Skip to main content
NCBI home page
UKPMC Funders Author Manuscripts logoLink to UKPMC Funders Author Manuscripts
. Author manuscript; available in PMC: 2024 Dec 8.
Published in final edited form as: Alzheimers Dement. 2023 Jun 11;19(12):5506–5517. doi: 10.1002/alz.13161

Predicting Amyloid-beta Pathology in the General Population

Phuong Thuy Nguyen Ho 1, Joyce van Arendonk 1,2, Rebecca ME Steketee 1, Frank JA van Rooij 2, Gennady V Roshchupkin 1, M Arfan Ikram 2, Meike W Vernooij 1,2, Julia Neitzel 1,2,3
PMCID: PMC7616996  EMSID: EMS177420  PMID: 37303116

Abstract

Introduction

Reliable models to predict amyloid beta (Aβ) positivity in the general aging population are lacking but could become cost-efficient tools to identify individuals at risk of developing Alzheimer’s disease.

Methods

We developed Aβ prediction models in the clinical Anti-Amyloid Treatment in Asymptomatic Alzheimer’s (A4) Study (n = 4,119) including a broad range of easily ascertainable predictors (demographics, cognition and daily functioning, health and lifestyle factors). Importantly, we determined the generalizability of our models in the population-based Rotterdam Study (n = 500).

Results

The best performing model in the A4 Study (area under the curve [AUC] = 0.73 [0.69–0.76]), including age, apolipoprotein E (APOE) ε4 genotype, family history of dementia, and subjective and objective measures of cognition, walking duration and sleep behavior, was validated in the independent Rotterdam Study with higher accuracy (AUC = 0.85 [0.81–0.89]). Yet, the improvement relative to a model including only age and APOE ε4 was marginal.

Discussion

Aβ prediction models including inexpensive and non-invasive measures were successfully applied to a general population–derived sample more representative of typical older non-demented adults.

Keywords: Alzheimer’s disease, amyloid-beta pathology, prediction models, machine learning, dementia

1. Background

In 2021, the World Alzheimer Report estimated there were > 55 million cases of dementia worldwide and forecasted that this number would increase up to 42% within the next 10 years.1 Alzheimer’s disease (AD) is the leading cause of dementia, defined by neurotoxic plaques forming amyloid beta (Aβ) peptides in the brain.2,3 Aβ can trigger the aggregation of neurofibrillary tangles and subsequently neurodegeneration resulting in progressive and irreversible cognitive decline. It was estimated that Aβ accumulation starts 15 to 20 years before the onset of clinical symptoms.4

Given the key role of Aβ accumulation in the pathophysiology of AD, enormous efforts are being undertaken to develop anti-amyloid drug treatments that remove Aβ plaques at the preclinical stage, before dementia symptoms manifest.5,6 A crucial step in study enrolment, as well as in translating treatment into clinical practice, is to identify patients at an early stage of AD when irreversible brain damage is still minimal. The only two clinically approved methods to confirm an elevated Aβ burden are a positive amyloid positron emission tomography (PET) scan or positive cerebrospinal fluid (CSF) markers, both of which are costly and invasive procedures with limited availability and restricted to hospitals. Identifying at-risk individuals via an algorithm predicting Aβ positivity is a cost-efficient, non-invasive method that could help screening patients in clinical trials and eventually in primary care before more elaborate confirmatory testing.

Ashford et al. provided a review of previously developed Aβ prediction models.7 They found that prior work mostly restricted to patients from highly specialized memory clinics, for example, 31.5% of studies included the Alzheimer’s Disease Neuroimaging Initiative (ADNI) cohort.7–12 Due to recruitment mechanisms (e.g., self-selection in response to advertisement) and strict inclusion criteria (e.g., no vascular disease), clinical studies like ADNI tend to include highly educated individuals who are more likely to report a family history of dementia and show fewer comorbidities as well as higher prevalence of Aβ pathology than observed in the general population.13–16 Aβ prediction models derived from clinical samples may not translate well to broader applications. Of particular concern is the lack of internal and external validation found in 41% and 71%, respectively, of previous studies.7 In contrast to clinical samples, epidemiological population-based studies invite all residents of a well-defined area to participate with less stringent inclusion criteria, and by design, are more representative of the general population.17 Of the 21 studies which performed external validation in Ashford et al.,7 only one study validated Aβ prediction models in a population-based sample. This study demonstrated good generalizability, but the validated model contained only three predictors (age, apolipoprotein E [APOE] ε4, memory performance) achieving moderately high performance (area under the curve [AUC] = 0.71).18 To extend previous work, we examined the generalizability of more complex Aβ prediction models in the current study.

Our goal was to determine how accurately prediction models, developed in a large convenience (clinical) sample with a wide range of easily ascertainable predictors, could identify amyloid-positive individuals from a population-based sample of older nondemented adults. To this end, we developed two Aβ prediction models (with and without APOE ε4 genotype) in cognitively unimpaired participants of the cross-sectional Anti-Amyloid Treatment in Asymptomatic Alzheimer’s (A4) Study (n = 4,119), considering 19 predictors. It should be noted that other authors had already developed Aβ prediction models in the A4 Study.19–21 Yet, new model development was necessary to include the largest possible set of predictors that was available in both the development and validation sample. Second, we internally validated our models by estimating how accurately they identified Aβ status in A4 Study participants not included in the model development, as well as how much prediction improved compared to “basic models” (age and APOE ε4), the two strongest known Aβ predictors.7 Third and critically, we assessed external validity and temporal stability in the prospective population-based Rotterdam Study (n = 500), which was recently enriched by amyloid PET (2018–2021), using predictors collected at three different timepoints (on average 12 years before, 7 years before, and 2 years after PET acquisition).

2. Methods

2.1. Participants

2.1.1. Anti-Amyloid Treatment in Asymptomatic Alzheimer (A4) Study

The A4 Study was a randomized clinical trial that tested whether solanezumab, an anti-amyloid antibody, slowed down cognitive decline at the preclinical stage.14 The study consisted of 67 sites across four countries (United States, Canada, Australia, and Japan) and collected data from 2014 to 2017.22 Inclusion criteria were age 65 to 85 years, living independently, normal cognition (Mini-Mental State Examination [MMSE] between 25 and 30, Clinical Dementia Rating [CDR] = 0, Logical Memory II between 6 and 18 depending on educational level) and having a study partner. Exclusion criteria were use of AD medication, significant depression or anxiety, and unstable medical condition. For the current study we used the screening data that was collected before the start of the clinical trial and included 4,486 participants who all underwent amyloid PET examination. We excluded participants without information on APOE ε4 genotype (n = 45). We further excluded participants with missing data regarding any of the 19 predictors (n = 322). The final sample contained 4,119 participants, which served as our training and test dataset. A flowchart of the participant inclusion is shown in Figure 1A.

Figure 1. Flow chart illustrating the study design of (A) the A4 Study and (B) the Rotterdam Study.

Figure 1

Open in a new tab

2.1.2. Rotterdam Study (RS)

The Rotterdam Study is an ongoing longitudinal population-based cohort study in the well-defined Ommoord district in the city of Rotterdam in the Netherlands.23 The Rotterdam Study started with 7,983 participants (RS-I) in 1990, extended with 3,011 participants (RS-II) in 2000, and 3,932 participants (RS-III) in 2006 (response rates were 78%, 68%, and 65%, respectively). Participants were re-examined every 3 to 4 years. Between 2018 and 2021, a subsample of RS-II and RS-III participants, who were ≥60 years, had a good-quality brain magnetic resonance imaging [MRI], no PET-related contraindications, no large cortical infarcts, or a clinical diagnosis of dementia were invited for PET examination. Out of 1,697 invited participants, 645 made an appointment (response rate 38%) and 639 PET scans were acquired (more details in Method S1 in supporting information and in van Arendonk et al.24). Figure 1B and Figure S1 in supporting information illustrate participant inclusion and study design of the Rotterdam Study. For the current study, we excluded participants with missing data for any predictor that was chosen in our models. Overall, 365, 500, and 351 participants had all predictors collected on average 12 years before (2006-2011), 7 years before (2010–2015), and 2 years (2021–2022) after PET, respectively. We used the largest dataset (n = 500) to investigate the external validity of our prediction models. Because only a subsample of the Rotterdam Study could receive a PET examination, we evaluated possible selection bias with respect to all individuals who were eligible for the PET study but did not participate (Table S1 in supporting information). PET participants were on average younger (69.0 vs. 71.7 years), more highly educated (34.0% vs. 24.7%), had slightly higher MMSE scores (28.6 vs. 28.3), and better Digit Symbol Substitution Test performance (32.3 vs. 30.9 pairs) than non-participants. All other variables, for example, APOE ε4 or family history of dementia, showed no significant group differences.

2.2. Study Outcome

18 F-florbetapir and 18 F-florbetaben amyloid PET imaging was, respectively, performed in the A4 Study22 and the Rotterdam Study24 and further processed according to an established pipeline in which average cortical standardized uptake value ratio (SUVR) was calculated within FreeSurfer-defined frontal, cingulate, lateral parietal, and lateral temporal regions and using the cerebellum as a reference (more details in Method S2 in supporting information). Aβ status was determined by an algorithm combining both quantitative SUVR and qualitative visual reads in both studies.25 Two tracer-specific SUVR thresholds were used to mark early and established Aβ accumulation, for example, 1.10 to 1.15 in the A4 Study26 and 1.10 to 1.24 in the Rotterdam Study.24,27 An SUVR > 1.15/1.24 was deemed positive while an SUVR < 1.10 was deemed negative independent of the visual rating. An SUVR between both thresholds was deemed positive only when the visual read was considered positive by two independent raters.

2.3. Study Predictors

We included all possible predictors that were collected identically or comparably and had no more than 30% missing values in both cohorts (for details regarding predictor inclusion/exclusion see Table S2 in supporting information).

2.3.1. Demographics

The demographic predictors included age, sex (female, male), education (lower [< 10 years of education], further [10–12 years], higher [> 12 years]), marital status (married/not married), and family history (0, 1, or 2 parents diagnosed with dementia).

2.3.2. Genetic measures

We included the number of APOE ε4 risk alleles (0, 1, 2) in our model, as it is a strong genetic predictor of late-onset AD.7 APOE genotyping (rs7412 and rs429358) was performed on the Illumina Global Screening Array in the A4 Study or on a biallelic Taqman assay in the Rotterdam Study.24,28

2.3.3. Objective measures of cognitive performance

We included a screening test for dementia (MMSE), a test for executive functions (Digit- or Letter-Symbol Substitution), and for memory performance (total free recall score from the Free and Cued Selective Reminding Test of the A4 Study29 and the 15-words learning test of the Rotterdam Study30). Both memory tests measured delayed word recall under controlled learning conditions.

2.3.4. Subjective measures of daily functioning

We also considered self-reports on cognitive complaints and daily activities. To this end, two independent evaluators (PTNH and JN) matched the content of different questionnaires across the two studies. They consistently identified four questions with comparable content reflecting “subjective memory difficulties” and “subjective word-finding difficulties” as well as the “need for assistance with finances or medication” (more details in Table S3 in supporting information).

2.3.5. Health and lifestyle measures

The seven health and lifestyle predictors we included were body mass index (BMI, kg/m2), current smoking (yes/no), alcohol consumption (number of glasses per day), sleep duration (number of hours per night), napping during the day (yes/no), and physical activity (time spent doing aerobics exercise and walking). In the A4 Study, physical activity was assessed using the two questions: “average number of minutes of walking per day” and “average number of hours of aerobic exercise per week.” In the Rotterdam Study, physical activity was assessed using the LASA Physical Activity Questionnaire.31

2.4. Data Analysis

Data analysis was performed in R statistical software (v4.1.3). To develop our Aβ prediction models, we split the A4 dataset into two parts, with 80% serving as the training set and 20% as the test set. The test set was not seen during model training. To select only the most informative predictors, we applied the least absolute shrinkage and selection operator (LASSO) technique (caret package, v6.0-92). Compared to standard logistic regression, LASSO constrains the sum of the regression coefficients to minimize overfitting and model misspecification, which are known problems for predicting rare events such as Aβ positivity.32 Specifically, LASSO can discard predictors from the final model (by shrinking their coefficients), thus reducing variance that is specific to the training data but would otherwise compromise generalizability. The strength of the coefficient shrinkage is determined by the lambda parameter. To choose the optimal values for lambda, we ran 10-fold cross-validation during which we oversampled amyloid-positive cases using the Synthetic Minority Oversampling Technique33 to prevent the algorithm from learning mainly to predict amyloid negativity. Because the coefficient shrinkage is sensitive to the variables’ unit, all predictors were centered and scaled.

For our second aim, evaluating internal validity, we determined the models’ calibration (by calibration slope and intercept [rms package v6.2.0]) and classification performance in the A4 test set. Classification performance was assessed by the area under the curve (AUC; pROC package v1.18.0). We calculated the 95% confidence intervals (CI) of the AUC values based on 1000 bootstrap samples. In addition to AUC, we also reported sensitivity, specificity, and positive and negative predictive value. We further estimated whether our models (hereafter referred to as “extended model”) added predictive performance beyond a “basic model” containing age and APOE ε4.

For our third aim, validating our models in an independent population-based sample, we compared the AUCs in the A4 test set to those in the Rotterdam Study. To estimate the models’ temporal stability, we compared the AUCs with predictors collected at the three different Rotterdam Study visits. In a supplementary analysis, we contrasted model performance across datasets that contained only those participants that had all predictors available at all three visits (n = 178). Finally, based on our models’ performance in the Rotterdam Study validation dataset, we estimated how many PET scans would need to be performed to find one amyloid-positive scan when using our models with and without APOE ε4 compared to a situation in which no prediction model is used.

3. Results

3.1. Sample characteristics

Table 1 summarizes the sample characteristics. The average age at PET acquisition was 71.3 (standard deviation [SD] = 4.7), 71.0 (SD = 4.8), and 69.0 years (SD = 5.10) in the A4 training set, A4 test set, and the Rotterdam Study validation set, respectively. All three datasets included slightly more women (59.8%, 59.7%, and 52.4%, respectively) than men and the majority of participants were White (93%, 93.6%, and 90%, respectively). In the A4 training and test sets, 34.7% and 36.5%, respectively, carried at least one APOE ε4 risk allele, which was not significantly different from the 31.6% in the Rotterdam Study. More than half of the A4 Study participants, 64.6% in the training set and 66% in the test set, had at least one parent diagnosed with dementia, in contrast to only 5.0% in the Rotterdam Study. Aβ positivity was more frequent in the A4 Study (29.9%) than in the Rotterdam Study (15.6%).

Table 1. Sample characteristics.

A4 Study Rotterdam Study
Training set Test set 12 years before PET 7 years before PET 2 years after PET
n 3296 823 365 500 351
Age at PET, mean (SD) 71.30 (4.66) 70.98 (4.75) 69.51 (5.33) 68.99 (5.10) 68.64 (5.04)
Years between predictors and PET, mean (SD) 0 (0) 0 (0) -12.32 (0.96) -7.03 (0.86) 1.86 (0.71)
Amyloid-PET status (%) Negative 2312 (70.1) 577 (70.1) 302 (82.7) 422 (84.4) 306 (87.2)
Positive 984 (29.9) 246 (29.9) 63 (17.3) 78 (15.6) 45 (12.8)
Demographic information
Sex (%) Female 1970 (59.8) 491 (59.7) 199 (54.5) 262 (52.4) 179 (51.0)
Male 1326 (40.2) 332 (40.3) 166 (45.5) 238 (47.6) 172 (49.0)
Race (%) White 3064 (93.0) 770 (93.6) 337 (92.3) 450 (90.0) 319 (90.9)
Asian 68 (2.1) 15 (1.8) 5 (1.4) 6 (1.2) 3 (0.9)
Black or African American 118 (3.6) 28 (3.4) 3 (0.8) 3 (0.6) 2 (0.6)
American Indian or Alaskan Native 6 (0.2) 3 (0.4) 0 (0) 0 (0.0) 0 (0.0)
Native Hawaiian or other Pacific Islander 2 (0.1) 0 (0) 0 (0) 0 (0.0) 0 (0.0)
Mixed 0 (0) 0 (0) 1 (0.3) 1 (0.2) 0 (0.0)
Not available 38 (1.2) 7 (0.9) 19 (5.2) 40 (8.0) 27 (7.7)
Education (%) Lower 13 (0.4) 4 (0.5) 66 (18.1) 91 (18.2) 55 (15.7)
Intermediate 311 (9.4) 73 (8.9) 166 (45.5) 239 (47.8) 162 (46.2)
Higher 2972 (90.2) 746 (90.6) 133 (36.4) 170 (34.0) 134 (38.2)
Married (%) No 999 (30.3) 232 (28.2) 57 (15.6) 101 (20.2) 95 (27.1)
Yes 2297 (69.7) 591 (71.8) 308 (84.4) 399 (79.8) 256 (72.9)
Family history (%) No parent had dementia 1166 (35.4) 280 (34.0) 347 (95.1) 475 (95.0) 332 (94.6)
One parent had dementia 1788 (54.2) 460 (55.9) 17 (4.7) 24 (4.8) 18 (5.1)
Both parents had dementia 342 (10.4) 83 (10.1) 1 (0.3) 1 (0.2) 1 (0.3)
Genetic measures
APOE e4 allele count (%) 0 2153 (65.3) 523 (63.5) 259 (71.0) 342 (68.4) 253 (72.1)
1 1038 (31.5) 273 (33.2) 92 (25.2) 141 (28.2) 87 (24.8)
2 105 (3.2) 27 (3.3) 14 (3.8) 17 (3.4) 11 (3.1)
Objective measures of cognitive performance
MMSE, mean (SD) 28.81 (1.21) 28.86 (1.17) 28.48 (1.47) 28.57 (1.29) 27.53 (5.58)
Delayed word-learning test, mean (SD) 28.91 (5.55) 29.23 (5.77) 8.18 (2.75) 8.65 (2.77) 7.25 (2.85)
Digit-symbol substitution test, mean (SD) 43.65 (8.91) 44.21 (9.09) 33.52 (5.79) 32.34 (5.90) 30.68 (6.27)
Subjective measures of daily functioning
Subjective memory difficulty (%) No 2547 (77.3) 633 (76.9) 199 (54.5) 281 (56.2) 162 (46.2)
Yes 749 (22.7) 190 (23.1) 166 (45.5) 219 (43.8) 189 (53.8)
Subjective word-finding difficulty (%) No 1236 (37.5) 324 (39.4) 273 (74.8) 381 (76.2) 240 (68.4)
Yes 2060 (62.5) 499 (60.6) 92 (25.2) 119 (23.8) 111 (31.6)
Need for assistance with finances or medication (%) No 3076 (93.3) 770 (93.6) 340 (93.2) 437 (87.4) 319 (90.9)
Yes 220 (6.7) 53 (6.4) 25 (6.8) 63 (12.6) 32 (9.1)
Lifestyle measures
BMI, mean (SD) 27.59 (5.18) 27.44 (4.78) 27.04 (4.13) 27.36 (4.14) 27.15 (3.96)
Aerobic exercise, hours/week, mean (SD) 2.87 (3.79) 3.03 (3.89) 2.98 (3.62) 3.14 (4.72) 2.76 (5.19)
Walking, minutes/day, mean (SD) 58.75 (60.47) 60.90 (66.40) 28.02 (36.06) 31.93 (40.10) 26.93 (25.91)
Sleep duration, hours/night, mean (SD) 7.12 (1.06) 7.09 (1.06) 6.71 (1.06) 6.75 (1.21) 6.64 (1.41)
Napping during the day (%) No 2062 (62.6) 504 (61.2) 321 (87.9) 433 (86.6) 298 (84.9)
Yes 1234 (37.4) 319 (38.8) 44 (12.1) 67 (13.4) 53 (15.1)
Current smoking (%) No 3245 (98.5) 812 (98.7) 300 (82.2) 411 (82.2) 328 (93.4)
Yes 51 (1.5) 11 (1.3) 65 (17.8) 89 (17.8) 23 (6.6)
Alcohol, glasses/day, mean (SD) 0.77 (1.12) 0.73 (1.19) 0.77 (0.93) 0.71 (0.91) 0.54 (0.80)
Open in a new tab

Abbreviations: APOE, Apolipoprotein E; BMI, Body Mass Index; y, years; MMSE, Mini Mental-State Examination; SD, standard deviation

3.2. Model development and predictor selection

The extended model included most predictors, except higher education, BMI, and alcohol consumption. The strongest predictors of Aβ positivity were age (β = 0.20), family history with both parents diagnosed with dementia (β = 0.18), and subjective memory (β = 0.14) and word-finding (β = 0.12) difficulties (Table 2). When APOE ε4 was included, carrying one (β = 0.59) or two (β = 0.44) APOE ε4 risk allele(s) became the strongest predictors, followed by age (β = 0.27), family history (β = 0.12), and subjective memory (β = 0.13) and word-finding (β = 0.08) difficulties. The predictor selection was consistent with the results of multivariate logistic regressions (Table S4 in support ing information) showing that all predictors with small (β > 0.05) to medium (β > 0.10) LASSO weights were significantly related to Aβ status.

Table 2. Standardized LASSO weights for Aβ prediction models.

Basic model Extended model Basic model with APOE4 Extended model with APOE4
Age 0.22 0.20 0.32 0.27
APOE, one e4 allele - - 0.62 0.59
APOE, two e4 alleles - - 0.46 0.44
Female - 0.03 - Not selected
Education, intermediate - -0.04 - Not selected
Education, higher - Not selected - Not selected
Married - 0.05 - Not selected
Family history, one parent had dementia - 0.04 - -0.02
Family history, both parents had dementia - 0.18 - 0.12
MMSE - -0.03 - -0.02
Delayed word-learning test - -0.09 - -0.05
Digit-symbol substitution test - -0.08 - -0.05
Subjective memory difficulty - 0.14 - 0.13
Subjective word-finding difficulty - 0.12 - 0.08
Need for assistance with finances or medication - 0.01 - Not selected
BMI - Not selected - Not selected
Aerobic exercise, hours/week - 0.02 - Not selected
Walking, minutes/day - -0.03 - -0.02
Sleep duration, hours/night - -0.06 - -0.04
Napping during the day - -0.09 - -0.06
Current smoking - 0.03 - Not selected
Alcohol, glasses/day - Not selected - Not selected
Open in a new tab

Abbreviations: APOE, Apolipoprotein E; BMI, Body Mass Index; MMSE, Mini Mental-State Examination; ‘-', Not included

3.3. Internal validity and added classification performance of the Aβ prediction models

Calibration plots are presented in Figure S2 in supporting information. The calibration slopes of the extended models were close to the target value of one in both the A4 training and test sets (range: 0.89– 1.06) suggesting that the predicted risks were not extreme (e.g., not too high for participants at high risk or not too low for participants at low risk). The calibration intercepts were all slightly negative (−0.82 to −0.85) implying that both models had a small tendency to overestimate the risk of Aβ positivity in all participants. Classification performance is shown in Table 3. The extended model showed moderate predictive performance with an AUC of 0.62 [95% CIs: 0.60–0.64] in the A4 training set and 0.61 [0.57–0.65]) in the A4 test set, which was higher relative to the performance of a basic model including only age (A4 training set: AUC = 0.56 [0.54–0.58]; A4 test set: AUC = 0.58 [0.54–0.63]). The extended model with APOE ε4 reached an AUC equal to 0.73 [0.71–0.75] in the A4 training set and to 0.73 [0.69–0.76] in the A4 test set. The improvement relative to the basic model with APOE ε4 was marginal (A4 training set: AUC = 0.71 [0.69–0.73]; A4 test set: AUC = 0.72 [0.68–0.76]).

Table 3. Performance of the Aβ prediction models in the training, test and external validation datasets.

Model n Prevalence Sensitivity Specificity PPV NPV Accuracy AUC [95%CI]
A4 training set
   Basic model 3296 0.30 0.48 0.61 0.34 0.73 0.54 0.56 [0.54-0.58]
   Extended model 3296 0.30 0.57 0.60 0.38 0.77 0.59 0.62 [0.60-0.64]
   Basic model with APOE4 3296 0.30 0.63 0.70 0.47 0.82 0.66 0.71 [0.69-0.73]
   Extended model with APOE4 3296 0.30 0.64 0.71 0.48 0.82 0.67 0.73 [0.71-0.75]
A4 test set
   Basic model 823 0.30 0.48 0.67 0.38 0.75 0.57 0.58 [0.54-0.63]
   Extended model 823 0.30 0.55 0.62 0.38 0.76 0.58 0.61 [0.57-0.65]
   Basic model with APOE4 823 0.30 0.66 0.68 0.47 0.82 0.67 0.72 [0.68-0.76]
   Extended model with APOE4 823 0.30 0.63 0.70 0.47 0.82 0.67 0.73 [0.69-0.76]
Rotterdam Study - 2 years after PET
   Basic model 351 0.13 0.62 0.57 0.18 0.91 0.60 0.60 [0.50-0.69]
   Extended model 351 0.13 0.60 0.59 0.18 0.91 0.60 0.63 [0.54-0.71]
   Basic model with APOE4 351 0.13 0.78 0.75 0.32 0.96 0.77 0.82 [0.75-0.88]
   Extended model with APOE4 351 0.13 0.78 0.73 0.30 0.96 0.75 0.82 [0.76-0.88]
Rotterdam Study - 7 years before PET
   Basic model 500 0.16 0.64 0.59 0.23 0.90 0.62 0.63 [0.56-0.69]
   Extended model 500 0.16 0.56 0.61 0.21 0.88 0.59 0.63 [0.56-0.70]
   Basic model with APOE4 500 0.16 0.82 0.72 0.35 0.96 0.77 0.84 [0.79-0.88]
   Extended model with APOE4 500 0.16 0.82 0.72 0.35 0.96 0.77 0.85 [0.81-0.89]
Rotterdam Study - 12 years before PET
   Basic model 365 0.17 0.63 0.61 0.25 0.89 0.62 0.64 [0.56-0.71]
   Extended model 365 0.17 0.60 0.62 0.25 0.88 0.61 0.61 [0.53-0.69]
   Basic model with APOE4 365 0.17 0.78 0.73 0.38 0.94 0.75 0.82 [0.76-0.87]
   Extended model with APOE4 365 0.17 0.75 0.74 0.37 0.93 0.74 0.82 [0.76-0.87]
Open in a new tab

Abbreviations: AUC, Area Under the Curve; CI, 95% confidence intervals; NPV, Negative Predictive Value; PPV, Positive Predictive Value.

3.4. External validity and temporal stability of the Aβ prediction models

External validity of the Aβ prediction models was tested in the Rotterdam Study using predictors collected 7 years before PET (Table 3). While the relative predictive accuracy across the different models was similar to that in the A4 Study, the absolute accuracy was higher in the Rotterdam Study indicating high external validity. For the extended model, AUC increased from 0.61 [0.57–0.65] in the A4 test set to 0.63 [0.56–0.70] in the Rotterdam Study. For the extended model with APOE ε4, performance improved considerably from an AUC of 0.73 [0.69– 0.76] in the A4 test set to 0.85 [0.81–0.89] in the Rotterdam Study. Table S5 in supporting information shows the performance at different probability thresholds. ROC curves are plotted in Figure 2. The models performed robustly across the three Rotterdam Study visits including predictors that were collected at three different timepoints (Figure S3 in supporting information). AUC values ranged from 0.61 to 0.63 for the extended model and from 0.82 to 0.85 for the extended model with APOE ε4 (Table 3). A supplementary analysis, in which we only included participants with complete data across the three visits (n = 178), yielded identical AUCs for the visits 12 and 7 years before PET, but a slightly higher AUC for the visit 2 years after PET (Table S6 in supporting information).

Figure 2. Receiver Operating Characteristic (ROC) curves display the performance of the (A) models without APOE4 and (B) models with APOE4 in the A4 Study test dataset and in the Rotterdam Study dataset used for external validation.

Figure 2

Open in a new tab

Finally, as a proof of concept, we estimated how many individuals from the general population (age range 60–90 years) would need to undergo PET imaging to find one amyloid-positive case (Table 4). When no prediction model is used, 8.1 PET scans would have to be acquired. This number was calculated as the inverse Aβ prevalence, which was estimated to be 18.9% in non-demented individuals similar to the whole Rotterdam Study cohort.24 By applying our extended model with APOE ε4 before PET, this number could be reduced to 6.5 PET scans (at a probability threshold set to achieve 90% sensitivity) or 4.1 PET scans (at a probability threshold set to achieve 90% specificity). For our extended model without APOE ε4, we estimated that 8.0 PET scans (at 90% sensitivity) or 6.7 PET scans (at 90% specificity) would be necessary. The numbers for the prediction models were calculated as the inverse positive predictive values which the models achieved in the Rotterdam Study validation dataset, assuming again an Aβ prevalence of 18.9%.

Table 4. Number of cognitively unimpaired participants necessary to undergo PET imaging to find one amyloid-positive case.

No prediction model1 Prediction model2 with APOE4 Prediction model2 without APOE4
Probability threshold3 - ≥0.42 ≥0.5 ≥0.68 ≥0.38 ≥0.5 ≥0.62
Number needed to scan 8.1 6.5 5.9 4.1 8.0 7.5 6.7
Open in a new tab
1

When no prediction model is applied the number to be scanned for identifying one amyloid-positive case was calculated as the inverse of the estimated Aβ prevalence. We assumed an average Aβ prevalence of 18.9% in cognitively unimpaired individuals aged between 60 and 90 years (this is an adjusted prevalence estimate we previously computed using an inverse probability weighting approach to adjust the proportion of amyloid-positive participants to the characteristics (age, sex, education and APOE e4 allele count) of all Rotterdam Study participants alive at the start of the PET study; see25, Table S2).

2

When our Extended models with and without APOE4 are applied the number to be scanned was calculated as the inverse of the positive prediction value (i.e. the likelihood to be truly amyloid-positive after a positive screen) derived for three different probability thresholds and assuming again a prevalence of 18.9%.

3

The probability thresholds were chosen to achieve at least 90% sensitivity (at ~40% probability), balanced sensitivity and specificity (at 50% probability) or 90% specificity (at 60-70% probability).

4. Discussion

In the current study, we developed two Aβ prediction models, one without and one with APOE ε4, based on the A4 Study, the largest amyloid PET study conducted to date (n = 4,119). When APOE ε4 was not considered, easily ascertainable predictors, such as family history of dementia or subjective cognitive complaints, improved predictive accuracy (AUC = 0.61) compared to a basic model including only age (AUC = 0.58). When APOE ε4 status was included, these predictors did not considerably increase predictive accuracy compared to using age and APOE ε4 only (AUC of 0.73 vs. 0.72). Importantly, these findings were validated in the prospective population-based Rotterdam Study (n = 500) with higher accuracy (e.g., AUC increased from 0.73 to 0.85).

Economic models only including predictors that are readily available in the clinical routine (without APOE ε4) have not achieved AUCs above 0.70.8,9,34–36 Two studies that also developed prediction models in the A4 Study classified amyloid-positive cases with an AUC of 0.61 (based on age, family history, BMI, free recall)19 or of 0.62 (based on age, education, sex, family history, activity of daily living, cognitive status [Cogstate, Cognitive Function Index, Preclinical Alzheimer Cognitive Composite]).20 Because our extended model’s performance (AUC of 0.61) was highly consistent with these reports, the inclusion of novel predictors, such as sleep duration,37 did not appear to aid predictive performance. Including APOE ε4 genotype improved prediction performance above AUCs of 0.7 in most prior work including the current and other A4-based studies (AUC = 0.73 in Petersen et al.19 and in current study or AUC = 0.74 in Langford et al.20). Our results further suggest that other readily ascertainable predictors did not increase predictive accuracy significantly beyond the strong effect of APOE ε4. Likewise, no considerable improvement above APOE ε4 was found for MMSE and objective memory performance in the Amyloid Biomarker Study (n = 2,908)38 or for subjective cognitive decline in the Harvard Aging Brain Study, ADNI, and Australian Imaging Biomarker and Lifestyle (AIBL) study (n = 890).35

To more accurately predict Aβ status, more sophisticated predictors are probably necessary. Structural MRI and blood-based markers (Aβ42/40, phosphorylated tau181), for example, helped to reach AUCs above 0.8 in multiple,9–11,39,40 but not all, previous studies.41,42 Because these models were developed in relatively small and highly selected patient samples and often lacked external validation, their performance in the wider population has yet to be determined (for first population-based data see Mielke et al.43). Although available, we decided not to include MRI in our models, because imaging is burdensome and expensive and therefore of limited use for screening purposes. Plasma biomarkers, on the other hand, were not available in the current cohorts, but seem to be promising minimally invasive predictors of Aβ positivity if inconsistencies in sample handling and untransparent usage of in-house assays are overcome.44 We are planning to enrich the Rotterdam Study with plasma biomarkers soon to then validate corresponding Aβ prediction models.

The key strength of this study was that we externally validated our developed models in an independent sample. The differences in sample characteristics between the A4 Study and the Rotterdam Study (multi-centric cross-sectional assessment of a convenience sample from North America, Australia, Japan versus mono-centric prospective assessment of a population-based sample from Northern Europe) allowed us to thoroughly determine the models’ performance across different populations. Somewhat unexpectedly, predictive performance was similar (model without APOE ε4) or higher (model with APOE ε4) in the population-based validation dataset indicating that Aβ prediction models can be applied to a broader population than the one in which they were developed. We can only speculate about what might have caused this performance boost. One explanation is that an accumulation of various genetic factors (other than APOE ε4) and/or environmental factors related to the high prevalence of family history of dementia in the A4 Study may have underestimated the predictive power of APOE ε4, while APOE ε4 is the main driver of Aβ in an unselected sample like the Rotterdam Study. One previous study that tested external validity in a population-based sample also found robust performance. The best-performing model (including age, APOE ε4, memory performance) reached an AUC of 0.75 and 0.72 in the clinical training cohorts (ADNI, AIBL) and 0.71 in the Mayo Clinic Study of Aging (MCSA) validation cohort.18 This result was similar to the performance of the best model developed directly in the MCSA cohort (AUC = 0.70; based on age, APOE ε4, family history, and subjective cognitive difficulties45), but lower than the performance observed in the current validation dataset (AUC = 0.85).

To our best knowledge the current study is one of the first to estimate the stability of Aβ prediction models over time. We found robust performance using predictors collected at three different timepoints before and after PET acquisition. This was not surprising for the models including static APOE ε4 status. However, even the model without APOE ε4, which contained comparably strong static (family history) and dynamic predictors (subjective memory difficulty), showed high temporal stability with a slight superiority for the timepoint closest to PET. Future studies should confirm whether Aβ positivity can be predicted with a time difference of up to 12 years as suggested by the current results.42

We suggest two scenarios in which Aβ prediction models may be useful in a general population setting: screening for clinical AD trials and in primary care. Clinical trials increasingly move toward the inclusion of asymptomatic subjects, because treatment may be more effective before notable cognitive impairment and brain damage have occurred. In trials which aim to include only amyloid-positive individuals (and thus would require a high specificity), prediction models could reduce the number of unnecessary (negative) PET scans. We calculated that half the number of PET scans (4.1 instead of 8.1) would be necessary for identifying one amyloid-positive individual, when applying our best performing model (extended model with APOE ε4) in individuals similar to the Rotterdam Study. In contrast, in a future scenario in which primary care would like to identify individuals for early disease management, this would require high sensitivity to miss as few amyloid-positive individuals as possible. Here, a prediction model could increase confidence of primary health-care providers to refer a patient to a specialized clinic. Such selective referrals may become even more critical in the future considering an increasing number of older adults and likely more approved treatments against AD that require confirmatory testing of underlying AD pathology as a first step.46

The current study has several limitations. First, not all predictors were measured in identical ways across the two cohorts, with the largest mismatch occurring between the different delayed recall tests. Misestimation of Aβ risk could be a possible consequence but should be marginal given the relatively small contribution of memory performance to Aβ prediction. Second, the Rotterdam Study validation sample was not free of selection or nonparticipation bias, which should be considered when interpreting the results. Third, although our models performed comparably well relative to previous models, the absolute performance was still insufficient for clinical use. Fourth, it is likely that blood-based biomarkers could have improved prediction, but they were not available in the two cohorts. Finally, although we involved two independent studies in geographically diverse populations, most participants were non-Latinx White and highly educated, and it therefore remains crucial to further validate the resulting models in other ethnocultural groups and more diverse educational backgrounds.

In conclusion, we confirmed that Aβ prediction models can be generalized to a population with very different characteristics than the convenience sample in which they were developed and which, importantly, was more representative of typical older non-demented adults.

Supplementary Material

Table S1 to S6, Figures S1 to S3, Methods S1 to S2

Acknowledgements

The A4 Study is a secondary prevention trial in preclinical Alzheimer's disease, aiming to slow cognitive decline associated with brain amyloid accumulation in clinically normal older individuals. The A4 Study is funded by a public-private-philanthropic partnership, including funding from the National Institutes of Health-National Institute on Aging, Eli Lilly and Company, Alzheimer's Association, Accelerating Medicines Partnership, GHR Foundation, an anonymous foundation and additional private donors, with in-kind support from Avid and Cogstate. The companion observational Longitudinal Evaluation of Amyloid Risk and Neurodegeneration (LEARN) Study is funded by the Alzheimer's Association and GHR Foundation. The A4 and LEARN Studies are led by Dr. Reisa Sperling at Brigham and Women's Hospital, Harvard Medical School and Dr. Paul Aisen at the Alzheimer's Therapeutic Research Institute (ATRI), University of Southern California. The A4 and LEARN Studies are coordinated by ATRI at the University of Southern California, and the data are made available through the Laboratory for Neuro Imaging at the University of Southern California. The participants screening for the A4 Study provided permission to share their de-identified data in order to advance the quest to find a successful treatment for Alzheimer's disease. We would like to acknowledge the dedication of all the participants, the site personnel, and all of the partnership team members who continue to make the A4 and LEARN Studies possible. The complete A4 Study Team list is available on: a4study.org/a4-study-team.

We would like to thank the entire staff of the Nuclear Medicine department of Erasmus Medical Center, for their help in acquiring the amyloid PET data in the Rotterdam Study, including but not limited to Dennis Kuijper, Annelies Schipper, Pieter Meppelink and Jean-Baptiste Aarssen for their coordinating roles. We would also like to acknowledge the immense contribution of the data management team of the Rotterdam Study, with Jolande Verkroost-van Heemst in particular, and of the Imaging Trialbureau. Lastly, we would like to thank the Rotterdam Study participants for their contribution.

Study funding

This project has received funding from the Alzheimer’s Association Research, Grant/Award Number: AARG-22-972229; ZonMw Memorabel, Grant/Award Number: 733050817; European Union’s Horizon 2020 research and innovation programme, Grant/Award Number: 101032288; ZonMW, Grant/Award Numbers: #73305095007, #10510032120003; Health Holland, Topsector Life Sciences & Health (PPP-allowance), Grant/Award Number: #LSHM20106.

Abbreviations

A4

Anti-Amyloid Treatment in Asymptomatic Alzheimer Study

Aβ

amyloid-beta

AD

Alzheimer’s disease

ADNI

Alzheimer’s Disease Neuroimaging Initiative

AIBL

Australian Imaging, Biomarker & Lifestyle Flagship Study of Ageing

APOE

Apolipoprotein E (gene)

APOE4

Apolipoprotein E ε4 allele

AUC

Area Under the Curve

BMI

Body Mass Index

CFI

Cognitive Function Index

CI

Confidence Interval

CSF

Cerebrospinal Fluid

LASSO

Least Absolute Shrinkage and Selection Operator

MCSA

Mayo Clinic Study of Aging

MMSE

Mini-Mental State Examination

MRI

Magnetic Resonance Imaging

PACC

Preclinical Alzheimer Cognitive Composite

PET

Positron Emission Tomography

ROC

Receiver Operating Characteristic

RS

Rotterdam Study

SUVr

Standard Uptake Value ratio

Footnotes

Author contributions

Julia Neitzel and Phuong Thuy Nguyen Ho developed the study design. Joyce van Arendonk, Rebecca Steketee and Frank van Rooij led data acquisition and management. Phuong Thuy Nguyen Ho and Julia Neitzel performed the statistical analysis and drafted the manuscript. All authors contributed to the interpretation of the results, critically revised the manuscript, and approved the final draft of this report.

Conflicts of interest

The authors report no competing interests.

Standard Protocol Approvals, Registrations, and Patient Consents

The A4 Study was approved by the institutional review boards of all participating institutions (NCT02008357). The Rotterdam Study has been approved by the Medical Ethics Committee of the Erasmus MC (MEC-2018-085) and by the Dutch Ministry of Health, Welfare and Sport (Population Screening Act WBO, license number 1071272-159521-PG). Written informed consent was obtained from all participants. This study followed the TRIPOD guidelines for reporting prognostic models 46.

Data Availability

A4 data used in this article are available for download from the Laboratory of NeuroImaging (LONI; loni.usc.edu). Rotterdam Study data can be obtained upon request. Requests should be directed towards the management team of the Rotterdam Study (secretariat.epi@erasmusmc.nl), which has a protocol for approving data requests. Because of restrictions based on privacy regulations and informed consent of the participants, data cannot be made freely available in a public repository. The Rotterdam Study has been entered into the Netherlands National Trial Register (NTR; www.trialregister.nl) and into the WHO International Clinical Trials Registry Platform (ICTRP; www.who.int/ictrp/network/primary/en/) under shared catalogue number NTR6831.

References

  • 1.Gauthier S, Rosa-Neto P, Morais JA, Webster C. World Alzheimer Report 2021: Journey through the Diagnosis of Dementia. 2021
  • 2.Price JL, Morris JC. Tangles and plaques in nondemented aging and “preclinical” Alzheimer’s disease. Ann Neurol. 1999;45(3):358–368. doi: 10.1002/1531-8249(199903)45:3<358::aid-ana12>3.0.co;2-x. [DOI] [PubMed] [Google Scholar]
  • 3.Jack CR, Jr, Bennett DA, Blennow K, et al. Contributors. NIA-AA research framework: toward a biological definition of Alzheimer’s disease. Alzheimers Dement. 2018;14(4):535–562. doi: 10.1016/j.jalz.2018.02.018. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4.Jack CR, Knopman DS, Jagust WJ, et al. Tracking pathophysiological processes in Alzheimer’s disease: an updated hypothetical model of dynamic biomarkers. Lancet Neurol. 2013;12(2):207–216. doi: 10.1016/S1474-4422(12)70291-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.van Dyck CH, Swanson CJ, Aisen P, et al. Lecanemab in early Alzheimer’s disease. N Engl J Med. 2023;388(1):9–21. doi: 10.1056/NEJMoa2212948. [DOI] [PubMed] [Google Scholar]
  • 6.Alexander GC, Knopman DS, Emerson SS, et al. Revisiting FDA approval of Aducanumab. N Engl J Med. 2021;385(9):769–771. doi: 10.1056/NEJMp2110468. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Ashford MT, Veitch DP, Neuhaus J, Nosheny RL, Tosun D, Weiner MW. The search for a convenient procedure to detect one of the earliest signs of Alzheimer’s disease: a systematic review of the prediction of brain amyloid status. Alzheimer’s Dementia. 2021;17(5):866–887. doi: 10.1002/alz.12253. [DOI] [PubMed] [Google Scholar]
  • 8.Insel PS, Palmqvist S, Mackin RS, et al. Assessing risk for preclinical β-amyloid pathology with APOE, cognitive, and demographic infor-mation. Alzheimer’s Dementia: DADM. 2016;4:76–84. doi: 10.1016/j.dadm.2016.07.002. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Palmqvist S, Insel PS, Zetterberg H, et al. Accurate risk estimation of β-amyloid positivity to identify prodromal Alzheimer’s disease: cross-validation study of practical algorithms. Alzheimer’s Dementia. 2019;15(2):194–204. doi: 10.1016/j.jalz.2018.08.014. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Tosun D, Veitch D, Aisen P, et al. Detection of β-amyloid positiv-ity in Alzheimer’s Disease Neuroimaging Initiative participants with demographics, cognition, MRI and plasma biomarkers. Brain Commun. 2021;3(2):fcab008. doi: 10.1093/braincomms/fcab008. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.Burnham SC, Faux NG, Wilson W, et al. A blood-based predictor for neocortical Aβ burden in Alzheimer’s disease: results from the AIBL study. Mol Psychiatry. 2014;19(4):519–526. doi: 10.1038/mp.2013.40. [DOI] [PubMed] [Google Scholar]
  • 12.Ezzati A, Harvey DJ, Habeck C, et al. Predicting Amyloid-β levels in amnestic mild cognitive impairment using machine learning tech-niques. J Alzheimers Dis. 2020;73(3):1211–1219. doi: 10.3233/JAD-191038. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13.Whitwell JL. Comparison of imaging biomarkers in the Alzheimerdis-ease neuroimaging initiative and the Mayo Clinic Study of Aging. Arch Neurol. 2012;69(5):614. doi: 10.1001/archneurol.2011.3029. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Doody RS, Thomas RG, Farlow M, et al. Alzheimer’s disease cooper-ative study steering committee; solanezumab study group. Phase 3 trials of solanezumab for mild-to-moderate Alzheimer’s disease. N Engl J Med. 2014;370(4):311–321. doi: 10.1056/NEJMoa1312889. [DOI] [PubMed] [Google Scholar]
  • 15.Brodaty H, Mothakunnel A, de Vel-Palumbo M, et al. Influence of population versus convenience sampling on sample characteristics in studies of cognitive aging. Ann Epidemiol. 2014;24(1):63–71. doi: 10.1016/j.annepidem.2013.10.005. [DOI] [PubMed] [Google Scholar]
  • 16.Gianattasio KZ, Bennett EE, Wei J, et al. Generalizability of findings from a clinical sample to a community-based sample: a comparison of ADNI and ARIC. Alzheimer’s & Dementia. 2021;17(8):1265–1276. doi: 10.1002/alz.12293. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Kokmen E, Özsarfati Y, Beard CM, O’Brien PC, Rocca WA. Impact of referral bias on clinical and epidemiological studies of Alzheimer’s disease. J Clin Epidemiol. 1996;49(1):79–83. doi: 10.1016/0895-4356(95)00031-3. [DOI] [PubMed] [Google Scholar]
  • 18.Maserejian N, Bian S, Wang W, et al. Practical algorithms for amyloid β probability in subjective or mild cognitive impairment. Alzheimer’s Dementia: DADM. 2019;11(1):710–720. doi: 10.1016/j.dadm.2019.09.001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Petersen KK, Lipton RB, Grober E, Davatzikos C, Sperling RA, Ezzati A. Predicting amyloid positivity in cognitively unimpaired older adults: a machine learning approach using A4 data. Neurology. 2022;98(24):E2425–E2435. doi: 10.1212/WNL.0000000000200553. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Langford O, Raman R, Sperling RA, et al. Predicting amyloid burden to accelerate recruitment of secondary prevention clinical trials. J Prev Alzheimers Dis. 2020;7(4):213–218. doi: 10.14283/jpad.2020.44. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21.Sato K, Ihara R, Suzuki K, et al. Predicting amyloid risk by machine learning algorithms based on the A4 screen data: application to the Japanese Trial-Ready Cohort study. Alzheimers Dement (N Y) 2021;7(1):e12135. doi: 10.1002/trc2.12135. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Sperling RA, Donohue MC, Raman R, et al. Association of factors with elevated amyloid burden in clinically normal older individuals. JAMA Neurol. 2020;77(6):735. doi: 10.1001/jamaneurol.2020.0387. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Ikram MA, Brusselle G, Ghanbari M, et al. Objectives, design and main findings until 2020 from the Rotterdam Study. Eur J Epidemiol. 2020;35(5):483–517. doi: 10.1007/s10654-020-00640-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24.van Arendonk J, Neitzel J, Steketee RME, et al. Diabetes and hyper-tension are related to amyloid-beta burden in the population-based Rotterdam Study. Brain. 2023;146(1):337–348. doi: 10.1093/brain/awac354. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Pontecorvo MJ, Arora AK, Devine M, et al. Quantitation of PET sig-nal as an adjunct to visual interpretation of florbetapir imaging. Eur J Nucl Med Mol Imaging. 2017;44(5):825–837. doi: 10.1007/s00259-016-3601-4. [DOI] [PubMed] [Google Scholar]
  • 26.Johnson KA, Sperling RA, Gidicsin CM, et al. Florbetapir (F18-AV-45) PET to assess amyloid burden in Alzheimer’s disease dementia, mild cognitive impairment, and normal aging. Alzheimer’s Dementia. 2013;9(5S) doi: 10.1016/j.jalz.2012.10.007. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Royse SK, Minhas DS, Lopresti BJ, et al. Validation of amyloid PET positivity thresholds in centiloids: a multisite PET study approach. Alzheimers Res Ther. 2021;13(1):99. doi: 10.1186/s13195-021-00836-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28.Deters KD, Napolioni V, Sperling RA, et al. Amyloid PET imag-ing in self-identified non-hispanic black participants of the anti-amyloid in asymptomatic Alzheimer’s disease (A4) Study. Neurology. 2021;96(11):e1491–e1500. doi: 10.1212/WNL.0000000000011599. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29.Grober E, Buschke H. Genuine memory deficits in dementia. Dev Neuropsychol. 1987;3(1):13–36. doi: 10.1080/87565648709540361. [DOI] [Google Scholar]
  • 30.Hoogendam YY, Hofman A, van der Geest JN, van der Lugt A, Ikram MA. Patterns of cognitive function in aging: the Rotterdam Study. Eur J Epidemiol. 2014;29(2):133–140. doi: 10.1007/s10654-014-9885-4. [DOI] [PubMed] [Google Scholar]
  • 31.Stel VS, Smit JH, Pluijm Saskia MF, Visser M, Deeg DJH, Lips P. Comparison of the LASA Physical Activity Questionnaire with a 7-day diary and pedometer. J Clin Epidemiol. 2004;57(3):252–258. doi: 10.1016/j.jclinepi.2003.07.008. [DOI] [PubMed] [Google Scholar]
  • 32.Pavlou M, Ambler G, Seaman SR, et al. How to develop a more accurate risk prediction model when there are few events. BMJ. 2015;351:h3868. doi: 10.1136/bmj.h3868. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33.Chawla Nv, Bowyer KW, Hall LO, Kegelmeyer WP. SMOTE: synthetic Minority Over-sampling Technique. J Artificial Intelligence Res. 2002;16:321–357. [Google Scholar]
  • 34.Ansart M, Epelbaum S, Gagliardi G, et al. Prediction of Amyloidosis from neuropsychological and mri data for cost effective inclusion of pre-symptomatic subjects in clinical trials. Springer; 2017. Deep learning in medical image analysis and multimodal learning for clinical decision sup-port; pp. 357–364. DLMIA ML-CDS 2017 2017. [DOI] [Google Scholar]
  • 35.Buckley RF, Sikkes S, Villemagne VL, et al. Using subjective cognitive decline to identify high global amyloid in community-based samples: a cross-cohort study. Alzheimer’s Dementia: Diagnosis, Assessment Disease Monitoring. 2019;11(1):670–678. doi: 10.1016/j.dadm.2019.08.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Lee JH, Byun MS, Yi D, et al. Prediction of cerebral amyloid with com-mon information obtained from memory clinic practice. Front Aging Neurosci. 2018;10 doi: 10.3389/fnagi.2018.00309. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37.Winer JR, Deters KD, Kennedy G, et al. Association of short and long sleep duration with amyloid-β burden and cognition in aging. JAMA Neurol. 2021;78(10):1187. doi: 10.1001/jamaneurol.2021.2876. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38.Jansen WJ, Ossenkoppele R, Tijms BM, et al. Association of cere-bral amyloid-β aggregation with cognitive functioning in persons without dementia. JAMA Psychiatry. 2018;75(1):84. doi: 10.1001/jamapsychiatry.2017.3391. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Verberk IMW, Slot RE, Verfaillie SCJ, et al. Plasma amyloid as pre-screener for the earliest Alzheimer pathological changes. Ann Neurol. 2018;84(5):648–658. doi: 10.1002/ana.25334. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Petrone PM, Casamitjana A, Falcon C, et al. Prediction of amy-loid pathology in cognitively unimpaired individuals using voxel-wise analysis of longitudinal structural brain MRI. Alzheimers Res Ther. 2019;11(1):72. doi: 10.1186/s13195-019-0526-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.ten Kate M, Redolfi A, Peira E, et al. MRI predictors of amyloid pathology: results from the EMIF-AD Multimodal Biomarker Discov-ery study. Alzheimers Res Ther. 2018;10(1):100. doi: 10.1186/s13195-018-0428-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42.de Rojas I, Romero J, Rodríguez-Gomez O, et al. Correlations between plasma and PET beta-amyloid levels in individuals with subjective cog-nitive decline: the Fundació ACE Healthy Brain Initiative (FACEHBI) Alzheimers Res Ther. 2018;10(1):119. doi: 10.1186/s13195-018-0444-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43.Mielke MM, Dage JL, Frank RD, et al. Author Correction: performance of plasma phosphorylated tau 181 and 217 in the community. Nat Med. 2022 doi: 10.1038/s41591-022-01822-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Teunissen CE, Verberk IMW, Thijssen EH, et al. Blood-based biomark-ers for Alzheimer’s disease: towards clinical implementation. Lancet Neurol. 2022;21(1):66–77. doi: 10.1016/S1474-4422(21)00361-6. [DOI] [PubMed] [Google Scholar]
  • 45.Mielke MM, Wiste HJ, Weigand SD, et al. Indicators of amyloid burden in a population-based study of cognitively normal elderly. Neurology. 2012;79(15):1570–1577. doi: 10.1212/WNL.0b013e31826e2696. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46.Liu JL, Hlavka JP, Hillestad R, Mattke S. Assessing the Preparedness of the US Health Care System Infrastructure for an Alzheimer’s Treatment. 2017. https://www.rand.org/pubs/research_reports/RR2272.html . [PMC free article] [PubMed]
  • 47.Collins GS, Reitsma JB, Altman DG, Moons KG. Transparent reporting of a multivariable prediction model for individual prognosis or diagno-sis (TRIPOD): the TRIPOD statement. Ann Intern Med. 2015;162(1):55–63. doi: 10.7326/M14-0697. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Table S1 to S6, Figures S1 to S3, Methods S1 to S2
EMS177420-supplement-Table_S1_to_S6__Figures_S1_to_S3__Methods_S1_to_S2.docx (1.5MB, docx)

Data Availability Statement

A4 data used in this article are available for download from the Laboratory of NeuroImaging (LONI; loni.usc.edu). Rotterdam Study data can be obtained upon request. Requests should be directed towards the management team of the Rotterdam Study (secretariat.epi@erasmusmc.nl), which has a protocol for approving data requests. Because of restrictions based on privacy regulations and informed consent of the participants, data cannot be made freely available in a public repository. The Rotterdam Study has been entered into the Netherlands National Trial Register (NTR; www.trialregister.nl) and into the WHO International Clinical Trials Registry Platform (ICTRP; www.who.int/ictrp/network/primary/en/) under shared catalogue number NTR6831.

RESOURCES