3rd Workshop on Research Definitions for Reserve and Resilience in Cognitive Aging & Dementia

In both research aims, human data will be used. Cross-sectional neuroimaging and clinical data, along with longitudinal cognitive data, from the Irish Longitudinal Study on Ageing (TILDA), will be used. TILDA is a randomly sampled, nationally representative sample. To-date, TILDA data has been collected for five time points/waves. For the purposes of this project, Wave 3 (only time point to-date with neuroimaging data) will be considered the baseline time point, and Wave 5 will be considered the follow-up time point.

In Aim 1, after listwise deletion, the analyses will have an N of 351 participants with a mean age of 67.9 years (SD = 7.3, range = 49 – 87) and with 51% females. In Aim 2, after listwise deletion, the analyses will have an N of 309 participants with a mean age of 67.5 years (SD = 7.3, range = 49 – 87) and with 50% females.

In both research aims, there are two separate predictor variables: cognitive reserve (CR) and brain maintenance (BM). CR will be measured using CR network strength which will be calculated from the application of a robust machine learning model, connectome-based predictive modelling, to functional MRI data. BM will be measured using the brain-predicted age difference, which is calculated from the application of robust machine learning techniques to structural MRI data. To minimise repetition, the specific detail describing these measure are outlined in Section 6 Operational definitions of concepts.

Three separate cognitive outcome variables measured at follow-up will be used in both research aims, with each variable reflecting different domains of cognitive function. These will include global cognition (MMSE), executive function/verbal fluency (Animal Naming Test), and episodic memory (composite measure of immediate and delayed recall of 10-word list).

In Aim 1, covariates will be comprised of the following variables measured at baseline:
• Age
• Sex
• cognitive function
• brain structure measured using Alzheimer’s disease signature cortical thickness. This reflects the mean cortical thickness across entorhinal, inferior temporal, middle temporal, and fusiform cortices and is not confounded by total intracranial volume (Jack Jr. et al., 2015; Neth et al., 2020).

In Aim 2, in addition to the covariates included in aim 1, the following baseline covariates will be added:
• white matter hyperintensity volume measured using FSL BIANCA
• white matter microstructural integrity measured by fractional anisotropy of the genu of the corpus collosum (Neth et al., 2020; Vemuri et al., 2018) generated using ExploreDTI
• physical frailty measured using grip strength and 5 metre walking speed
• alcohol use (as measured using the CAGE questionnaire)
• tobacco exposure (as measured using packyears which corresponds to the number of packs of cigarettes smoked per day by the number of years the person has smoked)
• body mass index
• depressive symptoms (as measured using the CES-D scale)
• systemic vascular health (as measured using a composite score of Cardiovascular and Metabolic Conditions; (2018)). The composite score of Cardiovascular and Metabolic Conditions score is calculated by summing the presence of the following conditions: hyperlipidemia, cardiac arrhythmia, coronary artery disease, congestive heart failure, diabetes mellitus, and stroke. Whereas Vemuri et al (2018) accounted for the presence of these conditions within the last five years, the present project will account for the presence of these conditions at the current or previous time-point which amounts to a three year period.

CR will be operationalised as CR network strength. CR network strength will be developed using a machine learning model, connectome-based predictive modelling (Rosenberg et al., 2015; Shen et al., 2017), applied to a task-based connectivity matrix obtained from the Paper Folding task in the CR/RANN dataset. CR network strength reflects the connectivity strength of connections between different brain regions-of-interest that are positively associated with CR. Using 10-fold cross validation, the summed connectivity strengths will be used to predict a proxy measure of CR, which will be defined by the residual from a multiple regression predicting episodic memory from hippocampal volume, age, and sex. This model will then be externally validated on resting-state fMRI data in TILDA.

BM will be operationalised as a robust machine learning residual, the brain-predicted age difference, which compares an individual’s structural brain health, reflected by their voxel-wise grey matter density, to the state typically expected at that individual’s age. This measure was developed using a robust machine learning model, the Elastic Net with 10-fold cross validation and a data resampling ensemble approach, applied to open-access structural MRI data from 1,359 participants. In this model, voxel-wise grey matter density values were used to predict chronological age. This model was then applied to three independent datasets, including the TILDA dataset, where it accurately predicted chronological age in each dataset (Boyle et al., 2020). Brain-predicted age difference scores were then created by subtracting the brain-predicted age from chronological age, with positive scores reflected worse brain health, or BM, and negative scores reflecting better brain health, or BM.

The key criterion for satisfying an operational definition of CR is that a candidate CR measure must exert a moderation effect on cognitive function, such that it moderates the relationship between brain structure or pathology and clinical status or cognitive function (Stern, Arenaza-Urquijo, et al., 2018). This criterion can be examined in a moderated multiple regression model where cognitive function is predicted using variables representing brain structure, a CR measure, and the interaction term comprising the product of brain structure and the CR measure.

Moderation effects suffer from greatly reduced statistical power when there is measurement error in the variables in the interaction term and/or when either variable are themselves associated with the outcome variable (McClelland & Judd, 1993; Whisman & McClelland, 2005). This is pertinent for functional neuroimaging measures of CR, which will inevitably contain considerable levels of noise (Murphy et al., 2013) and because cognitive function is often associated with brain structure variables (Hedden et al., 2014; Tsapanou et al., 2019) and neuroimaging measures of CR (Stern, Gazes, et al., 2018; van Loenhoud et al., 2020). Moreover, longitudinal neuroimaging studies suggest that CR delays onset of cognitive decline but might not alter the rate of decline (Soldan et al., 2020) which casts doubt on the existence of moderation effects for CR. As such, the moderation effect criterion may be too strict and could result in promising candidate CR measures being too easily discounted.

A less rigorous criterion, the independent effect, holds that a candidate CR measure should demonstrate a positive association with cognitive function, after accounting for brain structure (Stern, Arenaza-Urquijo, et al., 2018). However, the closely related concept of BM has been also positively associated with cognitive function (Boyle et al., 2020; Cole et al., 2018; Elliott et al., 2019; Habeck et al., 2017; Liem et al., 2017). Consequently, this criterion may not dissociate CR from BM, which could hinder attempts to uncover their neural bases.

In sum, the current operational definition of CR is impeded by a strict criterion which may dissociate between CR and BM but might not be statistically feasible and a lenient criterion which can be fulfilled but does not dissociate between CR and BM. If the moderation effect cannot be satisfied, one solution is to refine the independent effect criterion such that a measure of CR must be positively associated with cognitive function after accounting for additional brain structure variables as well as physiological and clinical health measures that are related to BM (Cole et al., 2018; Elliott et al., 2019; Franke et al., 2013, 2014; Han et al., 2019; Kolenic et al., 2018; Lange et al., 2020; Ning et al., 2020; Ronan et al., 2016) and cognitive function (Anstey et al., 2007; Auyeung et al., 2011; Cronk et al., 2010; Hooghiemstra et al., 2017; Sabia et al., 2014; Samieri et al., 2018; Vemuri et al., 2018; Wilson et al., 2004; Yaffe, 2007).

The specific aims of the research proposal are to address the following research questions:

Aim 1: Are the current moderation effect and independent effect criteria appropriately rigorous for operational definitions of CR such that they can dissociate CR from BM?

• Hypothesis 1a: The moderation effect criterion will be satisfied by a measure of CR, but not by a measure of BM, in the presence of appropriate statistical conditions.

• Hypothesis 1b: The independent effect criterion will be satisfied by a measure of CR, and by a measure of BM, and as such it will not be rigorous enough to dissociate CR from BM.

Aim 2: Can the independent effect criterion be refined such that it will be satisfied by a measure of CR but not by a measure of BM?

• Hypothesis 2: The independent effect criterion, adjusting for physiological and clinical health covariates related to BM and cognitive function, will be satisfied by a CR measure, but not by a measure of BM.

Confirmation of hypothesis 1a would demonstrate that the current independent effect criterion is too lenient such that it cannot dissociate between measures of CR and BM. Confirmation of hypothesis 1b would restrict the operational definition of neuroimaging measures of CR to measures that demonstrate moderation effects. This refined operational definition would enable measures of BM and CR to be dissociated. Given the noted difficulties in detecting moderation effects (McClelland & Judd, 1993), confirmation of the hypothesis 2 would refine the independent effect criterion so that the operational definition of CR can be satisfied without demonstrating moderation effects but that can dissociate putative CR measures from BM measures. This would allow candidate CR measures to be validated against a statistically achievable criterion which can dissociate CR from the closely related concept of BM. Ultimately, this would enable researchers to apply such measures in order to study the neural basis of CR without mistakenly focusing on the neural basis of BM.

WP1
– Participants for our discovery sample will be amyloid-positive human individuals from ADNI (estimated age, sex and race; 70±10 years, 50% female, primarily non-Hispanic white). Our replication sample will include amyloid-positive human individuals from the Amsterdam Dementia cohort (ADC) selected to match the discovery sample (estimated age, sex and race; 70±10 years, 50% female, primarily non-Hispanic white).
– Existing data to be collected is longitudinal cognitive data of at least 5 timepoints prior to diagnosis, at least two timepoints with CSF and MRI data prior to diagnosis.

WP2
– Participants will be selected based on the same criteria as in WP1. ADNI will again be used as the discovery sample and ADC as the replication sample
– Existing data to be collected will be longitudinal data from at least 2 timepoints after AD diagnosis with cognitive, CSF and MRI data

WP1:
Predictor variable measurement
– Structural properties of the brain
o Gray matter atrophy
o White matter hyperintensities
– AD pathology
o CSF amyloid
o CSF total tau
o CSF p-tau
Outcome measurement
The following characteristics of the inflection point (see statistical plan) will be used as outcomes:
– Concurrent level of cognitive performance at inflection point
– Time to dementia from inflection point
Covariates:
Age, sex, intracranial volume (for structural properties of the brain)

WP2:
Predictor variable measurement
– Years of education, dichotomized into high education vs low education using a mean split
Outcome measurement
– Cognitive decline after AD diagnosis
– Longitudinal structural properties of the brain
o Gray matter atrophy
o White matter hyperintensities
– AD pathology
o CSF amyloid
o CSF total tau
o CSF p-tau
Covariates
Age, sex, intracranial volume (for structural properties of the brain)

WP1:
Cognitive reserve will be operationalized by the concurrent level of cognitive performance at the inflection point and by time to dementia from the inflection point. Higher performance and longer time to dementia will indicate more reserve. Brain reserve will be operationalized as structural brain properties and brain maintenance will be operationalized as change over time in structural properties and AD pathology.

WP2
For this work package, cognitive reserve will be operationalized by years of education, and will be dichotomized into low and high reserve using a mean split for education. Brain reserve will again be operationalized as structural brain properties and brain maintenance will again be operationalized as change over time in structural properties and AD pathology.

WP1
Validation step: We hypothesize that higher education will be related to higher inflection points (higher initial performance) as well as delayed inflection points (longer time to dementia).

Hypothesis WP1: We hypothesize that structural brain properties (brain reserve) are more strongly associated with higher initial performance (higher inflection points) and that reduced pathological build-up (brain maintenance, e.g. CSF-tau) plays a larger role in delayed onset of cognitive decline (longer time to dementia).

WP2
Validation step: We hypothesize that higher education will be associated with faster cognitive decline, thereby replicating findings with regard to increased decline with higher reserve

Hypothesis WP2: If cognitive decline in individuals with high reserve is only related to baseline structural properties of the brain and not with change over time, this would support the notion of static effects of brain reserve. However, we hypothesize that accelerated cognitive decline in individuals with high education is coupled with increased rates of degeneration and build-up of AD pathology, supporting the notion of a (loss of) brain maintenance capability after cognitive decline has set in.

WP1:
Longitudinal cognitive measures (at least 5 timepoints) will be used to establish individual inflection point towards cognitive decline using Bayesian change point analysis. The concurrent level of cognitive performance at inflection point and time to dementia from inflection point are used to operationalize different aspects of reserve (see Operational definitions of concepts). In a validation step, we will first relate the inflection point characteristics to education using linear regression analyses. We will then use linear mixed model analyses to predict height and onset of inflection points from (changes in) structural properties of the brain (gray matter atrophy, white matter hyperintensities) and AD pathology (CSF-amyloid-beta 42, total tau and p-tau) using linear mixed effects models.

WP2:
First, we will try to replicate earlier findings regarding increased decline with higher reserve by assessing the association between education and rates of cognitive decline using linear mixed effects models. Dichotomized high reserve vs low reserve will then be used to assess the association between cognitive decline and high/low reserve, longitudinal degeneration of structural brain properties and longitudinal change in AD pathology using linear mixed effects models.

Independent variable: Training/Education in the repeated acquisition and performance chamber along with natural aging.

Dependent variable/outcome: Performance in cognitive contextual and spatial tasks (behavior). To determine the neural correlates of cognitive reserve and brain maintenance, we will tag a contextual memory in our male and female ArcCreERT2 x EYFP mouse model at 8 months of age. These mice will either have gone through 2 months of RAPC or have aged normally. As a marker of brain aging, we will also quantify senescent cells.

Cognitive reserve: The brain’s ability to buffer against neuropathology and cognitive stressors that lead to degeneration. Here, education or behavior training will be defined as boosting cognitive reserve.

Brain maintenance: A compliment of CR, refers to the relative absence of change in the brain over time with age. To test brain maintenance, we will have a control group that ages without training

AIM 1: Test the hypothesis that educational training buffers against age-related cognitive decline (ARCD) in ArcCreERT2 x EYFP mice: I hypothesize that education will protect against ARCD and therefore, enhance CR.
We have included a wide-array of behavioral paradigms and are not reliant on one task to accurately assess cognition. However, we are attempting training for the first time in the RAPC and this specific timeline may not produce robust results. If this occurs, we have the option to let the animals age until 12 or 24 months. Our AD mouse line is also bred to the ArcCreERT2, which we can test as an alternate model. We can also increase training to 8 weeks. These parameters can be modified. Additionally, we can use other types of training tasks such as the radial arm maze and hole board

AIM 2: Test the hypothesis that there is a specific neural signature of cognitive reserve in ArcCreERT2 x EYFP mice: I predict that memory trace activation in trained mice will be increased throughout the hippocampus. The number of reactivated memory trace cells will represent a specific neural signature of CR.
Although unlikely, it is possible that training does not affect memory trace cells in the hippocampus. If we do not observe differences in the hippocampus, then we will investigate other brain regions. We can also assess differences in encoding or retrieval alone; it is also possible that trained mice do not exhibit differences in whole brain memory traces during retrieval, but instead exhibit enhanced encoding of the memory. If this is the case, we can examine cells immediately after encoding of the memory

In order to benchmark the functional to a structural-based maintenance index we compare the validation against biomarkers and cognition to the BrainAGE index (Franke et al., 2010). Although this is highly exploratory work, we expect a stronger association of BrainAGE index to existing cognitive differences and a higher associatioo of the function-based index to Amyloid biomarker status, respectively.

3. Are young adults or high maintenance elderly individuals more suited as a reference (training) sample for inference about unseen subjects at risk of Alzheimer’s disease ?
4. How does the one-class approach compare to alternative two-class discriminative approaches ?

Work package 1: Operational definition of maintenance using a normative model of functional brain activity

Here we follow the established principle of casting patient classification as an outlier detection problem (Mourao-Miranda, 2011). Using multivariate metrics, measures of departure from normal brain activity (as originally proposed in Düzel et al., 2011) could extended using Euclidian or Mahalanobis distance from group mean activity in a young reference sample. However, the extremely large number of dimensions characteristic of neuroimaging data (hundreds of thousands of voxels) renders this difficult or impossible due to the small sample size typically available for neuroimaging datasets. One way of efficiently addressing this problem is to compute the probability density of brain activity contrast in a specific (high maintenance) reference class. When a new participant’s brain activity contrast falls below some density threshold this new example is considered abnormal (or more unusual). This is expected to happen when the spatial pattern of local networks recruited during task-related activity does change compared to the high maintenance reference sample. In addition, the distance of new unseen participant’s brain activity to the boundary (learning in the reference) can be used to quantify the degree of abnormality (or deviation) in single continuous index. Then, the latter index can be correlated with clinical outcomes (such as cognition and biomarkers) for validation. In this work package we aim to start adapt the previously established one-class Support Vector Machines (SVM) approach (Mourao-Miranda, 2011) for the quantification of brain maintenance using the FADE task paradigm.
More specifically, machine learning (or AI) kernel methods are used to define a similarity measure of whole brain functional maintenance using the novelty during encoding contrast two individuals performing the task. Linear or non-linear kernel mappings might be potentially useful to characterize different aspects activity pattern similarity relevant for the definition of maintenance. We therefore explore linear as well as non-linear relationships in the data. Using a nonlinear kernel matrix is equivalent to mapping the contrast maps from the original input space into a high dimensional feature space where the distance become more meaningful or the separation between the two classes can be easier (using a plain linear boundary).

Once the decision boundary is fitted during a model training process using a young reference sample, it can be used to classify and characterize unseen elderly participant’s maintenance as being either low (if they fall outside the boundary) or high (normal non-outliers). Alternative multivariate soft-margin (Shawe-Taylor & Christianini, 2004) and probabilistic approaches accounting for full model uncertainties (such as Gaussian Process Classification, Rasmussen & Williams, 2006) and Deep Learning (Pinaya et al., 2018) are explored and compared in this project. Reliability and robustness estimates of the proposed maintenance index will be assessed using longitudinal follow-up data in very low risk subsamples (of the well characterized DELCODE cohort). Accounting for known and undesired covariates is another important issue for single-subject inference methods to be considered in this project phase. The empirically optimized one-class machine learning approach using FADE paradigm’s novelty contrast will result in a new maintenance index using a sample of n=99 young (high maintenance) adults with ages 20-30 performing the scene-encoding (FADE) task.

Work package 2: Validation using elderly individuals with high maintenance and probing the hypothesis of brain reserve

In this work package we test whether the functional brain maintenance index as introduced in WP1 can be validated against a clinically defined high maintenance reference sample. These are older adults who are CSF biomarker and ApoE4 negative, do not have vascular lesions (based on White Matter Hyperintensities (WMH) and vascular risk scores), show absence of significant atrophy and remain cognitively stable (over follow-ups, which are available in DELCODE for several years). Based on previous findings it can be expected that this subsample shows high maintenance (Cabeza et al., 2018, Düzel et al., 2011). Furthermore we test whether there is evidence that it can be used as a determinant of preserved cognition (i.e. using individual level longitudinal cognitive trajectories over follow-ups in DELCODE).
In context of this call cognitive reserve is defined as multiple properties of the brain that might allow for sustained cognitive performance in the face of age-related changes and brain insult or disease (such as Aβ/tau or vascular pathology). Here we specifically focus on the potential reserve variables such as global and local (e.g. hippocampal network) gray matter volumes. In addition to the above validation, this WP explores the contribution (or dissociation) of several factors (when considered simultaneously) to cross-sectional cognitive performance in a single regression model. We include the predictors (1) functional maintenance; (2) structural brain reserve, (3) demographics, and the (4) presence amyloid/tau and vascular pathology. We further hypothesize independent contributions of maintenance and reserve to cognitive performance differences and a moderator effect of structural brain reserve on brain pathology indicator’s (Aβ/tau and WMH) effect on cognition.

Work package 3: Assessing the role of the reference sample.

One class approaches might be fundamentally biased by a somewhat arbitrary choice of the reference sample (and population). A crucial question is whether brain activity contrast differences induced by age and (often mixed) brain pathologies do align in the considered feature space. A related issue is whether one should assume that no functional reorganization e.g. as a consequence of decades of experience and structural brain ageing in high maintenance elderly individuals has happened. Both assumptions are implicitly made in the traditional FADE score (Düzel et al., 2011) and might play out differentially (1) when trying to assess maintenance in healthy functional ageing; or (2) for early identification of brain pathology.
Consequently, by using the traditional approach when extrapolating (from functionally healthy young individuals) ‘long-distance’ towards pathological ageing (i.e. by varying age & pathology simultaneously), we might not obtain an optimal characterization of individual differences and subtle changes (over follow-ups) for one or even both of these important scientific goals (1) and (2). As introduced above, the non-linear dependency of (early) hyper- and (later) hypo-activation along the healthy ageing to AD disease spectrum might also challenge performance of a naive (linear) extrapolation based on young individuals as a reference. In this work package we compare the above approach outlined in WP1 using a low risk (assumed high maintenance) elderly sample (n=100) as an alternative reference for the individual characterization of brain maintenance in unseen subjects. We consider both biomarker, vascular lesion load and cognition for stratification and compare those competing methods.

Work package 4: Does learning a discriminative model of functional brain maintenance offer a promising alternative ?

Our normative (one-class) approach to brain maintenance proposed in WP1 solves a more general (and difficult) problem than conventional two-class machine learning problems (e.g. classifying patients into NO vs. AD). However, two-class and one-class classification approaches address fundamentally different questions. The first finds the discriminative boundary between two classes while the second finds the boundary enclosing a specific class in relation to which activity patterns belonging to other classes can be detected as outliers. In clinical practice, if one is interested in training a classifier to discriminate two well-defined and homogeneous classes with high accuracy, the standard two class approach is expected to have the best performance. However, the proposed one-class approach will be more advantageous in situations where one class being more homogeneous and well defined and the other(s) being highly heterogeneous and/or with small sample sizes.
In this work-package we finally aim to compare the above one-class novelty detection approach to maintenance with an alternative approach explicitly learning the (non-linear) function characterizing the step by step transition of brain (activity) contrast maps from healthy ageing (Amyloid-,Tau-) towards pathological ageing individuals (Amyloid+, Tau+). Multiple kernel-based approaches (such as SVM or GPC) allow accounting for non-linear effects and assessing the predicted ‘distance travelled’ of a patient at risk during progression towards AD. We finally intend to validate both approaches by comparing their potential for early detection of those DELCODE participants that show cognitive decline over longitudinal follow-ups.

To compare the above function-based index with a brain structure-based maintenance index, we aim to complement the performance in the validation steps (WP2 & WP4) with the BrainAge index. The latter is described in detail in Franke et al., 2010, Neuroimage. The pre-trained regression-method will be applied to DELCODE subjects generating their individual Brain-Age which will be entered in all validation analyses using CSF biomarkers and cognition described above.