Optimal age and outcome measures for Alzheimer's disease prevention trials in people with Down syndrome

People with Down syndrome (DS) typically develop Alzheimer's disease (AD) neuropathology before age 40, but a lack of outcome measures and longitudinal data have impeded their inclusion in randomized controlled trials (RCTs).

in the brain as people age. Such plaques, alongside neurofibrillary tangles of tau protein, form the characteristic neuropathology of AD, which develops nearly universally in adults with DS before the age of 40 years. 5,6 Dementia is most frequently diagnosed between 50 and 55 years, and 90% of adults with DS are expected to develop dementia in their lifetime. [7][8][9] The predictable onset of AD neuropathology and symptoms in DS opens opportunities for early intervention that are unfeasible in sporadic AD. Yet despite advances in understanding the development of AD in DS, people with DS have historically been, and continue to be, excluded from randomized controlled trials (RCTs). 10 Aside from being a missed scientific opportunity, this exclusion limits access to health care that is appropriate and targeted to this population's needs.
A lack of data detailing the subtle cognitive changes that occur during the preclinical to prodromal (ie, symptomatic) stages of AD in DS has likely contributed to adults with DS being overlooked for inclusion in AD prevention trials. Impact on clinical outcomes remains a prerequisite for approvals of drugs targeting these early stages of AD (that is, progression from stage 1 to 3 in the FDA classification). 11 Fluid and neuroimaging biomarker studies are underway in people with DS, [12][13][14] and will undoubtedly help with future targeted trial recruitment. However, given difficulties in obtaining such data in this population, detailed examination of early cognitive signs may offer a complementary approach for monitoring disease progression.
A further limitation of current RCT design in AD is the absence of considering the earliest age at which decline in cognitive abilities can be detected. Identifying the inflection point in transition from preclinical to prodromal stage AD could help determine the ideal window to intervene with pharmacological treatments, such as amyloidtargeting drugs, that could slow or prevent further neuropathological progression.
With the advent of large, longitudinal DS cohort studies, data are now becoming available that can help address these concerns. [15][16][17] Such data can help accelerate progress in this high-dementia risk population, offering important insights into AD development, prevention, and treatment that could eventually improve AD outcomes for those both with and without DS.

Objectives
In keeping with the European Medicines Agency guidance highlighting the need to identify subscales and items that are sensitive during early stages of symptomatic AD, 11 we aimed to use longitudinal data from a DS cohort study to:

Study design
This was a longitudinal cohort study.

Setting
The study was conducted in a community setting in England. Baseline assessments were completed between October 2013 and September 2015, with follow-up assessments 2 years later. Assessments took place at participants' homes, day-care centers or university testing centers, according to participant preference.

Participants
Adults with DS were recruited from the LonDownS cohort. 16 Participants aged 36 years or older at baseline were eligible (n = 173). By this age, AD neuropathology is universally expected; 5,18 thus participants can be considered to be in at least a preclinical stage of AD. DS was confirmed genetically for 163 participants (details in supplementary material).

Demographic variables
The demographic data were sex, age, level of intellectual disability (ID; carer reported, based on ICD-10 descriptions).

Clinical variables
To ensure dementia diagnoses were independent of neuropsychological assessments undertaken for this analysis, we used diagnoses based on clinical assessments by each individual's psychiatrist. The CAMDEX-

DS (Cambridge Examination for Mental Disorders of Older People
with Down Syndrome and Others with Intellectual Disabilities) 19 was additionally used to identify symptoms of decline in cognition, adaptive functioning, or behavior indicative of early dementia-related change.

Eligibility criteria
Participants required data for at least one outcome measure at baseline and follow-up, complete CAMDEX data, and sufficient hearing and vision to comfortably engage with the cognitive tests (see supplementary material).
For the trajectory modeling of the outcomes using an E max model, individuals with a dementia diagnosis or performance at floor-level on a given outcome measure at baseline were excluded, in order to focus on decline in the preclinical and prodromal stages of AD only.

2.9
Statistical analysis

Event-based modeling
We used event-based models (EBMs) 20 to estimate the sequence in which cognitive markers become measurably abnormal, and to stage patients along this sequence. In brief, the EBM is a probabilistic model of observed data generated by an unknown sequence of events, where an event is defined as the transition of a marker from a normal to an abnormal state. The model learns distributions of normality and abnormality for each marker separately and enables estimation of the most likely sequence of abnormality over the whole population. The EBM has been applied extensively to progressive neurological diseases, including AD 21 and Huntington's disease. 22 We recently developed an EBM for AD in DS using baseline data from the LonDownS adult cohort. 17 Here, we used this model to test

Trajectory modeling of cognitive decline using E max models
To determine the earliest age-bands of change for each outcome measure, we examined dose-response relationships between performance change over 2 years, and increasing "doses" of age. We assumed a sigmoidal (ie, "S-shaped") relationship between performance decline and age in years. This constrained a baseline level of cognitive stability, followed by a period of decline that eventually plateaus.
To allow exploration of changes on a yearly basis, mean proportional change in performance between time points was calculated across participants in 5-year smooth moving-average baseline age bands, starting at age 36 and subsequently incrementing by 1 year. Age bands with fewer than four observations and individual change score outliers (>1.5 times the interquartile range of the static 5-year age band or due to clinical anomalies, such as substantial improvement in performance in an older adult) were excluded from the analysis.
A sigmoid E max model 24,25 was fitted to these change scores, using the "DoseFinding" package for R (version 0.9-16). In the context of our data, the model estimates the following: the baseline group performance in the absence of aging-related decline (E 0 ); the maximum effect of age on performance (E Max ); the age at which half of the effect of the E max is observed (EC 50 ); and an h parameter, the steepness of the curve at the EC 50 value. Jackknife resampling was performed on each model to estimate bias.
EC 1 values were calculated from the model results using a freely available online calculator (https://tinyurl.com/emaxcalc). These values give the age bands in which we can expect to see 1% of the maximum effect of age on performance and were used here as the earliest ages of decline. For reference, EC 5 and EC 10 values are also given in the supplementary pages.

2.9.3
Indicative effect size and sample size estimation Raw performance changes in the age band at EC 1 for each outcome measure were used to estimate required sample sizes to compare groups in hypothesized RCTs where pharmacological treatments would reduce aging-related decline by 35% or 75% compared to placebo, over a 2-year period. Cohen's d was used to show the effect sizes of these hypothetical group differences.
Sample size calculations were performed in GPower 3.1, using independent samples t-tests (α = 0.05), and 80% power.  Figure 1A shows the predicted individual EBM stage at baseline versus follow-up. To permit longitudinal staging and reduce staging uncertainty due to missing data, we required participants to have measurements at both baseline and follow-up and less than 50% missing data; these criteria removed 30 participants (Figure 2). To give an estimate of the uncertainty in the staging due to the sequence, the uncertainty in the sequence ordering estimated by 100 bootstraps of  or were stable (n = 37). Two participants regressed more than three stages due to improvements in cognitive test scores between baseline and follow-up; likely due to missing some assessments at baseline which they then completed at follow-up. We also observed general consistency between the sequences estimated separately using participants at baseline and follow-up ( Figure 1B), with the earliest changes in the PAL (visuo-spatial memory) and NEPSY car motorbike (sustained attention/ praxis) markers.

E max analysis
In order to focus on the earliest signs of decline, for these analyses we

Sample size calculations
EC 1 values and required sample sizes varied across the test battery (

DISCUSSION
Due to their known genetic risk for AD, people with DS are increasingly being considered for early AD intervention RCTs. However, progress has been impeded by a lack of data characterizing the trajectory of ADrelated change in these adults, as well as concerns regarding response to treatments such as anti-amyloid antibodies given immune system differences associated with DS. 26 Using longitudinal cognitive data from a DS cohort, we combined EBM with E max modeling approaches to determine optimal cognitive outcome measures and age bands for tracking the earliest signs of cognitive decline in this population. We demonstrated that the EBM model may be useful for tracking stages of cognitive decline in DS.
Considering a sub-sample that would be eligible for preventive RCTs, we found that cognitive decline could be observed over 2-years in participants 20 years younger than the average age of dementia diagnosis (around 55 years in those with DS). 9 The cognitive outcome measures showing greatest sensitivity at these early stages, with highest feasibility for use in RCTs, were the  Prevention trials in people with DS may ultimately benefit trial design in other populations with AD and provide proof of concept for trials in sporadic AD, which are urgently needed. 29

Limitations
The E max modeling had relatively small sample sizes in each age band.
We minimized this limitation by using moving average age bands.

Conclusion
The cognitive stages of AD in DS can be identified using an EBM staging model, and may have potential to track AD-related change over time. In addition, using a novel application of dose-response models, we determined optimal recruitment age bands and outcome measures for RCTs of drugs targeting the earliest stages of disease in a population at exceptionally high risk of developing AD.
Our results have allowed us to determine a short cognitive battery, taking less than 30 minutes to administer, that is capable of detecting cognitive decline in adults with DS up to 20 years before their average age of dementia diagnosis, to improve prospects for RCTs in individuals with DS. All co-authors contributed to the final article.

DATA SHARING
Deidentified cognitive data collected by the LonDownS consortium will be made available to researchers upon request and following approval of a protocol and a signed data access agreement. A data dictionary with variable descriptions will be made available with each request.
Data requests should be made to Professor Andre Strydom at King's College London: andre.strydom@kcl.ac.uk .