The Atherosclerosis Risk in Communities (ARIC) study is a population-based prospective study of 15,792 men and women aged 45–64 years at baseline (1987–1989) from four U.S. communities: Washington County, Maryland; Forsyth County, North Carolina; Jackson, Mississippi; and Minneapolis, Minnesota. Between baseline and the year 2013, participants completed 4 additional in-person visits. These visits were scheduled between midlife, in the years 1990–92 (visit 2), 1993–95 (visit 3) and 1996–98 (visit 4), and late life in the years 2011–13 (visit 5).

At visit 5 (late-life; ages 69–89 years), 6538 surviving participants underwent an extensive neuropsychological battery, as part of the ARIC Neurocognitive Study, and individuals were categorized as having normal cognitive function, mild cognitive impairment, or dementia [17]. A subset of the cohort was also selected for research brain MRI. Of those participants included in the MRI imaging sub study, 346 participants from 3 ARIC sites (Jackson, Mississippi; Washington County, Maryland; and Forsyth County, North Carolina) were also recruited for the ancillary ARIC-PET study as previously described [18]. Only individuals without an initial dementia diagnosis, heavy alcohol use, renal dysfunction or prolonged QT-c interval were included. 289 and 226 individuals had sufficient high-quality retinal imaging data to characterize the primary retinal measure of interest, i.e., retinopathy vs. no retinopathy, at visit 3 and visit 5 respectively. Additional participants at visit 3 (N = 4) and visit 5 (N = 6) were excluded if the diabetes or hypertension measure was missing (Fig. 1). Study procedures were approved by the Institutional Review Board for each field center.

Flow diagram for inclusion in the ARIC-PET retinopathy analytic sample. Cases with a valid retinopathy measure but with missing arteriovenous nicking and generalized arteriolar narrowing measures were imputed. At visit 5, the missing APOE-4 status (N = 2) was imputed to retain the maximum number of positive retinopathy cases. aCases imputed at visit 3: Arteriovenous nicking: N = 2; Generalized arteriolar narrowing: N = 16. bCases imputed at visit 5: Arteriovenous nicking: N = 2; Generalized arteriolar narrowing: N = 16; APOE-4 status: N = 2

Brain MRIs performed on a 3 T magnet and Florbetapir PET scans were carried out at each ARIC site, within 12 months but ideally within 6 months of each other. Participants without a contraindication of MRI were eligible for the MRI imaging substudy when meeting one of the following criteria: [1] prior brain MRI during an earlier ARIC brain MRI ancillary visit; [2] low scores of cognitive function on an extensive neuropsychological battery or a marked decline on cognition on tests repeated at ARIC visits 2, 4, and 5; or [3] an age-stratified random sample of individuals with normal cognitive function [17]. For the florbetapir PET scans, an isotope was injected through a butterfly needle and images were obtained from 50–70 min for a 20-min uptake scan. PET scans were centrally reviewed and quantified at the PET image analysis center (Johns Hopkins); brain MRI scans, all performed on a 3 T magnet at facilities associated with each field center, were centrally reviewed at the MRI reading center (Mayo Clinic). PET scans were quantified for standardized uptake value ratios (SUVRs) in distinct regions of interest, and co-registered to the MRI images. A global cortical measure of Aβ was acquired by computing the weighted average of the following brain regions: orbitofrontal, prefrontal, and superior frontal cortices, lateral temporal, parietal, and occipital lobes, precuneus, and anterior and posterior cingulate. Based on the sample median, the SUVR was dichotomized at a value of 1.2. Elevated Aβ burden was defined as a global cortical SUVR greater than 1.2.

45° non-stereoscopic color retinal photographs of 2 fields (one centered on the optic nerve and one on the macula) in a single randomly selected eye at visit 3 and in both eyes at visit 5 were acquired by trained technicians using nonmydriatic fundus cameras (Canon CR-45UAF; Canon CR-1 Mark II). At both visits, photographs were graded by certified graders masked to the participants’ characteristics at the Ocular Epidemiology Reading Center (OREC) at the University of Wisconsin-Madison based on written standardized protocol and digital photographic standards. Further details on the retinal image acquisition have been described previously [19, 20]. Four measures of retinal microvascular abnormalities were included: [1] retinopathy; [2] arteriovenous nicking; [3] focal arteriolar narrowing; [4] generalized arteriolar narrowing. The presence of retinopathy was determined employing the modified Airlie House classification, as used in the Early Treatment Diabetic Retinopathy Study (ETDRS) [19]. Retinopathy was classified as present (level of 14–87) vs. absent [10,11,12,13]. Retinopathy levels between 14–87 refer to a wide spectrum of vascular abnormalities in the eye which include microaneurysms, hemorrhages, soft and hard exudates, venous beading, neovascularization, and fibrosis. Arteriovenous nicking was seen as “definite” in case at least one venous blood column was tapered on both sides of its crossing underneath an arteriole. Focal arteriolar narrowing was determined as “definite” based on the grading and number of arterioles estimated to be ≥ 50 μm in diameter with a constricted area ≤ 2/3 the width of proximal and distal vessel segments. Generalized arteriolar narrowing was defined as lowest quartile of the central retinal arteriolar equivalent (CRAE). CRAE is quantified on the arteriolar diameters within a pre-specified zone surrounding the optic nerve [21]. The reproducibility of the retinal measures was moderate to good (Cohen’s kappa = 0.45–0.89) [19]. The individual reproducibility estimates were as follows: arteriovenous nicking, kappa = 0.61; focal arteriolar narrowing, kappa = 0.45; retinopathy, kappa = 0.89, CRAE, correlation coefficient = 0.74.

Demographic information such as age, sex, race, and education were self-reported and collected at study baseline (1987–1989). Other variables included in this study were APOE-4 genotype (TaqMan assay; Applied Biosystems, Foster City, CA) as well as status of hypertension (measured systolic blood pressure > 140 mm Hg, diastolic blood pressure > 90 mm Hg or use of antihypertensive medications) and of diabetes (fasting glucose > 125, non-fasting > 200, or self-report of diabetes, diagnosed by a physician, or use of antidiabetic medication) at visit 3 and 5. Cognitive status was classified according to expert-adjudicated cognitive outcomes as described previously in detail [17]. Cognitive functions were categorized into normal, mild cognitive impairment, or dementia by a panel of experts according to standardized criteria [17]. As described previously in detail, a battery of neuropsychological test scores was used to compute a global cognition factor score at visit 5 [22].

The statistical analysis was carried out using R (version 3.6.2). The cohort’s characteristics were summarized by descriptive analysis overall and stratified by retinopathy for visit 3 and visit 5. Using multivariate imputation by chained equations (MICE), missing measures of arteriovenous nicking and generalized retinal arteriolar narrowing at both visits were imputed in those cases where a participant had a valid retinopathy measure, but these other measures were missing. APOE-4 status were imputed in 2 cases at visit 5 to retain the maximum number of positive retinopathy cases for the analysis [23].

A comprehensive retinal scoring system was created that summarizes the progressive manifestations of retinal abnormalities in the vasoconstrictive, sclerotic, and exudative phases into a single measure. Whereas generalized arteriolar narrowing, focal arteriolar narrowing and arteriovenous nicking can be seen as early as in the vasoconstrictive and sclerotic phases, features of retinopathy such as tissue ischemia, microaneurysms, hemorrhages and exudates are observed only in the more advanced exudative phase [24]. The scoring system ranged from 0–3. A score of 3 was given to a participant if retinopathy as assessed by the modified Airlie House classification was evident or all other three retinal signs (i.e., arteriovenous nicking, focal arteriolar narrowing, generalized arteriolar narrowing) were present. Individuals were assigned a score of 2 or 1 when two or one retinal sign(s) other than retinopathy were present, respectively. Participants received a score of 0 if they had none of the 4 retinal microvascular features.

Recognizing that a retinal score version that assumes equivalent weighting of these microvascular retinal abnormalities may not be the optimal measure for assessing greater amyloid burden at visit 5, we also created a weighted score version in a sensitivity analysis in which the four retinal measures received different weights based on their association with amyloid burden. To derive the weights, a ridge logistic regression model was employed with the retinal measures as predictors and elevated amyloid burden (SUVR > 1.2) as the dichotomous outcome. To generalize the weights beyond the sample of the ARIC-PET study, the optimal lambda value of the regularized method was selected that minimized the cross-validation prediction error rate for elevated amyloid burden. The weighted score distribution was normalized and transformed to a whole number score range of 0–3 and these numbers were then assigned to each participant.

In the main analysis, the associations between the retinal measures and elevated Aβ burden were tested using separate logistic regression models. A two-step model building process for adjustment was employed. Covariates in model 1 included age, sex, education, APOE4 status and race. In model 2, the presence of hypertension and diabetes was added as additional confounders. We also ran a statistical power analysis to assess the probability of detecting a significant effect of the retinal measures on amyloid burden in the ARIC-PET study if it truly exists in the population.

To assess whether a significant survival bias exists among individuals with retinopathy, the proportions of retinopathy vs. no retinopathy cases at visit 3 stratified by mortality and dementia conversion by visit 5 were computed in the overall ARIC cohort seen at visit 3. It was furthermore tested whether participants who were later enrolled in the ARIC-PET study at visit 5, which specifically excluded individuals with dementia, had a healthier profile at visit 3 than those participants not being recruited for the ancillary study. This was done by comparing the clinical and demographic characteristics between the 2 groups using the Wilcoxon rank sum and Pearson’s Chi-squared tests.

In a secondary analysis employing a robust linear regression analysis with 2000 bootstrap replicates, we also assessed whether retinal signs and the composite scores at visit 5 were associated with increased amyloid SUVR, defined as a continuous outcome measure. The model was adjusted by the covariates age, sex, education, race, and APOE-4 status in model 1. In model 2, the clinical risk factors hypertension and diabetes were further added to the model. Grouped boxplots were also employed to visualize the distributions of amyloid SUVR by retinal signs and scores. We furthermore assessed the association between continuous CRAE and amyloid SUVR.