Introduction

Brain pathologies and changes in brain structure are commonly seen in aging [1]. The two main age-related pathological changes are related to cerebrovascular disease and Alzheimer’s disease (AD) [2]. These pathologies often develop simultaneously [3, 4] and explain a substantial proportion of cognitive impairment in older age [2, 5]. Importantly, both neuropathological changes can be identified decades before clinical symptoms occur [6, 7] and emerging evidence suggests that AD- and vascular-related pathologies can have detrimental effects on brain structure and cognition already at low pathology burden [8,9,10,11]. Thus, studying healthy individuals to identify and distinguish age- and pathology-related changes in cognition in the earliest stages will pave the way to earlier detection, treatment, and preventive strategies [12, 13].

AD and cerebrovascular pathologies have differential impact on brain health early in the disease process. In AD, amyloid (Aβ) pathology has been shown to trigger tau-mediated neuronal death, thereby altering grey matter structure [14]. The deposition of tau in the medial temporal lobe (MTL), specifically in the entorhinal cortex, is consistently found in cognitively healthy older individuals, including those without concurrent Aβ pathology [15,16,17]. Tau pathology is closely related to local cortical atrophy [18, 19] and the pattern of tau accumulation and neurodegeneration mirror cognitive domain-specific dementia symptoms in later disease stages [20]. Cerebrovascular pathologies are commonly observed as white matter hyperintensities (WMHs) on MRI scans. The underlying pathology of WMHs mostly reflects demyelination and axonal loss as a consequence of chronic ischaemia [21]. WMHs may be clinically silent in many individuals but increasing WMH volume is associated with cognitive impairment and AD [22]. Although there are exceptions [23], it is commonly observed that AD-related and cerebrovascular-related pathological processes target distinct cognitive domains. Episodic memory (MEM) relies on MTL structures that are preferentially affected in early stages of AD [5, 24]. In turn, decline in executive function (EXE) are often observed together with cerebrovascular pathologies and are associated with frontal-striatal atrophy [5, 24]. Because cognitive decline in different domains generally occurs together as individuals age [25], it is important to examine age- and disease-related brain changes and cognitive decline in the same context to assess the extent to which they are independent of each other [3, 25].

We used multimodal imaging and statistical frameworks for modeling complex relationships among multiple variables [26] to simultaneously quantify early age- and pathology-related changes in cognition. We hypothesized that already low amounts of brain pathologies in individuals who may be considered “normal aging” carry meaningful information when domain- and sex-specific vulnerabilities are considered. Specifically, we demonstrate how Aβ, tau, WMH volume, lateral ventricle volume, and volume/thickness of five a priori-defined brain regions associate with MEM and EXE through multiple pathways and how age acts as a common factor influencing all investigated variables.

Methods

Study participants

Study participants are part of the ID-cog cohort, an ongoing prospective cohort study at the University of Zurich, Switzerland that started in 2016. The participants are volunteers recruited through newspaper advertisements. To be enrolled in the study, participants had to be at least 50 years of age and German-speaking. Exclusion criteria included inadequate visual and auditory capacities for neuropsychological assessment, presence of clinically significant depression, presence of a medical condition that is seen as the predominant cause of cognitive impairment (e.g., history of stroke), and presence of diseases that would interfere with study procedures in subsequent years. The study was approved by the ethics committee of the Canton Zurich. All participants gave written informed consent prior to the first study procedure.

We recruited 179 cognitively unimpaired (CU) participants and 54 participants with mild cognitive impairment (MCI). The diagnosis of CU or MCI was made following published diagnostic guidelines [27]. In the present study, we included all participants who had both Aβ-PET imaging, MRI, and APOE genotype assessment, leading to the exclusion of one participant due to missing APOE genotype information. Starting in 2017, a subsample of 93 participants of this cohort obtained tau-PET imaging in addition to Aβ-PET. The median time difference between Aβ-PET and tau-PET imaging was 12 months (range: 0–38 months). For analyses involving both Aβ and tau-PET, we used the T1 image closest to the tau-PET scan (maximum 6 months prior to tau-PET). Baseline neuropsychological data obtained during the Aβ-PET acquisition visit were used in all analyses because an additional neuropsychological examination during the tau-PET acquisition visit was not performed in all participants.

Demographic and clinical data

Data on age and sex (self-reported as women or men) were ascertained at the clinical visit. APOE genoty** was performed by commercially available Sanger Sequencing (Microsynth AG). Participants were dichotomized into individuals carrying at least one copy of the ε4 allele and APOE4 non-carriers.

PET/MRI acquisition

We acquired MR and PET images on a 3T Signa PET/MR GE Healthcare scanner. Aβ-PET images were acquired from 90 to 110 min post-injection using approximately 140MBq [18F]-flutemetamol with 4 frames of 5 min each. Tau-PET images were acquired from 80 to 100 min post-injection with 10mCi of [18F]-flortaucipir. A BRAVO 3D T1 MRI sequence (8-channel coil) with voxel size 1 mm in sagittal slice orientation, repetition time (TR) = 8.4 ms, echo time (TE) = 3.2 ms, inversion time (TI) = 450 ms, and flip angle = 12° was acquired in parallel to PET acquisition together with the 3D T2 weighted FLAIR image. A standard 3D CUBE FLAIR sequence with voxel size of 0.48 × 0.48 × 0.6 mm was acquired sagittal with TR/TE/TI = 6502/130.7/1962 ms and flip angle = 90°.

A standard high-resolution T1-weighted fast spoiled gradient recalled acquisition (FSPGR) with inversion recovery scanned on a 750W 3T (32-channel coil) or Premier 3T (48-channel coil) scanner was used for brain parcellation (0.5 mm isotropic voxel size, axial slice orientation, TR/TE/TI = 11/5.2/600 ms, flip angle = 8°).

T1 and FLAIR MRI processing

Cortical thickness measurements were obtained by processing FSPGR 3D T1-weighed images with FreeSurfer image analysis pipelines (version 7.1.1 for CentOS8 (Linux), surfer.nmr.mgh.harvard.edu). FreeSurfer parcellation of each participant was visually inspected for accuracy, and segmentation errors were manually corrected. Cortical thickness regions of interest (ROIs) included entorhinal cortex, parahippocampal cortex, and mean cortical thickness of a large neocortical composite ROI [28], hereafter referred to as NEOcomp. Additional ROIs were volumes of the hippocampus and striatum (average of putamen and nucleus caudatus). These ROIs were selected because they are likely mediators of age-related variation in cognition and were previously used in a study with a similar research question than the current study [3]. Lateral ventricle volume was included as an additional potential predictor of cognitive performance in the elderly [29, 30]. Thickness and volume measures were averaged across left and right hemisphere estimates. Detailed composition of each ROI by FreeSurfer label is shown in SFig. 1.

White matter lesion volume was estimated from T2-weighted FLAIR images. WMH were segmented on FLAIR images using the lesion prediction algorithm as implemented in the LST toolbox version 3.0.0 (www.statistical-modelling.de/lst.html) for SPM. Lesion masks were created by binarizing lesion probability maps at a threshold of 0.65. An optimal threshold was selected after applying four predefined thresholds (0.3; 0.5; 0.65; 0.75) to 20 randomly selected subjects and visually expecting the generated binarized lesion masks for accuracy. WMH clusters smaller than 2.5 mm3 were removed. The minimum cluster is defined as 80% of the smallest lesion size that was consistently detected by three manual raters in a study comparing manual and automated lesion segmentation [31]. Finally, lesion masks were visually inspected and corrected if necessary. In eight participants, WMH volume estimation was not possible due to artifacts on the FLAIR image.

PET processing

We used PMOD NeuroTool (Version 3.9 and 4.1, PMOD Technologies LLC) for processing and analyzing Aβ-PET and tau-PET images together with the corresponding anatomy from the BRAVO sequence. For Aβ-PET, a global neocortical standardized uptake value ratio (SUVR) was estimated from [18F]-flutemetamol uptake in a large cortical composite ROI, which includes frontal, temporal, and parietal cortices and precuneus. The cerebellar grey matter was used as reference region. For descriptive purposes and certain analyses, we used two previously established Centiloid [32] (CL) cut-off values that mark two relevant inflection points denoting different stages of Aβ pathology [33]: a CL of 12 that marks the transition from the absence of pathology to subtle pathology; and a CL of 30 that marks the presence of established pathology.

The tau-PET scan was coregistered to the participant’s T1-weighted MRI scan using rigid body registration. FreeSurfer parcellation of the T1-weighted MRI scan was then applied to the PET data to extract mean regional [18F]-flortaucipir retention. Average uptake was calculated for a MTL ROI that covered bilateral entorhinal cortex and amygdala and for a neocortical (NEO) ROI that covered bilateral inferior temporal and middle temporal gyri [34]. These regions were selected because most normal elderly adults demonstrate elevated binding confined to the MTL, whereas neocortical binding, particularly in the inferior temporal lobe, is often associated with clinical impairment and the presence of Aβ [34]. An eroded inferior cerebellar grey matter mask was used as the reference region [35]. We report results of analyses using non-partial volume-corrected tau PET measurements but note that the results are virtually identical for partial volume-corrected data (geometric transfer matrix method). Off-target binding was addressed in a sensitivity analysis. For this purpose, we created a skull/meninges mask surrounding the brain (SFig. 6) as previously described [36] to control for a potential effect of off-target binding in the skull/meninges.

Neuropsychological examination

Participants completed a battery of neuropsychological tasks. We assessed performance in EXE, MEM, visual construction, and working memory using 16 cognitive tasks as previously described [37]. We converted each individual test score to z-scores using the mean and standard deviation of the cohort. Composite scores for each domain were obtained by averaging the corresponding test z-scores.

Statistical analysis

We specified a series of models that describe pathways to cognitive function with increasing specificity. Structural equation modeling was used to test the validity of the hypothesized age- and pathology-related pathways on cognitive performance. The following paths were included in models including Aβ but not tau-PET: age was specified as a predictor of global Aβ burden, entorhinal cortex thickness, parahippocampal cortex thickness, NEOcomp thickness, hippocampal volume, striatal volume, lateral ventricle volume, WMH volume, and cognitive performance. Aβ burden was specified as predictor of entorhinal cortex thickness, parahippocampal cortex thickness, NEOcomp thickness, hippocampal volume, and striatal volume. As there is evidence that Aβ pathology may contribute to WMH volume [21], Aβ burden was also specified as a predictor of WMH volume. All structural and pathological measures were specified as predictors of cognitive performance. Cognitive performance was measured using latent variables, similar to our previous work [37]. Here, a latent variable represents the commonality of all neuropsychological tests assigned to it and thus reduces the influence of measurement errors inherent in each individual test [26]. Neuropsychological tests and composite scores assigned to each construct are indicated in Fig. 1A and STable 3.

Fig. 1: Cognitive domain structural equation model demonstrating a near-universal effect of age on all variables in the model and the mediating roles of pathological and structural measures on cognitive performance in episodic memory and executive function.
figure 1

A shows significant and non-significant associations in the specified model. Solid lines represent significant paths at *P < 0.05, **P < 0.01, and ***P < 0.001. A continuous amyloid burden variable was included in the presented model. The left-right-headed arrow indicates the residual covariance between episodic memory and executive function performance or correlated residuals of indicators. Indicators were allowed to correlate as the same words were used in these tasks. Residual correlations among indicator variables were highly significant (all P’s < 0.001). In B, significant residual covariances among mediating variables are shown. All model parameter estimates are standardized. The R2 values denote the variance of the corresponding variables that the model was able to explain. n.s. non-significant, EC entorhinal cortex, PhC parahippocampal cortex. N = 232.

Full size image

In a first model, global cognition was set as outcome variable (named global cognition model). Then, the global cognition variable was replaced with MEM and EXE latent variables (cognitive domain model). We focused on MEM and EXE cognitive domains because we suspected distinguishable pathways towards impairment in the two domains [5, 24]: AD-related pathways towards memory impairment and vascular-related pathways toward executive function impairment. Based on the result of this model, but also considering previous literature [24], we subsequently split the model into an MEM and EXE sub-model to investigate sex differences. Sex differences were examined in these sub-models after establishing measurement invariance of the MEM and EXE constructs (details in Supplementary Materials). Finally, MTL and NEO tau were added to the MEM sub-model (tau sub-model). Given that Aβ-PET signal is no longer associated with atrophy measures after accounting for tau-PET signal [19], Aβ burden was specified as a predictor of MTL and NEO tau, but no longer as a predictor of structural measures. MTL tau was specified as predictor of entorhinal and parahippocampal thickness, and hippocampal volume. Based on evidence suggesting tau in the entorhinal cortex is associated with MEM independently of atrophy [3, 64]. However, individuals in these cohorts were older and probably had a greater pathological burden and greater variability in cognitive performance than the present cohort. Our recruitment strategy, coupled with the cross-sectional study design, may also explain the unexpected direction of the association between NEOcomp thickness and MEM. Higher NEOcomp thickness might be significantly associated with poorer MEM performance because older participants in the cohort were increasingly selected to have preserved, well-functioning MEM, but age- and AD-related processes continued to adversely affect NEOcomp thickness. Particularly older individuals with high Aβ pathology or other pathologies associated with reduced cortical thickness will need to have well-preserved MEM to participate in the study. This in turn would suggest compensatory or reserve mechanisms that allow these individuals to maintain cognitive performance [65].

There are several limitations of this study. The sample size of the sub-group with both Aβ and tau-PET was relatively small. The good model fit to data and the fact that paths in our structural equation models were based on previous literature increase our confidence in the correctness of our results despite the limited sample size. Furthermore, many of our conclusions are consistent with a recent study with a different statistical approach and a novel tau PET-tracer [66]. Nevertheless, we were limited in investigating potentially moderating effects of sex and APOE genotype or Aβ-independent effects of the APOE genotype [46, 67]. Furthermore, as the participants were selected to represent a relatively healthy aging cohort, the variability in Aβ and tau pathology was low with many participants, particularly women in the tau sub-cohort, having very low Aβ pathology. As for our statistical approach, we chose structural equation modeling because it allowed us to examine multiple variables simultaneously; however, this likely weakened the predictive power of many mediating variables in the model due to suppression of their non-shared variance. Furthermore, the individual structures do not act independently of each other to predict cognition. For example, the MTL is a complex system in which the ensemble likely behaves in ways not predicted by its components [68]. Future research should therefore investigate functional properties of these regions and how they respond to structural atrophy and pathological measures [38]. Finally, we took a closer look into AD-related processes as the a priori knowledge of the pathophysiological processes is broader than is the case, for example, with WMH. However, detrimental effects of WMH on brain structure or cognition would likely be detectable even in healthy adults if more regional WMH were studied [69].

In conclusion, we showed that with increasing age, multiple, often sex-modified pathological pathways begin to develop that ultimately may lead to cognitive decline. The contribution of these underlying pathological processes to cognitive alterations were subtle in this cohort and a large proportion of the effect of age on cognition was not mediated by any of the pathological and structural measures included in this analysis. Recognizing these early signs of pathology-related cognitive decline may be critical for selecting individuals for therapies before degenerative processes have progressed too far.