Skip to main content
NCBI home page
Alzheimer's & Dementia : Diagnosis, Assessment & Disease Monitoring logoLink to Alzheimer's & Dementia : Diagnosis, Assessment & Disease Monitoring
. 2026 May 25;18(2):e70368. doi: 10.1002/dad2.70368

Age‐related macular degeneration severity and risk of Alzheimer's disease in older adults in a prospective, community‐based cohort study

Chloe A Krakauer 1, Luke Reilly 2, Yu‐Ru Su 1,3, Rod L Walker 1, Dana Lee 1, Vilaipone A Xaiyamuangchanh 4, David J Cronkite 1, Nicole M Gatto 1, Paul K Crane 5, Emily Y Chew 6, David E Arterburn 1, Cecilia S Lee 7,✉
PMCID: PMC13239676  PMID: 42255955

Abstract

INTRODUCTION

Age‐related macular degeneration (AMD) is associated with higher risk of Alzheimer's disease (AD), but how AD risk changes as AMD progresses remains unclear.

METHODS

Among participants in the Adult Changes in Thought study, we characterized the worst dry and wet AMD severities experienced to date on a scale of 0 to 10 and assessed whether they were associated with AD risk.

RESULTS

Among 2611 participants (550 AD cases, median 6.8 years follow‐up), AD hazard was estimated to increase by 4% per one unit increase in dry and wet scores (dry: 95% confidence interval [CI] for hazard ratio [HR] 1.01 to 1.07, p = 0.01; wet: 95% CI 1.00 to 1.08, p = 0.03). HR estimates were similar among people without diagnosed AMD and were attenuated among those with diagnosed AMD.

DISCUSSION

Worse AMD severity was associated with higher AD risk, with severity scores potentially offering greater clinical utility for AD risk characterization among individuals without diagnosed AMD.

Keywords: age‐related macular degeneration, Alzheimer's disease, biomarker, clinical epidemiology, dementia, dry age‐related macular degeneration, electronic health records, ophthalmology, risk surveillance, wet age‐related macular degeneration

Highlights

  • Dry and wet AMD severity scores were derived from EHRs.

  • Higher (worse) scores (individual and together) were associated with greater AD risk.

  • AMD severity scores – using only EHRs – may be useful for AD risk surveillance.

  • Associations were strongest among people without diagnosed AMD, for example, subclinical.

  • Scores did not outperform binary indicators of AMD diagnosis in predicting AD risk.

1. BACKGROUND

Dementia, with Alzheimer's disease (AD) being the most common type, is a leading cause of morbidity in older adults. 1 , 2 With limited effective screening, prevention, and treatment modalities, advancing our understanding of biomarkers and modifiable risk factors for AD/all‐cause dementia is therefore essential to inform potential screening tools and improve public health. Assessments of eye diseases over time – given potential effects of eye diseases on AD/all‐cause dementia risk and shared etiologic pathways 3 – may act as useful AD/all‐cause dementia risk biomarkers. Proposed mechanisms of common etiology include shared underlying oxidative stress, protein aggregation, pro‐inflammatory states, 4 and microvascular endothelial senescence. 5 While vision loss was recently added to the list of modifiable risk factors for dementia in the 2024 Lancet Commission report 6 and prior studies identified associations between ophthalmic disease and dementia risk, 3 , 7 , 8 , 9 , 10 , 11 , 12 the full utility of ophthalmic features in explaining dementia risk has yet to be explored.

To fill this gap, we applied recently developed granular dry and wet AMD severity scoring systems (AMD‐D and AMD‐W) that were developed using item response theory (IRT). 13 These two scoring systems apply to individuals with and without diagnosed AMD, using features documented in routine electronic health record (EHR)‐based ophthalmic care, enabling continuous risk monitoring. Leveraging extensive EHR data from Adult Changes in Thought (ACT) study participants, we investigated associations between these IRT‐based measures of dry and wet AMD severity (individually and together) and incidence of each of AD and all‐cause dementia. We also compared the predictive accuracy of AD and all‐cause dementia risks using AMD severity scores to more conventional AMD diagnosis history. Additionally, by assessing associations between severity scores and AD/all‐cause dementia risks separately among people with and without diagnosed AMD, we evaluated whether information was gained about AD/all‐cause dementia risks using severity scores once AMD diagnostic history is known.

2. METHODS

2.1. Study cohort

2.1.1. ACT and Eye ACT studies

The ACT study is a prospective, longitudinal cohort study that recruits dementia‐free adults aged 65+ who are members of Kaiser Permanente Washington (KPWA). 14 Their cognitive function is assessed biennially until they are diagnosed with dementia, die, or disenroll. The Eye ACT study, an ACT ancillary study that supported the current analyses, obtained data on ophthalmic features in EHRs via natural language processing (NLP) from ACT participants seeking regular ophthalmic care in KPWA. 15

2.1.2. Analytic cohort

ACT participants were included if they had at least two ACT study visits between December 31, 2002 and October 1, 2022 (when KPWA EHR ophthalmic data were available), and in the 730 days prior to at least one of those visits that was not their last, they had at least one eye specialist visit and continuous KPWA enrollment. The analytic cohort was further restricted to individuals with complete covariate data (Figure 1, Sections 2.4 and 3).

FIGURE 1.

FIGURE 1

Open in a new tab

Flow chart of final analytic sample inclusion. ACT, adult changes in thought study; KPWA, Kaiser Permanente Washington. *Eye visits are routine visits with optometrists and ophthalmologists.

2.2. Exposures

Primary time‐varying exposures were the worst dry and wet AMD severities experienced to date across both eyes: AMD‐Dwe and AMD‐Wwe, respectively. These were derived from previously developed IRT‐based AMD severity scores 13 (AMD‐D and AMD‐W) based on AMD‐related features recorded in the EHR from any individual routine visit with optometrists or ophthalmologists – hereafter referred to as eye visits – and abstracted using NLP. The scores may be used to characterize both subclinical and clinical AMD in ranges of 0, indicating no evidence of that AMD subtype, to 10, indicating evidence of very severe AMD of that subtype. 13 See Supporting Information Methods 1 for more details on the previously developed scoring systems.

In primary analyses, we elected to use the worst ever score to reflect irreversible underlying AMD progression, despite potential symptomatic improvement with treatment. In sensitivity analyses, alternative definitions of exposures were considered (Section 3.3).

For comparison, we also used a binary indicator of prior AMD diagnosis (ever vs never), which is a crude but more conventionally studied biomarker of dementia in the literature. 9 To identify the history of AMD diagnosis, we searched for diagnostic codes throughout all available prior claims data (ICD9 codes: 362.57, 362.50, 362.51, 362.52; ICD10 codes: H35.36*, H35.30*, H35.31*, H35.32*).

2.3. Outcomes

We used ACT study criteria to identify incident AD (primary) and all‐cause dementia (secondary) outcomes. 14 For more details on the diagnostic process and defining the age of onset, see Supporting Information Methods 2.

2.4. Other covariates

Potential confounders included sex (man/woman), education reported at ACT baseline (bachelor's degree or more vs not), apolipoprotein E (APOE) genotype (𝜀4 carrier/non‐carrier), smoking status reported at ACT baseline (current/former/never smoker), birth cohort 16 (years 1913 and earlier, 1914 to 1920, 1921 to 1928, 1929 to 1939, and 1940+), and time‐varying indicators of histories of various health conditions at biennial ACT visits (hypertension, congestive heart failure, cerebrovascular disease [composite of stroke, transient ischemic attack, and carotid endarterectomy], and cardiovascular disease [composite of myocardial infarct, angina, coronary artery bypass graft, and angioplasty]). We also used an indication of race (non‐white/white), matching prior analyses 7 ; with over 90% of participants self‐reporting as White (Table 1), other race categories were collapsed for analytic purposes but individually described in Table 1. All covariates were calculated using self‐reports, except for APOE genotype.

TABLE 1.

Demographic and clinical characteristics summarized among all individuals in the final analytic cohort and by AMD diagnosis status (yes/no) at the end of analytic follow‐up.

AMD No AMD Overall
n = 826 (32%) n = 1785 (68%) N = 2611 (100%)
Age at analytic baseline in years: median (IQR)  80 (75,85) 74 (70,80) 76 (71,82)
Categorical covariates: n (%)
Women 527 (64%) 1009 (57%) 1536 (59%)
Race
White 770 (93%) 1589 (89%) 2359 (90%)
Black or African American  10 (1%) 54 (3%) 64 (3%)
American Indian or Alaska Native  1 (0.1%) 5 (0.3%) 6 (0.2%)
Asian  26 (3%) 61 (3%) 87 (3%)
Other/more than one race  19 (2%) 76 (4%) 95 (4%)
Birth cohort
Born 1913 or earlier (pre‐WWI) 32 (4%) 14 (0.8%) 46 (2%)
Born 1914–1920 (WWI and Spanish flu) 159 (19%) 113 (6%) 272 (10%)
Born 1921–1928 (interwar prosperity/roaring 20s) 318 (39%) 447 (25%) 765 (29%)
Born 1929–1939 (Great Depression) 244 (30%) 588 (33%) 832 (32%)
Born 1940+ (during and post‐WWII) 73 (9%) 623 (35%) 696 (27%)
Completion of bachelor's or more at ACT baseline  424 (51%) 1016 (57%) 1440 (55%)
At least one APOE ϵ4 allele  172 (21%) 522 (29%) 694 (27%)
Smoking status at ACT baseline 
Current smoker  30 (4%) 58 (3%) 88 (3%)
Never smoker  420 (51%) 910 (51%) 1330 (51%)
Past smoker  376 (46%) 817 (46%) 1193 (46%)
Self‐reported history of heart/circulatory conditions at end of analytic follow‐up 
Hypertension  503 (61%) 1001 (56%) 1504 (58%)
Congestive heart failure 84 (10%) 118 (7%) 202 (8%)
Cerebrovascular disease a 165 (20%) 262 (15%) 427 (16%)
Cardiovascular disease b 210 (25%) 332 (19%) 542 (21%)
Open in a new tab

Abbreviations: ACT, adult changes in thought; AMD, age‐related macular degeneration; APOE, apolipoprotein E; IQR, interquartile range.

a

Composite indicator of stroke, transient ischemic attack, and carotid endarterectomy.

b

Composite indicator of myocardial infarction, angina, coronary artery bypass grafting, and angioplasty.

2.5. Statistical analysis

2.5.1. Primary analyses

We fit three separate Cox regression models 17 , 18 , 19 for AD diagnosis, all adjusting for covariates listed earlier and using age as the time scale: (1) AMD‐Dwe as the sole exposure, (2) AMD‐Wwe as the sole exposure, and (3) AMD‐Dwe and AMD‐Wwe in the same model, thus allowing us to estimate the increase in AD hazard associated with the increase in one AMD severity score holding the other constant. Because AMD‐D and AMD‐W can only be updated at the time of an eye visit, we used the last‐value‐carried‐forward (LVCF) approach to impute the severity scores at time points between eye visits with the scores from the most recent preceding eye visit to determine values of AMD‐Dwe and AMD‐Wwe.

RESEARCH IN CONTEXT

  1. Systematic review: Literature identified using conventional publication search engines shows associations between AMD and AD risk; however, it typically fails to account for AMD subtype (dry and wet) and extent (or severity) of subclinical or clinical AMD.

  2. Interpretation: Using an AMD severity scoring system that uses EHRs, we estimated increased risk of AD with more severe dry and wet AMD. When stratified by the diagnosis of AMD, associations between worsening scores and AD risk remained strongest among those without a diagnosis of AMD.

  3. Future direction: Characterizing the severity of AMD using EHRs may help surveille and distinguish risks of AD in older adults, but the impact may be stronger among those without a diagnosis of AMD. Future studies will focus on determining the specific features driving this elevated risk, which may inform the shared mechanisms of AMD and AD.

We implemented delayed entry (left truncation), where each participant's first age at risk was the age at their earliest ACT study visit satisfying inclusion criteria (Section 2.1.2), hereafter referred to as analytic baseline. Participants’ final age at risk is the first of AD diagnosis (if any) and any of the following censoring events: non‐AD dementia diagnosis, complete disenrollment from KPWA (defined to be >60 days of contiguous KPWA disenrollment), death, and the last dementia‐free ACT follow‐up visit prior to October 1, 2022. At‐risk participants contributed to risk sets (defined by age) if they had complete covariate data at the age at which the risk set was defined.

We reported the point estimate and 95% confidence interval (CI) of the hazard ratio (HR) for AD for each one‐unit increase in an AMD severity score, accompanied by Wald tests to assess the strength of association between an AMD severity score and time to AD (using an omnibus Wald test for both score types when included in the same model).

To compare the predictive abilities of AD hazard using different descriptors of AMD, we also retained the (generalized) Akaike information criterion (AIC) from each model. 20 We used previously developed benchmarks 21 to determine meaningful differences in AIC between two models. We also assessed differences in the predictive accuracy of AD hazard using AMD severity scores relative to the indicator of prior AMD diagnosis (Section 2.2) by comparing generalized AIC from primary models with the generalized AIC from a model using prior AMD diagnosis (ever/never) as the exposure.

The entire primary analytic process was replicated for the secondary outcome of time to all‐cause dementia diagnosis.

2.5.2. Sensitivity analyses

We conducted two separate sensitivity analyses. First, in place of LVCF of AMD‐Dwe and AMD‐Wwe throughout follow‐up, we employed a Bayesian joint modeling framework, 22 , 23 which is known to reduce bias by simultaneously modeling the time to outcome events and longitudinal trajectories of multiple exposures (Sensitivity Analysis 1). For further details, see Supporting Information Methods 3. Second, we replaced AMD‐Dwe and AMD‐Wwe – measures of the worst ever AMD‐D and AMD‐W scores across both eyes – with measures of the worst current AMD‐D and AMD‐W scores across eyes at each time point (Sensitivity Analysis 2). We continued to use the LVCF approach for this exposure. See Figure S1 for illustration.

2.5.3. Secondary analyses

We assessed associations between AMD‐Dwe and AMD‐Wwe – in separate models and within the same model – and each of AD and all‐cause dementia risk within subgroups defined by history of AMD diagnosis (Section 2.2). Separate Cox models were fit within each subgroup defined by AMD diagnosis history: never AMD diagnosis and prior AMD diagnosis. Participants who started follow‐up without AMD but were diagnosed with incident AMD during follow‐up contributed to both analyses: (1) prior to incident AMD, participants contributed to the never AMD diagnosis subgroup analysis and (2) at the time of AMD diagnosis, such participants were right‐censored from the never AMD diagnosis subgroup analysis and began contributing to the prior AMD diagnosis subgroup analysis (see Figure S1 for illustration). As in primary analyses, we compared the predictive accuracy of AD and all‐cause dementia risks by different configurations of severity scores (AMD‐Dwe only, AMD‐Wwe only, and AMD‐Dwe together with AMD‐Dwe) using generalized AICs within each subgroup.

All statistical tests were two‐sided and conducted at a significance level of 0.05, and all primary and secondary model and Sensitivity Analysis 2 output included global tests for proportional hazards assumptions violations. For additional proportional hazards considerations, see Supporting Information Methods 4. R version 4.5.2 was used for all analyses. 24

3. RESULTS

A total of 3277 participants had at least one ACT visit meeting inclusion criteria. The analytic cohort was further reduced to 2611 due to a lack of APOE genotyping data (reduction by 18.2%) with small further reductions due to other missing covariates (0.5% missing race, 0.2% missing smoking status, and an additional 1.4% with incomplete health histories throughout all of follow‐up) (Figure 1). See Table S1 for comparisons of the analytic sample and target population.

Of the final analytic cohort of 2611 participants, by the end of follow‐up, a total of 826 people (31.6%) had been diagnosed with AMD, 550 (21.2%) developed AD, and 628 (24.1%) developed all‐cause dementia. Analyses included 19,008 person‐years (PYs) of follow‐up, with 24% of PYs among those with a history of AMD diagnosis at the time and a median 6.8 years at risk per person (interquartile range [IQR] 4.0 to 10.4, min 0.2, max 18.9).

The median age at the beginning of time at risk in analyses was 76.4 years (IQR 70.8, 82.0), skewing older among individuals with AMD diagnoses at the end of follow‐up (median 80.4 [IQR 75.1, 85.2] vs median 74.0 [IQR 69.7, 80.2]) (Table 1). Among the final analytic cohort, 59% were women, 90% self‐identified as White, and 55% completed a bachelor's degree or higher education. Compared to the participants who never received an AMD diagnosis, those with AMD diagnosis were more likely to be women (64% vs 57%), self‐identify as White (93% vs 89%), and born before 1929 (62% vs 32%); proportions reporting heart or circulatory problems were 3% to 6% higher (Table 1). Participants with AMD were less likely to have completed at least a bachelor's degree (51% vs 57%) and have at least one APOE ε4 allele (21% vs 29%) (Table 1).

The correlation between AMD‐Dwe and AMD‐Wwe was 0.50 across risk sets used in analyses. For further summaries of AMD‐Dwe and AMD‐Wwe during follow‐up, see Supporting Information Results 1.

3.1. Primary analyses

The hazard of AD was estimated to be 4% higher for each one unit increase (out of 10) in AMD‐Dwe (HR 1.04 [95% CI 1.01, 1.07]; p = 0.01) and AMD‐Wwe (HR 1.04 [95% CI 1.00, 1.08]; p = 0.03) when used in separate models (see Table 2: Everyone), with higher scores indicating worse severity. The use of both scores together was also significantly associated with the hazard of AD (p = 0.02), with their effects attenuated slightly from when used alone (AMD‐Dwe: HR 1.03 [95% CI 1.00, 1.07]; AMD‐Wwe: HR 1.02 [95% CI 0.98, 1.07]). Results were similar with all‐cause dementia as the outcome (Table 2: Everyone). See Tables S2A and S2B for adjustment covariate coefficient estimates.

TABLE 2.

Results of primary and secondary analyses of associations between worst ever dry and/or wet AMD severity scores – AMD‐Dwe and AMD‐Wwe, respectively – and hazards of each of AD (primary) and all‐cause dementia (secondary).

AD (primary) All‐cause dementia (secondary)
HR per unit increase in score c (95% CI) Generalized AIC d HR per unit increase in score c (95% CI) Generalized AIC d
AMD‐Dwe AMD‐Wwe p‐value a PH test e AMD‐Dwe AMD‐Wwe p value a PH test e
Dry score only
Everyone 1.04 (1.01,1.07) 0.01 5008 0.81 1.04 (1.01,1.07) 0.02 5690 0.88
With AMD 1.02 (0.97,1.07) 0.51 1744 0.38 1.02 (0.97,1.07) 0.47 1885 0.50
Without AMD 1.05 (1.00,1.10) 0.06 2642 0.85 1.04 (0.99,1.09) 0.09 3100 0.78
Wet score only
Everyone 1.04 (1.00,1.08) 0.03 5010 0.79 1.04 (1.00,1.08) 0.03 5691 0.88
With AMD 1.00 (0.95,1.06) 0.91 1744 0.44 1.00 (0.95,1.06) 0.94 1885 0.56
Without AMD 1.05 (0.99,1.10) 0.10 2643 0.84 1.05 (1.00,1.10) 0.04 3099 0.80
Both scores
Everyone 1.03 (1.00,1.07) 1.02 (0.98,1.07) 0.02 5009 0.82 1.02 (0.99,1.06) 1.03 (0.99,1.07) 0.03 5691 0.91
With AMD 1.02 (0.97,1.07) 1.00 (0.94,1.06) 0.80 1746 0.44 1.02 (0.97,1.07) 1.00 (0.94,1.05) 0.76 1887 0.58
Without AMD 1.04 (0.98,1.09) 1.03 (0.97,1.09) 0.11 2643 0.83 1.02 (0.97,1.08) 1.04 (0.98,1.10) 0.09 3100 0.79
Open in a new tab

Abbreviations: AD, Alzheimer's disease; AIC, Akaike Information Criterion; AMD, age‐related macular degeneration; AMD‐Dwe, worst ever observed dry AMD severity score; AMD‐Wwe, worst ever observed wet AMD severity score; CI, confidence interval; HR, hazard ratio; PH, proportional hazards.

a

Results of (omnibus) Wald test of all AMD severity scores included in model.

b

All models have baseline hazards stratified by indicators of presence of at least one apolipoprotein E allele and time‐varying self‐reported histories of cerebrovascular disease (composite indicator of stroke, transient ischemic attack, and carotid endarterectomy) and congestive heart failure and adjust for indicators of female gender, completion of a Bachelors or more at ACT baseline, self‐reported white race, smoking status (separate indicators for never smoking and being a past smoker at ACT baseline, with current smoking as the referent group), and time‐varying self‐reported histories of hypertension and cardiovascular disease (composite indicator of myocardial infarction, angina, coronary artery bypass grafting, and angioplasty).

c

All severity scores re‐scaled to be between 0 (least severe) and 10 (most severe).

d

Lower AIC = better; differences in AIC between two models provide the following support for discrepancy in predictive accuracy of the outcome: <2 = (effectively) no support, 2–4 = low support, 4–7 = moderate support, 7–10 = moderate/high support, >10 = high support; generalized AIC using history of binary AMD diagnosis as the exposure (among everyone) was 5005 for AD and 5689 for all‐cause dementia.

e

Global test of proportional hazards assumptions violations in final models using scaled Schoenfeld residuals.

Overall, the predictive accuracy of AD hazard using AMD‐Dwe only (generalized AIC 5008) was nearly identical to the use of AMD‐Dwe and AMD‐Wwe together (generalized AIC 5009) and very similar to using AMD‐Wwe alone (generalized AIC 5010) (Table 2). Despite severity scores containing more granular information about AMD progression than an indicator of AMD diagnosis, there is low support for binary indicators of prior AMD diagnosis being more informative than severity scores with respect to AD hazard (generalized AIC 5005, with HR 1.34 [95% CI 1.11, 1.62]). The informativeness of different measures of AMD with respect to all‐cause dementia hazard was even more similar, with differences in generalized AIC of no more than 2 (Table 2).

3.2. Sensitivity analyses

Estimates of association in both sensitivity analyses were similar to or stronger than primary analyses. See Supporting Information Results 2 and Table S3 for further details.

3.3. Secondary analyses

Among participants with no history of AMD diagnosis, the estimated associations with AD hazard of AMD‐Dwe (HR 1.05 [95% CI 1.00, 1.10]; p = 0.06) and AMD‐Wwe (HR 1.05 [95% CI 0.99, 1.10]; p = 0.10) remained positive and with magnitudes similar to those observed in the entire population, albeit failing to reach statistical significance. The same relative findings were found for associations when both scores were used together in the same model (Table 2: Without AMD). Among people with no history of AMD, all individual and combinations of scores provide nearly identical predictive accuracy of AD hazard. Results were similar in this subpopulation with the dementia outcome (Table 2: Without AMD).

Among individuals previously diagnosed with AMD, relative to the entire population, estimates of association between AMD‐Dwe and AD hazard were attenuated and estimates of association between AMD‐Wwe and AD hazard were near null when scores were used both alone and together (Table 2: With AMD). Individual scores showed low support for superior informativeness in predicting AD hazard compared to using both scores together. Associations and relative informativeness of scores were similar for the all‐cause dementia outcome (Table 2: With AMD).

See Tables S2A and S2B for adjustment covariate coefficient estimates.

4. DISCUSSION

Based on the analyses of 19,008 PY of follow‐up data among 2611 ACT study participants, worse maximal severity of dry and wet AMD experienced to date – quantified using severity scores AMD‐Dwe and AMD‐Wwe, respectively – were each associated with higher risk of developing AD and all‐cause dementia with similar magnitudes of association across the AMD subtypes. Although scores were individually and jointly associated with AD/all‐cause dementia risk significantly, their predictive accuracy of AD/all‐cause dementia hazard were similar to indicators of prior AMD diagnosis. Associations between each severity score and AD/all‐cause dementia risk were attenuated or near null among individuals with a prior AMD diagnosis and were similar in magnitude to those observed in the full cohort among individuals without diagnosed AMD. These results suggest that the more granular scores do little to further distinguish AD/all‐cause dementia risk among people who have been diagnosed with AMD, though the scores may provide some explanatory power in people without a documented AMD diagnosis.

Differences in AD/all‐cause dementia hazard were presented for one‐unit severity score contrasts. However, the development of individual or small numbers of AMD‐related eye features often leads to jumps in severity scores far greater than one. 13 For example, individuals experience a 5‐point increase in dry AMD score with the development of pigmentation change only and in wet AMD score with the development of subretinal hemorrhage only; using results from models with individual severity scores, this jump yields an estimated 22% (=1.045) increase in AD hazard.

Several studies in well‐characterized prospective cohorts have found positive associations between AMD and risks of AD or all‐cause dementia, such as the Rotterdam Health Study, 25 the Cardiovascular Health study, 26 and the UK Biobank. 27 Results from administrative claims‐based studies have been mixed, 28 , 29 , 30 , 31 underscoring important limitations in relying solely on claims data, such as variability in diagnostic coding and selection bias. 32 , 33 Most studies to date have characterized AMD using diagnostic history (yes/no), generally failing to distinguish between dry and wet AMD subtypes, with few exceptions, 34 , 35 thereby disallowing characterization of risk by AMD subtype (dry and wet) and possible changes in risk as AMD progresses (both subclinically and clinically). Both of these limitations are addressed by our use of continuous AMD subtype severity scores as explanatory measures.

Our findings of positive associations among participants with no AMD diagnosis and attenuated associations among those with an AMD diagnosis were somewhat unexpected but may reflect pathology. We hypothesize that a clinical diagnosis of AMD may reflect a threshold of underlying shared pathology, beyond which increases in AMD severity provided limited increases in information about dementia risk.

To help interpret the results among participants without AMD, it is important to note that we do expect variation in severity scores among people without AMD – including individuals with subclinical AMD – thereby fulfilling one requirement for a useful biomarker of AD risk. As previously mentioned, dry and wet severity scores are determined based on the presence of AMD‐related eye features, which are potentially present among people without an AMD diagnosis. With a median time between eye specialist visits of 10 weeks (interquartile range, 3 to 29 weeks), highly accurate identification of AMD by eye specialists, and extensive testing and updating of the NLP algorithm used to abstract eye features from the EHR, 13 the risks of misclassification of AMD diagnosis or related features in our database are largerly mitigated.

Relative to prior studies examining associations between AMD and AD/all‐cause dementia, our study offers several important strengths. We combined high‐quality research data with real‐world clinical data in a well‐characterized cohort. Many existing studies rely on brief cognitive tests (e.g., the MMSE, Hasegawa, or MoCA scales) or ICD codes to approximate or identify dementia diagnoses. In contrast, the ACT cohort is monitored biennially to identify incident cases of AD/all‐cause dementia by consensus with research‐quality protocols. Furthermore, the prospective nature of the ACT study enables updates to time‐varying covariates (such as cardiovascular conditions) along with AMD status over follow‐up, allowing us to quantify more accurate attribution of risk of AD/all‐cause dementia to AMD status among otherwise similar individuals. Finally, the consistency of primary results across sensitivity analyses increases confidence in our primary findings.

Findings should be interpreted in the context of our study's limitations. Notably, the cohort is primarily composed of non‐Hispanic White participants, potentially limiting generalizability. Healthcare utilization requirements for inclusion to ensure sufficient data for analyses may introduce selection bias. The utilization of EHRs provides a unique opportunity to examine the history and progression of ophthalmic findings in ACT participants. However, these observations are highly dependent on consistency in the recording of retinal features in the EHRs, particularly given the sizeable impact individual features have on AMD severity scores; inconsistent recording practices could lead to spurious variability in calculated severity scores, lessening the precision of estimates of association. Incomplete recording of retinal features could also lead to underestimation of AMD severity, spuriously increasing point estimates of association. Our look back in the EHRs for information about eye features that were missing data at individual eye visits mitigates the impact of features missed at subsequent visits that were previously documented. However, this does not directly address the subjective nature of documentation practices or the possibility that certain visits may not have warranted retinal examinations that would have captured AMD‐related features. Although our application of a NLP algorithm to extract eye features is a unique strength of our study, it may also introduce potential sources of error. Misclassification of documented features could lead to inaccuracy in severity scores, yielding similar problems to inconsistent and incomplete recording of eye features by clinicians in the EHR.

Our future research will examine the utility of both simplified and expanded characterizations of AMD severity in describing AD/all‐cause dementia risk. We have already begun evaluating whether individual retinal features observed among participants both with and without AMD diagnosis (most commonly pigmentation change and retinal scarring) are stronger predictors of AD/all‐cause dementia than overall AMD severity scores. These analyses may help identify which aspects of retinal aging or pathology are most relevant to neurodegenerative risk. As additional imaging modalities and biomarker measures are incorporated into the growing Eye ACT dataset and broader EHR, we aim to revise our AMD severity scoring systems to incorporate these more recently available biomarkers and assess associations between these updated scores and risks of AD and all‐cause dementia. 36 , 37 , 38 , 39 , 40 , 41

In conclusion, our study contributes to the understanding of the association between AMD and AD/all‐cause dementia by demonstrating the information gained about AD risk from early retinal changes that occur prior to a formal diagnosis of AMD. Although the underlying mechanisms remain unclear, our findings underscore the importance of characterizing specific retinal pathological features that may contribute to elevated AD/all‐cause dementia risk. Further investigation of these retinal changes may help elucidate shared or convergent pathways underlying neurodegeneration. Our study highlights the promise of a granular retinal severity scoring system using available EHR data for population‐level surveillance, mostly among individuals without diagnosed AMD, and the role that routine eye exams and longitudinal follow‐up can play in monitoring ophthalmic risk factors for AD and all‐cause dementia in at‐risk populations.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the supporting information.

CONSENT STATEMENT

All participants included in this study provided informed consent.

Supporting information

Supporting Information

DAD2-18-e70368-s001.pdf (987.7KB, pdf)

ICMJE Disclosure Form

DAD2-18-e70368-s002.pdf (682.4KB, pdf)

ACKNOWLEDGMENTS

We thank the participants of the Adult Changes in Thought (ACT) study for the data they provided and the many ACT investigators and staff who steward that data. You can learn more about ACT at: actagingresearch.org. This work was funded by National Institute on Aging R01AG060942, U19AG066567, and National Institute of Health OT2OD32644, Gates Venture, and Alzheimer's Drug Discovery Foundation (ADDF) Diagnostic Accelerator. Data collection for this work was additionally supported, in part, by prior funding from the National Institute on Aging U01AG006781. The funders had no role in the study design; in the collection, analysis, and interpretation of data; in the writing of the report; or in the decision to submit the paper for publication.

DATA AVAILABILITY STATEMENT

Data from this analysis cannot be made publicly available for ethical and legal reasons. In order to replicate our findings, a researcher may need access to personal health identifiers (PHI) Version 8/6/2024, including dates of birth and death, dates of diagnoses, and ages over 89. These are required variables for the analysis, and we cannot publicly release this information without IRB approval and a Data Use Agreement with interested researchers. However, the datasets used and/or analyzed in the current study are available upon reasonable request and execution of appropriate human subjects review and data sharing agreements by following the process described on the Adult Changes in Thought website: actagingresearch.org.

REFERENCES

  • 1. Nichols E, Steinmetz JD, Vollset SE, et al. Estimation of the global prevalence of dementia in 2019 and forecasted prevalence in 2050: an analysis for the Global Burden of Disease Study. Lancet Public Health. 2022;7(2):e105‐e125. doi: 10.1016/S2468-2667(21)00249-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2. Alzheimer's Association . 2025 Alzheimer's disease facts and figures. Alzheimers Dement. 2025;21(4):e70235. [DOI] [PubMed] [Google Scholar]
  • 3. London A, Benhar I, Schwartz M. The retina as a window to the brain‐from eye research to CNS disorders. Nat Rev Neurol. 2013;9(1):44‐53. doi: 10.1038/nrneurol.2012.227 [DOI] [PubMed] [Google Scholar]
  • 4. Butovsky O, Rosenzweig N. Alzheimer's disease and age‐related macular degeneration: shared and distinct immune mechanisms. Immunity. 2025;58(5):1120‐1139. doi: 10.1016/j.immuni.2025.04.013 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5. Kim SY, Cheon J. Senescence‐associated microvascular endothelial dysfunction: a focus on the blood‐brain and blood‐retinal barriers. Ageing Res Rev. 2024;100:102446. doi: 10.1016/j.arr.2024.102446 [DOI] [PubMed] [Google Scholar]
  • 6. Livingston G, Huntley J, Liu KY, et al. Dementia prevention, intervention, and care: 2024 report of the Lancet standing commission. Lancet. 2024;404(10452):572‐628. doi: 10.1016/S0140-6736(24)01296-0 [DOI] [PubMed] [Google Scholar]
  • 7. Lee CS, Larson EB, Gibbons LE, et al. Associations between recent and established ophthalmic conditions and risk of Alzheimer's disease. Alzheimers Dement J Alzheimers Assoc. 2019;15(1):34‐41. doi: 10.1016/j.jalz.2018.06.2856 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8. Chung SD, Ho JD, Chen CH, Lin HC, Tsai MC, Sheu JJ. Dementia is associated with open‐angle glaucoma: a population‐based study. Eye. 2015;29(10):1340‐1346. doi: 10.1038/eye.2015.120 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Tsai HR, Lo RY, Liang KH, et al. Risk of subsequent dementia or Alzheimer disease among patients with age‐related macular degeneration: a systematic review and meta‐analysis. Am J Ophthalmol. 2023;247:161‐169. doi: 10.1016/j.ajo.2022.11.005 [DOI] [PubMed] [Google Scholar]
  • 10. Exalto LG, Biessels GJ, Karter AJ, Huang ES, Quesenberry CP, Whitmer RA. Severe diabetic retinal disease and dementia risk in type 2 diabetes. J Alzheimers Dis. 2014;42 Suppl 3(0 3):S109‐117. doi: 10.3233/JAD-132570 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Whitmer R, Exalto L, Biessels GJ, Quesenberry C Jr.. O2‐10‐06: microvascular damage in the kidney and eye and dementia risk: an epidemiological perspective. Alzheimers Dement. 2012;8(4S_Part_7):P256‐P256. doi: 10.1016/j.jalz.2012.05.683 [DOI] [Google Scholar]
  • 12. Chung SD, Lee CZ, Kao LT, Lin HC, Tsai MC, Sheu JJ. Association between neovascular age‐related macular degeneration and dementia: a population‐based case‐control study in Taiwan. PLOS ONE. 2015;10(3):e0120003. doi: 10.1371/journal.pone.0120003 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Lee CS, Su YR, Walker RL, et al. A data‐driven age‐related macular degeneration severity scoring system leveraging the AREDS studies and clinical electronic medical records. Ophthalmology. 2025;132(10):1076‐1087. doi: 10.1016/j.ophtha.2025.04.026 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Kukull WA, Higdon R, Bowen JD, et al. Dementia and Alzheimer disease incidence: a prospective cohort study. Arch Neurol. 2002;59(11):1737‐1746. doi: 10.1001/archneur.59.11.1737 [DOI] [PubMed] [Google Scholar]
  • 15. Lee CS, Ferguson AN, Gibbons LE, et al. Eye adult changes in thought (Eye ACT) study: design and report on the inaugural cohort. J Alzheimers Dis. 2024;100(1):309‐320. doi: 10.3233/JAD-240203 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Tom SE, Phadke M, Hubbard RA, Crane PK, Stern Y, Larson EB. Association of demographic and early‐life socioeconomic factors by birth cohort with dementia incidence among US adults born between 1893 and 1949. JAMA Netw Open. 2020;3(7):e2011094. doi: 10.1001/jamanetworkopen.2020.11094 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17. Therneau TM, Grambsch PM. The Cox model. In: Therneau TM, Grambsch PM, eds. Modeling Survival Data: Extending the Cox Model. Springer; 2000:39‐77. doi: 10.1007/978-1-4757-3294-8_3 [DOI] [Google Scholar]
  • 18. Andersen PK, Gill RD. Cox's regression model for counting processes: a large sample study. Ann Stat. 1982;10(4):1100‐1120. doi: 10.1214/aos/1176345976 [DOI] [Google Scholar]
  • 19. Therneau T. survival: A Package for Survival Analysis in R. Version 3.8‐3. 2024. https://CRAN.R‐project.org/package=survival
  • 20. Venables W, Ripley B. Modern Applied Statistics with S‐Plus. 2nd ed. Springer; 1997. [Google Scholar]
  • 21. Burnham KP, Anderson DR. Multimodel Inference: understanding AIC and BIC in model selection. Sociol Methods Res. 2004;33(2):261‐304. doi: 10.1177/0049124104268644 [DOI] [Google Scholar]
  • 22. Rizopoulos D. Joint Models for Longitudinal and Time‐to‐Event Data: With Applications in R. Chapman and Hall/CRC; 2012. doi: 10.1201/b12208 [DOI] [Google Scholar]
  • 23. Rizopoulos D, Miranda Afonso P, Papageorgiou G. JMbayes2: extended joint models for longitudinal and time‐to‐event data. Version 0.6‐0. 2026. doi: 10.32614/CRAN.package.JMbayes2 [DOI]
  • 24. R:A Language and Environment for Statistical Computing. Version 4.5.2. R Foundation for Statistical Computing; 2025. Accessed April 2026. https://www.r‐project.org/ [Google Scholar]
  • 25. Klaver CC, Ott A, Hofman A, Assink JJ, Breteler MM, de Jong PT. Is age‐related maculopathy associated with Alzheimer's disease? The Rotterdam Study. Am J Epidemiol. 1999;150(9):963‐968. doi: 10.1093/oxfordjournals.aje.a010105 [DOI] [PubMed] [Google Scholar]
  • 26. Hwang PH, Longstreth WT, Thielke SM, et al. Ophthalmic conditions associated with dementia risk: the Cardiovascular Health Study. Alzheimers Dement J Alzheimers Assoc. 2021;17(9):1442‐1451. doi: 10.1002/alz.12313 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Shang X, Zhu Z, Huang Y, et al. Associations of ophthalmic and systemic conditions with incident dementia in the UK Biobank. Br J Ophthalmol. 2023;107(2):275‐282. doi: 10.1136/bjophthalmol-2021-319508 [DOI] [PubMed] [Google Scholar]
  • 28. Wen LY, Wan L, Lai JN, et al. Increased risk of Alzheimer's disease among patients with age‐related macular degeneration: a nationwide population‐based study. PLOS ONE. 2021;16(5):e0250440. doi: 10.1371/journal.pone.0250440 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Choi S, Jahng WJ, Park SM, Jee D. Association of age‐related macular degeneration on Alzheimer or Parkinson disease: a retrospective cohort study. Am J Ophthalmol. 2020;210:41‐47. doi: 10.1016/j.ajo.2019.11.001 [DOI] [PubMed] [Google Scholar]
  • 30. Lishinsky‐Fischer N, Cnaany Y, Nitzan I, Chowers I, Levy J. Longitudinal assessment of age‐related macular degeneration and risk of dementia using clinical setting data. Ophthalmol Sci. 2025;5(6):100891. doi: 10.1016/j.xops.2025.100891 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Keenan TDL, Goldacre R, Goldacre MJ. Associations between age‐related macular degeneration, Alzheimer disease, and dementia: record linkage study of hospital admissions. JAMA Ophthalmol. 2014;132(1):63‐68. doi: 10.1001/jamaophthalmol.2013.5696 [DOI] [PubMed] [Google Scholar]
  • 32. Hackl CM, Lee VA, Dunnigan JK, et al. A narrative review of TriNetX in ophthalmology: applications, advantages, and pitfalls. Ann Eye Sci. 2024;9. doi: 10.21037/aes-24-27 [DOI] [Google Scholar]
  • 33. Olaker VR, Fry S, Terebuh P, et al. With big data comes big responsibility: strategies for utilizing aggregated, standardized, de‐identified electronic health record data for research. Clin Transl Sci. 2025;18(1):e70093. doi: 10.1111/cts.70093 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34. Tsai HR, Lo RY, Liang KH, et al. Risk of subsequent dementia or Alzheimer disease among patients with age‐related macular degeneration: a systematic review and meta‐analysis. Am J Ophthalmol. 2023;247:161‐169. doi: 10.1016/j.ajo.2022.11.005 [DOI] [PubMed] [Google Scholar]
  • 35. Tsai DC, Chen SJ, Huang CC, Yuan MK, Leu HB. Age‐related macular degeneration and risk of degenerative dementia among the elderly in Taiwan: a population‐based cohort study. Ophthalmology. 2015;122(11):2327‐2335.e2. doi: 10.1016/j.ophtha.2015.07.033 [DOI] [PubMed] [Google Scholar]
  • 36. Sadda SR, Guymer R, Holz FG, et al. Consensus definition for atrophy associated with age‐related macular degeneration on OCT: classification of atrophy report 3. Ophthalmology. 2018;125(4):537‐548. doi: 10.1016/j.ophtha.2017.09.028 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Guymer RH, Rosenfeld PJ, Curcio CA, et al. Incomplete retinal pigment epithelial and outer retinal atrophy in age‐related macular degeneration: classification of atrophy meeting report 4. Ophthalmology. 2020;127(3):394‐409. doi: 10.1016/j.ophtha.2019.09.035 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38. Wu Z, Luu CD, Hodgson LAB, et al. Prospective longitudinal evaluation of nascent geographic atrophy in age‐related macular degeneration. Ophthalmol Retina. 2020;4(6):568‐575. doi: 10.1016/j.oret.2019.12.011 [DOI] [PubMed] [Google Scholar]
  • 39. Wu Z, Pfau M, Blodi BA, et al. OCT signs of early atrophy in age‐related macular degeneration: interreader agreement: classification of atrophy meetings report 6. Ophthalmol Retina. 2022;6(1):4‐14. doi: 10.1016/j.oret.2021.03.008 [DOI] [PubMed] [Google Scholar]
  • 40. Agrón E, Domalpally A, Cukras CA, et al. Reticular pseudodrusen: the third macular risk feature for progression to late age‐related macular degeneration: age‐related eye disease study 2 report 30. Ophthalmology. 2022;129(10):1107‐1119. doi: 10.1016/j.ophtha.2022.05.021 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41. Waldstein SM, Vogl WD, Bogunovic H, Sadeghipour A, Riedl S, Schmidt‐Erfurth U. Characterization of drusen and hyperreflective foci as biomarkers for disease progression in age‐related macular degeneration using artificial intelligence in optical coherence tomography. JAMA Ophthalmol. 2020;138(7):740‐747. doi: 10.1001/jamaophthalmol.2020.1376 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

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

Supplementary Materials

Supporting Information

DAD2-18-e70368-s001.pdf (987.7KB, pdf)

ICMJE Disclosure Form

DAD2-18-e70368-s002.pdf (682.4KB, pdf)

Data Availability Statement

Data from this analysis cannot be made publicly available for ethical and legal reasons. In order to replicate our findings, a researcher may need access to personal health identifiers (PHI) Version 8/6/2024, including dates of birth and death, dates of diagnoses, and ages over 89. These are required variables for the analysis, and we cannot publicly release this information without IRB approval and a Data Use Agreement with interested researchers. However, the datasets used and/or analyzed in the current study are available upon reasonable request and execution of appropriate human subjects review and data sharing agreements by following the process described on the Adult Changes in Thought website: actagingresearch.org.

Articles from Alzheimer's & Dementia : Diagnosis, Assessment & Disease Monitoring are provided here courtesy of Wiley

RESOURCES