength of all connections between any pair of 2 modules. For detailed definitions of network metrics, see Supplementary Text S4.
Statistical analysis
All calculations except voxel-wise statistical analysis were performed by using MATLAB R2016a. Firstly, 2 unpaired T-tests were applied to investigate the differences in microstructural and network-based properties between the young–old, and old–old while controlling for the sex and years of education. All outcomes were corrected for multiple comparisons using false discovery rate (FDR). For the voxel-wise statistical analysis, a general linear model was built with group factor as independent factors, demeaned sex and years of education as nuisance covariates. Between-groups differences in white matter skeleton obtained in TBSS were tested using permutation-based nonparametric inference on statistic maps with 10,000 random permutations. Resultant maps were then thresholded using threshold-free cluster enhancement with P < 0.05 corrected for multiple comparisons.
Linear regressions were used to further interrogate the effect of age on microstructural/network-based topological metrics for whole sample and each age group. To allow for meaningful comparisons between within-group alteration rates per year (i.e. slope) regarding different microstructural/network-based metrics, the original regression coefficients were normalized by dividing them using the baseline value (i.e. the intercept of regression line) of each metrics. For between-group comparisons, slope differences were tested in 2 steps: a permutation test was used to evaluate the association between each metrics and age, 1,000 permutations were repeatedly conducted for young–old and old–old, respectively. Then, an independent sample t-test was applied to analyze differences in permutation results between groups for each metrics separately, adjusting for sex and years of education. Outcomes were corrected for multiple tests using FDR.
For the purposes of clarifying the neural mechanisms underlying age-related cognitive changes in old age, potential relationships between white matter connectivity in microstructure/structural networks and global cognition were firstly examined using partial correlation analysis for whole-sample and each age group, after adjusting for the confounding effects of sex and years of education. Mediation analyses were then conducted to estimate whether and how white matter connectivity in microstructure and brain networks mediated the effect of age on global cognition for the whole sample and each age group. Direct and indirect effects between variables have been estimated while adjusting for sex and years of education. In addition, we proposed an overall index to represent the global white matter connectivity in microstructure and structural networks by summing weighted values of connection strength selected from the above linear regressions for each white matter tract or network. These summary statistics can be represented as follows:
- (1) microstructural connectivity:$$ {SF}_i=\frac{1}{N}\sum_{j\in G}\left({F}_{ij}\ast{\beta}_j\right) $$
- (2) network-based connectivity:$$ {SSC}_i=\frac{1}{N}\sum_{j\in G}\left({S}_{ij}\ast{\beta}_j\right) $$
where SFi, SSCi are the global microstructural connectivity and global network-based connectivity for participant i; Fij, Sij are the average FA value of white matter tract j for participant i, and average strength of all connections within network j for participant i, respectively; βj is the normalized regression coefficient for white matter tract j or network j acquired in the above linear regressions.
Methodological illustration. a) An illustration of three types of orientational distortions (i.e. splay, bend, and twist). b) Each row shows a synthetic tensor field with the 6 index maps (splay, bend, twist, total distortion, FA and MD) calculated from the tensor field. All tensors in the tensor fields share the same shape but have different orientations, resulting in constant FA and MD maps. c) All tracts-of-interest included in the further analysis. Ten white matter tracts were selected, including ATR, CG, CST, forceps major/minor, UF, ILF and SLF, and IFOF. d) AAL-90 atlas.
Prediction of cognition
To investigate the predictive effect of the white matter connectivity declines observed at the wave 2 for global cognition four years later at the wave 4, we chose an epsilon supported vector regression (epsilon-SVR) with linear kernel LIBSVM to build regression prediction models by using microstructural and/or network-based metrics, respectively. A 10-fold cross validation framework was applied to evaluate the performance of training features. Before the model was built, we regressed out subjects’ sex and years of education from the training set, then scaled the resultant residuals of training dataset. The acquired parameters were also applied for scaling the test dataset. A random seed was predefined as 1 to minimize the impact of the randomness in the 10-folds algorithm. Pearson correlation between the actual scores and the predictive scores was used to evaluate the predictive effect of model. Permutation tests were applied to determine the validity and prediction accuracy of established models.
Reproducibility analyses
To evaluate the potential confounding effect including different parcellation schemes and cognitive domain on the results, we performed reproducibility analyses with a different brain parcellation scheme and cognitive domains. Detailed descriptions can be found in Supplementary Text S5.
Results
Age effects on cognitive function
Demographic information and neuropsychological characteristics of each group are summarized in Table 1. The cognitive performances at the wave 2 and in the 4-year follow-up for the young–old and old–old are presented in Table S2. As expected, the old–old had worse cognitive performance in all the domains including attention/process speed, language, executive function, visuospatial function, memory, and verbal memory (P < 0.05, FDR corrected). In addition, 2-way repeated measured analysis of variance analysis showed a significant interaction effect between group (young–old vs. old–old) and time (wave 2 vs. 4) in the attention/process speed, executive function, and global cognition, which indicated that old–old had an accelerated cognitive decline in these cognitive domain and global cognition compared with the young–old (Fig. S1 and Table S3).
Age effects on white matter microstructure
Compared with young–old, the old–old had extensive age-related deterioration in microstructural integrity (i.e. FA, MD). The changes in the white matter fiber orientation were mainly observed in the forceps minors, forceps major, and CCb (Fig. 2a). Further linear regression on tracts of interest verified that commissural bundles had the largest magnitude of age-related degradation in microstructural integrity compared with projection and association fibers. On the other hand, only commissural fibers (i.e. forceps minor and body of callosum corpus) demonstrated significant declines in orientation with increased age. The forceps minor showed most prominent age-related degeneration in both orientation and integrity of microstructure (Fig. 2b).
Changes in white matter microstructure in nondemented old–old. a) Age-related deterioration in the orientational properties (i.e. distortion, splay, bend, and twist) and integrity (i.e. FA and MD) of white matter microstructure in the old–old compared with young–old. b) Linear regression on tracts of interest of white matter fibers in nondemented older. Light dots in left panel represent the microstructural metrics which reached significant level after multiple comparison correction, P < 0.05, FDR adjusted. Dotted line in right panel indicate the average values of microstructural indices (e.g. FA values) at each age. c) Estimated rates of age-related microstructural degeneration in white matter fibers (e.g. forceps minor) in the young–old and old–old.
In addition, rates of microstructural degeneration were estimated in both young–old, and old–old (Fig. 2c and Tables S4–S9). Scatter plots with the estimated line of the best fit between the microstructural measures and age are shown in Fig. 2c. Of note was that age-related microstructural degeneration in white matter fibers was relatively stable in the young–old, while an accelerating trend was observed in the old–old. It is worth noting that forceps minor had an accelerated deterioration in both orientation and integrity of microstructure in the old–old compared with those in young–old (Fig. 2c).
Age effects on structural connectivity network
Compared with young–old, the old–old had a significantly weaker structural connection strength (young–old = 0.294 ± 0.021; old–old = 0.287 ± 0.021; P < 0.001) of the SCN. For globally topological organization, the nondemented old–old are worse in the ability for global information communication, manifested as higher normalized characteristic path length, Lambda (P = 0.003, FDR adjusted, Fig. 3a). Our analysis on the modular organization showed that the old–old was lower in the intramodular strength within SMN, VN, and ECN (all P < 0.05, FDR adjusted) and in the intermodular strength between SMN–ECN, SMN–DMN, SAN–ECN, VN–ECN, VN–DMN, SAN–ECN, SAN–DMN, and ECN–DMN (all P < 0.05, FDR adjusted, Fig. 3a). Of the edges of SCN, compared with young–old, the old–old demonstrated lower connection strength between right supplementary motor area (SMA) and left olfactory cortex within SMN (P < 0.001, FDR adjusted, Fig. 3a). Further linear regression verified that nondemented older had significant declines in small-world properties (i.e. Gamma and Lambda), modular organization, and edges with increased age (all P < 0.05, FDR adjusted, Fig. 3b). In addition, nondemented older showed age-related decreases in the right SMA and orbital part of middle frontal gyrus, and age-related increase in the right lingual gyrus in the degree and betweenness centrality (all P < 0.05, FDR adjusted, Fig. 3b).
Changes in structural connectivity network in nondemented old–old. a) Age-related deterioration in topological organization of SCNs in the old–old compared with young–old. b) Linear regression on topological organization of SCNs in nondemented older. Light dots represent for the topological metrics, which reached significant level after multiple comparison correction, P < 0.05, FDR adjusted. Abbreviations: SC, structural connection. *P < 0.05, FDR adjusted; **P < 0.01, FDR adjusted; ***P < 0.001, FDR adjusted.
We analyzed the differences of slopes of estimated deterioration rate per year between young–old and old–old in structural connectivity. We observed that young–old kept a relatively stable trend in all topological metrics of SCN, whereas old–old exhibited an accelerating trend for age-related deterioration in the intramodule connection strength within SMN, VN, and DMN and in the intermodule connection strength between SMN–SAN, SMN–ECN, VN–ECN, VN–DMN, SAN–ECN, SAN–DMN, and ECN–DMN (all P < 0.05, FDR corrected, Fig. 4 and Tables S10–S14).
Clinical and demographic characteristics of young–old and old–old in wave 2.
| Young–Old | Old–Old | Old–Old vs. Young–Old | |||||
|---|---|---|---|---|---|---|---|
| Mean | SD | Mean | SD | t | P | Cohen’s d | |
| Sex (male/total, %) | 0.550 | / | 0.495 | / | −1.131 | 0.259 | −0.111 |
| Age (years) | 76.250 | 2.003 | 83.978 | 2.970 | 31.649 | <0.001a | 3.106 |
| Education (years) | 11.962 | 3.717 | 11.928 | 3.749 | −0.095 | 0.925 | −0.009 |
| MMSE Scoresb | 29.055 | 1.381 | 28.663 | 1.434 | −2.828 | 0.005a | −0.280 |
| Attention/Process Speed Scorec | 0.096 | 1.337 | −0.356 | 1.046 | −3.719 | <0.001a | −0.375 |
| Language Scorec | −0.065 | 1.029 | −0.488 | 1.096 | −4.045 | <0.001a | −0.402 |
| Executive Function Scorec | 0.009 | 1.148 | −0.369 | 1.271 | −3.067 | 0.002a | −0.316 |
| Visuospatial Function Scorec | 0.263 | 1.059 | −0.092 | 1.111 | −3.306 | 0.001a | −0.331 |
| Memory Scorec | 0.117 | 1.077 | −0.449 | 1.060 | −5.347 | <0.001a | −0.528 |
| Verbal Memory Scorec | 0.067 | 1.061 | −0.459 | 1.032 | −5.067 | <0.001a | −0.498 |
| Global Cognition Scorec | 0.097 | 1.163 | −0.469 | 1.121 | −4.987 | <0.001a | −0.497 |
| Young–Old | Old–Old | Old–Old vs. Young–Old | |||||
|---|---|---|---|---|---|---|---|
| Mean | SD | Mean | SD | t | P | Cohen’s d | |
| Sex (male/total, %) | 0.550 | / | 0.495 | / | −1.131 | 0.259 | −0.111 |
| Age (years) | 76.250 | 2.003 | 83.978 | 2.970 | 31.649 | <0.001a | 3.106 |
| Education (years) | 11.962 | 3.717 | 11.928 | 3.749 | −0.095 | 0.925 | −0.009 |
| MMSE Scoresb | 29.055 | 1.381 | 28.663 | 1.434 | −2.828 | 0.005a | −0.280 |
| Attention/Process Speed Scorec | 0.096 | 1.337 | −0.356 | 1.046 | −3.719 | <0.001a | −0.375 |
| Language Scorec | −0.065 | 1.029 | −0.488 | 1.096 | −4.045 | <0.001a | −0.402 |
| Executive Function Scorec | 0.009 | 1.148 | −0.369 | 1.271 | −3.067 | 0.002a | −0.316 |
| Visuospatial Function Scorec | 0.263 | 1.059 | −0.092 | 1.111 | −3.306 | 0.001a | −0.331 |
| Memory Scorec | 0.117 | 1.077 | −0.449 | 1.060 | −5.347 | <0.001a | −0.528 |
| Verbal Memory Scorec | 0.067 | 1.061 | −0.459 | 1.032 | −5.067 | <0.001a | −0.498 |
| Global Cognition Scorec | 0.097 | 1.163 | −0.469 | 1.121 | −4.987 | <0.001a | −0.497 |
Notes: SD, standard deviation; old–old vs. young–old, the scores in the old–old are higher than in the young–old. MMSE, Mini–Mental State Examination.
aP < 0.05, The results survived FDR correction appeared in bold.
bThe MMSE scores have been adjusted for age and years of education.
cThese domain scores were calculated by averaging the z-scores of component tests; these scores were subsequently transformed.
Clinical and demographic characteristics of young–old and old–old in wave 2.
| Young–Old | Old–Old | Old–Old vs. Young–Old | |||||
|---|---|---|---|---|---|---|---|
| Mean | SD | Mean | SD | t | P | Cohen’s d | |
| Sex (male/total, %) | 0.550 | / | 0.495 | / | −1.131 | 0.259 | −0.111 |
| Age (years) | 76.250 | 2.003 | 83.978 | 2.970 | 31.649 | <0.001a | 3.106 |
| Education (years) | 11.962 | 3.717 | 11.928 | 3.749 | −0.095 | 0.925 | −0.009 |
| MMSE Scoresb | 29.055 | 1.381 | 28.663 | 1.434 | −2.828 | 0.005a | −0.280 |
| Attention/Process Speed Scorec | 0.096 | 1.337 | −0.356 | 1.046 | −3.719 | <0.001a | −0.375 |
| Language Scorec | −0.065 | 1.029 | −0.488 | 1.096 | −4.045 | <0.001a | −0.402 |
| Executive Function Scorec | 0.009 | 1.148 | −0.369 | 1.271 | −3.067 | 0.002a | −0.316 |
| Visuospatial Function Scorec | 0.263 | 1.059 | −0.092 | 1.111 | −3.306 | 0.001a | −0.331 |
| Memory Scorec | 0.117 | 1.077 | −0.449 | 1.060 | −5.347 | <0.001a | −0.528 |
| Verbal Memory Scorec | 0.067 | 1.061 | −0.459 | 1.032 | −5.067 | <0.001a | −0.498 |
| Global Cognition Scorec | 0.097 | 1.163 | −0.469 | 1.121 | −4.987 | <0.001a | −0.497 |
| Young–Old | Old–Old | Old–Old vs. Young–Old | |||||
|---|---|---|---|---|---|---|---|
| Mean | SD | Mean | SD | t | P | Cohen’s d | |
| Sex (male/total, %) | 0.550 | / | 0.495 | / | −1.131 | 0.259 | −0.111 |
| Age (years) | 76.250 | 2.003 | 83.978 | 2.970 | 31.649 | <0.001a | 3.106 |
| Education (years) | 11.962 | 3.717 | 11.928 | 3.749 | −0.095 | 0.925 | −0.009 |
| MMSE Scoresb | 29.055 | 1.381 | 28.663 | 1.434 | −2.828 | 0.005a | −0.280 |
| Attention/Process Speed Scorec | 0.096 | 1.337 | −0.356 | 1.046 | −3.719 | <0.001a | −0.375 |
| Language Scorec | −0.065 | 1.029 | −0.488 | 1.096 | −4.045 | <0.001a | −0.402 |
| Executive Function Scorec | 0.009 | 1.148 | −0.369 | 1.271 | −3.067 | 0.002a | −0.316 |
| Visuospatial Function Scorec | 0.263 | 1.059 | −0.092 | 1.111 | −3.306 | 0.001a | −0.331 |
| Memory Scorec | 0.117 | 1.077 | −0.449 | 1.060 | −5.347 | <0.001a | −0.528 |
| Verbal Memory Scorec | 0.067 | 1.061 | −0.459 | 1.032 | −5.067 | <0.001a | −0.498 |
| Global Cognition Scorec | 0.097 | 1.163 | −0.469 | 1.121 | −4.987 | <0.001a | −0.497 |
Notes: SD, standard deviation; old–old vs. young–old, the scores in the old–old are higher than in the young–old. MMSE, Mini–Mental State Examination.
aP < 0.05, The results survived FDR correction appeared in bold.
bThe MMSE scores have been adjusted for age and years of education.
cThese domain scores were calculated by averaging the z-scores of component tests; these scores were subsequently transformed.
Estimated rates of age-related degeneration in modular organization of SCNs in the young–old, and old–old. Abbreviations: aP < 0.05, FDR adjusted.
White matter connectivity mediates the age effect on the cognition
There was a strong correlation between global cognition and white matter connectivity in both microstructure (all-sample: r = 0.229, P < 0.001; young–old: r = 0.163, P < 0.014; old–old: r = 0.218, P = 0.003) and structural network (all-sample: r = 0.226, P < 0.001; young–old: r = 0.146, P = 0.028; old–old: r = 0.237, P = 0.001). Further mediation models assessed how microstructural/network-based connectivity contributed to global cognition in all nondemented older, young–old, and old–old, respectively. Results indicated that white matter connectivity significantly mediated the effect of age on global cognition in all sample (mediation effect size: SF = 12.4%; SSC = 11.2%) and old–old (mediation effect size: SF = 16.5%; SSC = 18.4%); however, this mediation effect vanished in the young–old (Fig. 5a).
Relationship between age, white matter connectivity and cognition in nondemented older adults. a) Mediation effect of white matter connectivity on the relationship between chronological age and global cognition in all nondemented older, young–old, and old–old. Solid lines represent for the significant relationship between latent variables (P < 0.05, FDR adjusted); dot lines represent for nonsignificant relationship between latent variables. b) Prediction of longitudinal cognition from white matter connectivity based on the support vector regression models.
Prediction of cognition
We constructed global cognition predictive models based on the white matter connectivity. As shown in Fig. 5b and Table S15, results indicated that white matter connectivity showed good predictive ability for individual global cognition in 4 years at the wave 4 in nondemented older. Specifically, for microstructural metrics, the predicted scores were significantly correlated with actual global cognitive scores at the wave 4 (FA: r = 0.235, P < 0.001; mean absolute error, MAE = 1.052, P = 0.037; distortion: r = 0.164, P = 0.008; MAE = 1.060, P = 0.001). For SCNs, intra−/inter-modular connectivity showed good predictive effect for global cognition at the wave 4 (r = 0.198, P = 0.003; MAE = 1.081, P < 0.001). When combining all metrics, the predictive model had a substantively improvement in predictive ability (r = 0.314, P < 0.001; MAE = 1.02, P < 0.001). The results supported the validation of the acquired predictive models based on the age-related alterations in the white matter microstructure and brain networks to the global cognitive trajectory.
Reproducibility analyses
Reproducibility analyses revealed that the results were reproducible with a different brain parcellation scheme and cognitive domains. For details, see Figures S2–S6 and Supplementary text S5.
Discussion
This is a study to systematically investigate the degenerative profiles of white matter microstructure and brain networks in the young–old (<80 years old) and old–old (≥80 years old). Specifically, white matter connectivities in microstructure and brain networks were found to decline with increased age in both young–old and old–old. An accelerated rates of white matter connectivity in microstructure and networks were observed in the old–old compared with young–old. The white matter connectivity mediated the age effect on global cognition in the old–old but not the young–old and had notable predictive effect for global cognition in 4 years in the nondemented older. Of note, our older sample only included those who were willing and able to undergo an MRI scan, which means these older participants were considered to be physically healthier and slightly higher functioning compared with the general older population. This needs to be taken account in the interpretation of the patterns observed.
Age effects on white matter microstructure
Previous studies on age trajectory have reported that the onset of age-related decline in white matter microstructure begins in the fifth and sixth decade, with an acceleration of decline after 60 years (Sexton et al. 2014). In line with prior studies (Burzynska et al. 2010; Sexton et al. 2014), we observed that the old–old had extensive accelerated deterioration in white matter microstructure with increased age. The largest magnitude of age effect was found in the commissural bundles, especially for forceps minor. Forceps minor, as the most late-myelinating region in brain development, contains smaller axons with fewer myelin lamellae (Chia et al. 1983). In accordance with the last-in-first-out hypothesis (Raz 2000), it tended to be more vulnerable to neurodegenerative processes (e.g. demyelination). The degeneration of forceps minor may subserve the disconnection of structural and functional networks within prefrontal regions and contribute to cognitive decline in the nondemented older population (Sheline and Raichle 2013).
It is interesting to note that a relatively stable trend of the white matter microstructure was observed in the young–old. Different from lifespan studies with wide age range and limited sample size of older adults (Westlye et al. 2010), this study was strengthened by our large sample size with age ranges spanning from the young–old to the old–old. Recent evidence has also documented a similar nonlinear degradation of white matter microstructure in the older adult lifespan from the 7th to 10th decades (Merenstein et al. 2021). Of the lifespan of older adults, white matter microstructure remained relatively stable in the young–old (Merenstein et al. 2021). An alternative explanation of the observed stable pattern in the young–old may be the protracted development of white matter (Westlye et al. 2009). Developmental studies have highlighted asymptotic white matter maturation during midlife, followed by a relatively stable and slowly declining phase in late adulthood or young–old (Sexton et al. 2014; Slater et al. 2019). The observed differential pattern in white matter microstructure in the young–old and old–old is likely driven by different levels of pathologies (Auriel et al. 2014). Relative to young–old, the old–old are more susceptive to higher prevalence and more serious white matter diseases and other pathologies (e.g. microinfarcts and white matter hyperintensity) (Pereira et al. 2019).
Age effects on SCNs
Human brain is a highly complex system, optimized and balanced between network segregation and network integration (Cohen and Esposito 2016). During the process of normal aging in human brain, one of the most evident hallmarks is the age-related modulation of segregation and integration of brain networks (Koen et al. 2020). In line with prior findings (Li et al. 2020), our results revealed significant age-related degeneration in both network integration and network segregation, implying disruption in global topological organization of SCNs in the nondemented aging brain. Specifically, the observed increased normalized characteristic path length reflects the declined ability for interregional effective information integration in the brain networks in the old–old. Such growth of path length would reduce the efficiency of neural information transfer in global brain networks (He and Evans 2010). Second, modules are believed to support specific functioning, and they dynamically interact with each other to maintain complex higher-level cognition (Liang et al. 2014). Therefore, our observed decreased intra-/intermodular connectivity in the old–old highlighted the loss of structural specialization (Allen et al. 2011; Dai and He 2014; Geerligs et al. 2015). Finally, we found decreased degree/betweenness centrality in the old–old, which was in line with prior aging findings (Wang et al. 2018). Degree/betweenness centrality characterizes the capability of parallel information communication in the brain networks, which implied the magnitude of integration of brain networks (Hagmann et al. 2008). In brief, the aforementioned topological alterations supported the disconnection hypothesis of brain aging and reflected the disruption of integration and segregation of brain networks in the nondemented old–old (Bennett and Madden 2014).
A noteworthy finding was the significant accelerating weakening of modular organization (i.e. intra-/intermodular connectivity) in the old–old compared with the young–old. Such differentiating profiles between the young–old and old–old in connectivity were in line with one of our previous longitudinal studies, which indicated the turning points for an accelerated decline of brain structures at round 75–80 years (Shen et al. 2018). Previous studies have indicated an accelerating decline in white matter maturation during this period (Westlye et al. 2010; Sexton et al. 2014). This would contribute to the accelerating disconnection in the topological organization of structural networks in the nondemented old–old. Therefore, in conjunction with previous studies (Shu et al. 2017; Li et al. 2020), we postulated that the accelerating decline in structural topology may be a consequence to rapid white matter degeneration in advanced age.
Degeneration of white matter connectivity serves as structural substrate of cognitive aging
Our mediation models provided the essential evidence that cognitive aging in the nondemented older participants was in part through the age effect on the degeneration of white matter connectivity. This highlighted that the degeneration of white matter microstructure and brain networks was not solely brain aging characteristics, but also a neural underpinning of cognitive decline in advanced age. Notably, our results indicated that white matter connectivity constituted only a small portion of the mediation effect on cognitive aging. Several recent studies have also documented close relationship between white matter connectivity, cognitive aging, and other brain features, such as gray matter volume, white matter hyperintensity, and cerebrovascular risk factor (Brown et al. 2019; Veldsman et al. 2020). Together with these findings, current finding offered a clue that the degenerative process of white matter connectivity likely underlain the behavioral and cognitive declines in advanced age, which need to be further explored.
In addition, our machine-learning predictive models established a neurocognitive association between white matter connectivity and longitudinal cognition in our nondemented older participants. This affirmed the prior report on the predictive ability of white matter connectivity for cognitive status in advanced age (Madole et al. 2021). However, none of those studies have focused on the microstructural organization of white matter to construct the predictive models. Our studies revealed a substantive improvement in predictive ability when combining white matter connectivity in both microstructure and brain networks into one model. This provided robust evidence for the claim that these microstructures and network-based topologies were specialized to be neural fingerprints for cognitive aging.
Limitations
Several inherent issues of this study need to be addressed. First, potential confounds that have been reported in cross-sectional studies include cerebrovascular risk factors, white matter hyperintensity, and cohort effects, and our study did not include these factors due to lack of FLAIR sequence and T2 weighted MRI images (Veldsman et al. 2020). Further longitudinal studies need to consider these potential variables to validate the current findings with individual trajectories. Second, the older population tends to have the larger head motions due to their relatively lower cooperation in the scanner. Thus, the motion correction and DTI tractography algorithm commonly used in the pipeline cannot ensure the enough efficiency, which may result in the potential false-negative fiber tractography result, especially in the regions of crossing fibers (Jones et al. 2013). Third, the modular organization of brain networks was constructed based on previous studies (Power et al. 2011; Yeo et al. 2011). Data-driven approach can be another option to identify the modular structures using modularity algorithms such as spectral optimization. Fourth, we only investigated the topological properties of SCNs. Unlike structural connectivity, functional connectivity is deemed to be the direct biological representation of cognitive activities (Zhao et al. 2022). In this vein, combination of functional and structural connectivity may facilitate a mechanistic understanding of the neural substrates for cognitive declines in advanced age. Finally, although our results showed robust predictive effect of white matter connectivity for global cognition using the cross-validation approach, the test–retest reliability of predictive models needs to be further validated for its stability and robustness in future studies using another independent large-sample older population (Zhao et al. 2015).
Conclusion
The present study delineated the age-related alteration patterns of white matter connectivity, including a relatively stable period in the young–old and an accelerating deterioration in the old–old. The white matter connectivity mediated the effect of age on global cognition and had relatively notable predictive effect for longitudinal cognition in nondemented older participants. These findings facilitate our understanding of the degenerative profiles of white matter connectivity in nondemented older adults and shed light on the novel white matter connectivity-based imaging markers for the prediction of cognitive aging.
Funding
This work was supported by the National Natural Science Foundation of China (Grant No. 61971017 and 81871434) and Beijing Natural Science Foundation (Grant No. Z200016). MAS (The Sydney Memory and Aging Study) cohort was supported by 3 National Health & Medical Research Council (NHMRC) Program Grants (ID No. ID350833, ID568969, and APP1093083). We thank the study participants and the MAS research team.
Conflict of interest statement: None declared.
Author contributions
HZ contributed to conceptualization, methodology, formal analysis, original draft, and interpretation of manuscript. WW contributed to methodology, data resources, review, and editing of manuscript. JC contributed to methodology, revision, and funding. JJ and HN revised the manuscript. NK and HB contributed to data resources and funding. PS contributed to data resources, revision, and funding. TL contributed to conceptualization, supervision, methodology, data resources, review, and funding.
