Retinal BioAge is associated with indicators of cardiovascular-kidney-metabolic syndrome in UK and US populations

Abstract

Accelerated biological aging, as well as cardiovascular, kidney, and metabolic (CKM) diseases, contribute to shortened healthspan. We studied a deep-learning model, retinal BioAge, and multiple indicators of CKM syndrome in participants from UK Biobank and the US-based EyePACS dataset. Retinal BioAge was trained on 77,887 retinal images and then used to analyze separate retinal images from UK Biobank (10,976) and EyePACS (19,856). In both datasets, CKM biomarker profiles were significantly worse for the top vs. bottom quartiles of BioAgeGap (retinal BioAge—chronological age), including measures of blood pressure, kidney function, adiposity, and glycemia. The top BioAgeGap quartile also had a significantly higher prevalence of clinical CKM indicators, including hypertension, kidney disease, and diabetes (UK Biobank) or suboptimally controlled diabetes (EyePACS). Thus, analysis of retinal images for accelerated biological aging may provide opportunistic screening to help identify individuals who could benefit from formal CKM assessment, potentially contributing to earlier detection and management of CKM syndrome.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-41465-8.

Introduction

Cardiovascular, kidney, and metabolic diseases are major contributors to global morbidity and mortality, and healthcare costs^1–4. There is also a growing recognition of the pathophysiologic overlap of these conditions and their interplay in risk for adverse clinical events. These combined conditions have been described recently as the cardiovascular-kidney-metabolic (CKM) syndrome by the American Heart Association, with a staging system based on the presence and severity of indicators for all three diseases^5,6. Thus, detecting indicators of CKM syndrome presents an opportunity to enhance interventions for prevention and risk-factor management to attenuate progression to cardiovascular events, kidney failure, and diabetes complications.

The retina is unique in being the only part of the human vasculature that is directly visible. It is widely recognized that ocular biomarkers can indicate the risk for, or presence of, important systemic diseases, including the well-known retinal signs of hypertensive and diabetic retinopathy^7–10. With progress in artificial intelligence (AI), and deep learning (DL) in particular, there are now many studies using AI to analyze retinal images^11–13. We and others have shown that these DL models can detect cardiovascular risk, chronic kidney disease, and diabetic retinopathy^14–24.

DL models analyzing retinal images have also been developed to estimate biological age. Since its initial inception in the late 1960s, there is increasing recognition that biological aging impacts the risk of an individual developing a major disease and its severity^25–28. Traditional regression-based risk equations typically do not account for individuals aging at different rates as they assume all individuals in the population age at the same rate. Thus, a retinal biological age may provide a broader and more personalized indicator of health for an individual. We and others have shown previously that biological age derived from AI analysis of retinal images correlates with telomere length and higher cardiovascular and cancer mortality^29–32. Here we apply our retinal BioAge DL model to indicators of CKM syndrome in two different cohorts – a UK-based general population and a diverse, US-based cohort of persons living with diabetes.

Methods

Retinal BioAge model development

The retinal BioAge DL model (Toku, Inc.) was developed as an extension of prior work determining atherosclerotic cardiovascular disease (ASCVD) risk from retinal images and limited demographic data¹⁵. The retinal BioAge was derived by comparing the AI-predicted 10-year ASCVD risk score of an individual to those of their peers of higher and lower chronological age, as published previously²⁹. The Pooled Cohort Equations (PCE) were used for the 10-year ASCVD risk, as this was the guideline-endorsed tool available at the time of model development³³. The equation for calculating retinal BioAge is provided in the Supplementary Methods. Retinal BioAge was calculated as the expected chronological age of individuals with similar model-predicted ASCVD risk, by comparing each participant’s AI-predicted risk to that of their closest peers in the retinal feature space. Additional robustness steps were used to limit the influence of outliers in both the feature space and ASCVD risk predictions. Finally, each participant’s retinal BioAgeGap was calculated as retinal BioAge minus chronological age.

Training and validation datasets

The composition of the UK Biobank and EyePACS 10 K datasets used in this study is shown in Table 1. Participants in the UK Biobank were recruited, with informed consent, from a UK general population with 90% non-Hispanic White participants and only 5% self-identified as having diabetes “diagnosed by doctor.” The data from the UK Biobank were accessed via a direct request to the UK Biobank (per University of Auckland IRB protocol UOA-86299) in accordance with relevant guidelines and regulations. 88,318 retinal images from 51,955 unique participants from the UK Biobank with acceptable quality images were available. Participants with any missing data required to calculate the 10-year ASCVD risk score were excluded. Validation testing was limited to ages 41–70, as there were few participants outside this age range. This left 86,247 images from 48,825 participants: 75,271 images from 43,349 participants were used for training purposes and 10,976 images from 5476 participants were preserved for validation testing (Fig. 1).

Table 1.

Demographic and risk factor makeup of the training and test datasets.

		Training (N = 44,731): UK Biobank (N = 43,349) & EyePACS 10 K (N = 1382)		Test: UK Biobank (N = 5476)		Test: EyePACS 10 K (N = 9786)
		Mean	Std	Mean	Std	Mean	Std
Age (years)		56.6	8.8	56.1 **	8.0	57.0**	7.2
Systolic blood pressure (mmHg)		137.1	18.5	135.3**	17.2	131.2**	12.8
Diastolic blood pressure (mmHg)		81.4	10.2	81.2	9.6	71.3**	8.9
Hemoglobin A1c (%)		5.5%	0.7%	5.4%**	0.5%	8.1%**	1.8%
Total cholesterol (mg/dL)		220.0	44.5	219.9	42.8	177.5**	44.0
Non-high-density lipoprotein cholesterol (mg/dL)		162.5	41.6	162.2	40.2	130.9**	42.8
Body mass index (kg/m2)		27.3	4.9	27.0**	4.6	33.0**	9.7
Urine albumin-to-creatinine ratio (mg/g)		32.9	175.0	29.8	131.2	145.4**	600.3
Estimated glomerular filtration rate (mL/min/1.73 m2)		95.0	14.6	96.1**	14.1	97.9**	21.3
Sex assigned at birth		Male	Female	Male	Female	Male	Female
		20,134 (45%)	24,597 (55%)	2470 (45%)	3006 (55%)	4115 (42%)†	5671 (58%)†
Current smoker		TRUE	FALSE	TRUE	FALSE	TRUE	FALSE
		3617 (8%)	40,886 (92%)	459 (8%)	5017 (92%)	575 (6%)†	9211 (94%)†
Diabetes (%)		3445 (8%)	41,069 (92%)	247 (5%)†	5213 (95%)†	9782 (> 99%)†	4 (< 1%)†
Race and ethnicity	Black	2%		2%		7%
	East Asian	< 1%		< 1%		7%
	Hispanic	2%		N.A.		67%
	Native American	< 1%		N.A.		< 1%
	Non- Hispanic White	90%		94%		6%
	South Asian	2%		2%		< 1%
	Multi-racial	< 1%		< 1%		< 1%
	Other	1%		1%		2%
	Declined/unknown	1%		1%		12%

Significance testing was performed between training and both test datasets (**P < 0.001 z test. †P < 0.01 Chi square). The UK Biobank did not include Hispanic ethnicity as an option and the White participants were predominantly of British and Irish origin. EyePACS 10 K included choices of Hispanic, Black, or White, so separating Hispanic Black participants from Hispanic White participants was not possible.

Fig. 1.

Flow chart illustrating the composition of the training and test datasets. ASCVD, atherosclerotic cardiovascular disease.

EyePACS 10 K is a diverse, US-based dataset of persons living with diabetes undergoing diabetic retinopathy screening (Table 1). The EyePACS dataset was accessed and analyzed per University of California, Berkeley IRB protocol UCB 2017-09-10340 in accordance with relevant guidelines and regulations, receiving IRB exemption from informed consent as it was a retrospective de-identified dataset. A subset of 19,856 retinal images from 9786 individuals in EyePACS was set aside for validation testing, based on having at least one retinal image of good quality from each eye and within the same age range of 41 to 70 years old as tested in UK Biobank. As the retinal BioAge model requires comparing individuals to peers of higher and lower chronological age, training data for those aged 30–40 and 71–80 were required. UK Biobank is limited to individuals over 40 years old with only a small representation of those over 70. Therefore, 1382 participants from EyePACS who were 30–40 or 71–80 years old were included only to support model training and calibration at the extremes of the age distribution. Specifically, 853 participants with 1535 images aged 30–40 (46.3% female, mean age: 36.1 years) and 529 participants with 1081 images aged 71–80 (65% female, mean age: 73.8 years) with adequate image quality were selected to provide coverage of those ages. These individuals were not included in the main CKM association analyses, i.e., for both UK Biobank and EyePACS, the holdout datasets used to validate the retinal BioAge model were restricted to individuals between 41 and 70 years of age. UK Biobank data were derived from baseline assessment center visits, where retinal imaging and biomarker measurements were obtained in a standardized clinical setting. EyePACS data came from a diabetic retinopathy screening program in community-based settings. We emphasize that our analyses are based on cross-sectional data (imaging and biomarker) rather than longitudinal follow-up. Data cleaning and quality control included reliance on cohort-provided biomarker quality control procedures, exclusion of retinal images with poor quality scores, and appropriate handling of missing data through exclusion of participants without complete data for the relevant analyses.

Retinal BioAge model assessment

Assessment of BioAgeGap associations with CKM biomarkers

For both UK Biobank and EyePACS, the following biomarkers were analyzed: systolic blood pressure (SBP), diastolic blood pressure (DBP), estimated glomerular filtration rate (eGFR), body mass index (BMI), hemoglobin A1c (HbA1c), and non-HDL cholesterol (non-HDL-C). Waist circumference and arterial stiffness index were also available in UK Biobank but not EyePACS. To account for the association between chronological age and BioAgeGap, participants in both datasets were grouped into 5-year age bins (i.e., 41–45, 46–50, etc.) to calculate the population mean and the difference (Δ) between the mean and the actual value of each biomarker for every participant. We then compared the Δ of each biomarker for the top (Q4) and bottom (Q1) quartiles of BioAgeGap in both datasets.

Assessment of BioAgeGap associations with clinical CKM indicators

For both UK Biobank and EyePACS, the prevalence of the following conditions was analyzed: hypertension, elevated cholesterol, impaired kidney function, obesity, and diabetes. Hypertension was defined as SBP ≥ 140 mmHg or DBP ≥ 90mmHg for UK Biobank (based on UK guidelines³⁴ and SBP ≥ 130 mmHg or DBP≥80mmHg for EyePACS (based on US guidelines³⁵. Elevated cholesterol was defined as non-HDL ≥ 130 mg/dL³⁶. Impaired kidney function was defined as estimated GFR < 60 mL/min per 1.73 m² or urine albumin-to-creatinine ratio (UACR) ≥ 30 mg/g (per KDIGO guideline³⁷ of moderate or greater risk). Diabetes was defined as HbA1c ≥ 6.5%³⁸. As the EyePACS participants were known to have diabetes, the threshold for well-controlled diabetes (HbA1c < 7.0%) was used³⁹. Obesity was defined as BMI of ≥ 30 kg/m² or waist circumference (available in UK Biobank only) of ≥ 88 cm for women or ≥ 102 cm for men⁴⁰. For EyePACS, the prevalence of diabetic retinopathy (DR) was also analyzed, defined as the presence of any DR (i.e., grade > R0)⁴¹. The prevalence of the above clinical indictors was compared between participants in the top and bottom quartiles of BioAgeGap across the 5-year age bins for both UK Biobank and EyePACS.

Data analysis

Statistical comparison of mean biomarker values between the top and bottom retinal BioAgeGap quartiles was performed using two-sided unpaired t-tests. Statistical comparison of the prevalence of CKM indicators between the top and bottom retinal BioAgeGap quartiles was performed by constructing a contingency table followed by chi-squared tests. A Bonferroni correction was applied to adjust for multiple comparisons. A family-wise significance level of α = 0.05 was used. The 95% confidence intervals were generated using 10,000 bootstrap resamples to assess variability and compute interval limits. An additional subgroup analysis by sex was conducted to examine the relationships separately for males and females. As levels of blood pressure and non-HDL-C can be impacted by anti-hypertensive and cholesterol-lowering medications, respectively, CKM indicator prevalence was analyzed by medication status. Medication variables were derived from prescription or medication fields where available. In UK Biobank, self-reported medication use for antihypertensive, lipid-lowering, and diabetes medications was available from the participant interview data fields. In EyePACS, insulin use was recorded as part of the diabetic retinopathy screening data collection. We note that medication data were incomplete in some settings and were not the primary focus of this analysis. Statistical analyses were conducted using Python, employing the SciPy and statsmodels libraries.

Results

Study population characteristics

The characteristics of the study populations are summarized in Table 1. The UK Biobank validation cohort included 5476 participants (53.3% female) with a mean age of 56.7 years; 90% were of non-Hispanic White race and ethnicity, and 5% reported physician-diagnosed diabetes. The EyePACS validation cohort included 9786 participants (54.4% female) with a mean age of 54.0 years; this was a diabetic retinopathy screening population with 100% having diabetes, and 77.6% were of Hispanic/Latino ethnicity. The association between chronological age and BioAgeGap showed R² = 0.14 for UK Biobank and R² = 0.27 for EyePACS, supporting our age-stratified analytical approach.

Biomarkers

In the UK Biobank, individuals in the top vs. bottom quartiles of retinal BioAgeGap had significantly higher biomarker levels (Table 2), including SBP (difference = + 8.01 mmHg), DBP (+ 4.63 mmHg), arterial stiffness (+ 0.39), HbA1c (+ 0.06%), BMI (+ 0.72 kg/m²), and waist circumference (+ 2.45 cm). Individuals in the top quartile of retinal BioAgeGap had a significantly lower non-HDL-C compared to those in the bottom quartile (− 4.11 mg/dL). They were also more likely to be on cholesterol-lowering medication (20% vs. 14%) (Supplementary Table 1). The difference in eGFR (− 0.14 mL/min/1.73 m²) was not statistically significant.

Table 2.

UK Biobank: Difference between mean CKM biomarker values and population mean for top and bottom retinal BioAgeGap quartiles.

CKM biomarker	Top retinal BioAgeGap quartile (95% CI)	Bottom retinal BioAgeGap quartile (95% CI)	p-value
SBP (mmHg)	3.69 (2.86 to 4.53)	− 4.32 (− 5.11 to − 3.52)	< 0.001*
DBP (mmHg)	2.17 (1.64 to 2.70)	− 2.46 (− 2.95 to − 1.99)	< 0.001*
Arterial stiffness index	0.16 (− 0.00 to 0.32)	− 0.23 (− 0.38 to − 0.08)	< 0.001*
HbA1c (%)	0.03 (0.00 to 0.06)	− 0.03 (− 0.06 to − 0.00)	0.003*
BMI (kg/m2)	0.34 (0.10 to 0.59)	− 0.38 (− 0.63 to − 0.14)	< 0.001*
Waist circumference (cm)	0.86 (0.20 to 1.54)	− 1.59 (− 2.28 to − 0.92)	< 0.001*
eGFR (mL/min/1.73 m2)	− 0.16 (− 0.84 to 0.51)	− 0.02 (− 0.75 to 0.58)	0.769
Non-HDL cholesterol (mg/dL)	− 2.40 (− 4.43 to − 0.19)	1.71 (− 0.35 to 3.86)	0.007*

SBP, systolic blood pressure; DBP, diastolic blood pressure; HbA1c, - Hemoglobin A1c; BMI, Body Mass Index; eGFR, estimated glomerular filtration rate; HDL, high-density lipoprotein. *Denotes statistical significance after Bonferroni correction.

For the biomarkers available in the EyePACS dataset (Table 3), the top vs. bottom quartiles of retinal BioAgeGap also had significantly higher SBP (+ 3.53 mmHg) and HbA1C (+ 0.73%), plus a significantly lower eGFR (− 3.47 mL/min/1.73 m²), compared to those in the bottom quartile. The differences in the DBP (− 0.07 mmHg), BMI (− 0.27 kg/m²), and non-HDL-C (− 0.66 mg/dL) were not statistically significant.

Table 3.

EyePACS 10 K: Difference between mean CKM biomarker values and population mean for top and bottom retinal BioAgeGap quartiles.

CKM biomarker	Top retinal BioAgeGap quartile (95% CI)	Bottom retinal BioAgeGap quartile (95% CI)	p-value
SBP (mmHg)	2.10 (1.52 to 2.66)	− 1.43 (− 1.92 to − 0.94)	< 0.001*
DBP (mmHg)	0.36 (− 0.00 to 0.72)	0.43 (0.09 to 0.80)	0.771
HbA1c (%)	0.50 (0.43 to 0.58)	− 0.23 (− 0.29 to − 0.17)	< 0.001*
BMI (kg/m2)	− 0.40 (− 0.86 to 0.13)	− 0.13 (− 0.46 to 0.19)	0.394
eGFR (mL/min/1.73 m2)	− 2.82 (− 3.80 to − 1.85)	0.45 (− 0.29 to 1.20)	< 0.001*
Non-HDL cholesterol (mg/dL)	− 1.07 (− 2.84 to 0.76)	− 0.41 (− 2.05 to 1.33)	0.618

SBP, systolic blood pressure; DBP, diastolic blood pressure; HbA1c, Hemoglobin A1c; BMI, Body Mass Index; eGFR, estimated glomerular filtration rate; HDL, high-density lipoprotein. *Denotes statistical significance after Bonferroni correction.

Graphing the observed mean biomarker values (with 95% confidence intervals) vs. chronological age for the top (Q4) and bottom (Q1) quartiles in UK Biobank (Fig. 2, Supplementary Table 2) showed profiles consistent with the findings above. These curves represent descriptive summaries of observed data within age groups, not fitted model-based predictions. SBP, DBP, arterial stiffness, HbA1c, BMI, and waist circumference increased with age, more so for the top BioAgeGap quartile. eGFR decreased with age while non-HDL-C increased then decreased, neither showing significant differences between Q4 and Q1 quartiles. Interestingly, in EyePACS, all biomarkers except SBP decreased with age (Fig. 3, Supplementary Table 3), with notable differences in Q4 vs. Q1 for SBP, HbA1c, and eGFR.

Fig. 2.

Observed CKM biomarker profiles stratified by retinal BioAgeGap quartiles in UK Biobank. Graphs showing observed mean values and 95% confidence intervals for key CKM biomarkers across chronological age groups, comparing the top (Q4) and bottom (Q1) retinal BioAgeGap quartiles. Higher BioAgeGap is associated with worse biomarker profiles across most age ranges. SBP, systolic blood pressure; DBP, diastolic blood pressure; HbA1c, Hemoglobin A1c; eGFR, estimated glomerular filtration rate; HDL, high-density lipoprotein.

Fig. 3.

Observed CKM biomarker profiles stratified by retinal BioAgeGap quartiles in EyePACS. Graphs showing observed mean values and 95% confidence intervals for key CKM biomarkers across chronological age groups, comparing the top (Q4) and bottom (Q1) retinal BioAgeGap quartiles. Higher BioAgeGap is associated with worse biomarker profiles across most age ranges. SBP, systolic blood pressure; DBP, diastolic blood pressure; HbA1c, Hemoglobin A1c; eGFR, estimated glomerular filtration rate; HDL, high-density lipoprotein.

Clinical indicators of CKM syndrome

In UK Biobank (Table 4), participants in the top quartile of retinal BioAgeGap had a significantly greater prevalence of hypertension compared to those in the bottom quartile (51.3% vs. 30.0%, p < 0.001). They were also more likely to have kidney disease (5.9% vs. 3.9%, p = 0.02), diabetes (2.9% vs. 1.5%, p = 0.017) and obesity by both BMI and waist circumference. Elevated non-HDL-C was highly prevalent in both quartiles (> 75%) without a significant difference.

Table 4.

UK Biobank: Prevalence of clinical indicators of CKM syndrome for top and bottom retinal BioAgeGap quartiles.

CKM indicator	Top retinal BioAgeGap quartile	Bottom retinal BioAgeGap quartile	p-value
Hypertension: SBP ≥ 140 mmHg or DBP ≥ 90 mmHg	701/1366 = 51.3%	411/1372 = 30.0%	< 0.001*
Elevated cholesterol: Non-HDL cholesterol ≥ 130 mg/dL	1053/1366 = 77.1%	1079/1372 = 78.6%	0.349
CKD: eGFR < 60 mL/min/1.73 m2 or UACR ≥ 30 mg/g	81/1366 = 5.9%	54/1372 = 3.9%	0.02*
Diabetes: HbA1c ≥ 6.5%	39/1366 = 2.9%	20/1372 = 1.5%	0.017*
Obesity: BMI ≥ 30 kg/m2	329/1366 = 24.1%	267/1372 = 19.5%	0.004*
Obesity: Waist Circumference F ≥ 88 cm/M ≥ 102 cm	467/1366 = 34.2%	382/1372 = 27.8%	< 0.001*

SBP, systolic blood pressure; DBP, diastolic blood pressure; HDL, high-density lipoprotein; CKD, chronic kidney disease; eGFR, estimated glomerular filtration rate; UACR, urine albumin-creatinine ratio; HbA1c, Hemoglobin A1c; BMI, Body Mass Index. *Denotes statistical significance after Bonferroni correction.

In EyePACS (Table 5), participants in the top BioAgeGap quartile had a higher prevalence of hypertension (54.3% vs. 46.7%, p < 0.001), kidney disease (35.8% vs. 20.8%, p < 0.001), suboptimally controlled diabetes (69.0% vs. 57.8%, p < 0.001), and DR (57.8% vs. 9.4%, p < 0.001). Those in Q4 were also twice as likely to be on insulin (30% vs. 15%) (Supplementary Table 1). There was no significant difference in the prevalence of elevated cholesterol. Notably, and in contrast to UK Biobank findings, individuals in the bottom quartile of BioAgeGap in EyePACS had a higher prevalence of obesity compared to those in the top quartile (52.6% vs. 48.9%, p = 0.01). This discordant pattern is discussed further in the Discussion section.

Table 5.

EyePACS 10 K: Prevalence of clinical indicators of CKM syndrome for top and bottom retinal BioAgeGap quartiles.

CKM indicator	Top retinal BioAgeGap quartile	Bottom retinal BioAgeGap quartile	p-value
Hypertension: SBP ≥ 130 mmHg or DBP ≥ 80 mmHg	1327/2444 = 54.3%	1143/2449 = 46.7%	< 0.001*
Elevated cholesterol: Non-HDL Cholesterol ≥ 130 mg/dL	952/2444 = 38.9%	978/2449 = 39.9%	0.501
CKD: eGFR < 60 mL/min/1.73 m2 or UACR ≥ 30 mg/g	876/2444 = 35.8%	510/2449 = 20.8%	< 0.001*
Suboptimally controlled diabetes HbA1c ≥ 7.0%	1686/2444 = 69.0%	1415/2449 = 57.8%	< 0.001*
Obesity: BMI ≥ 30 kg/m2	1195/2444 = 48.9%	1289/2449 = 52.6%	0.01*
Diabetic retinopathy: DR grade > 0	1413/2444 = 57.8%	230/2449 = 9.4%	< 0.001*

SBP, systolic blood pressure; DBP, diastolic blood pressure; HDL, high-density lipoprotein; CKD, chronic kidney disease; eGFR, estimated glomerular filtration rate; UACR, urine albumin-creatinine ratio; HbA1c, Hemoglobin A1c; BMI, Body Mass Index; DR, diabetic retinopathy. *Denotes statistical significance after Bonferroni correction.

Subgroup analyses

Separate analyses for female and male participants again showed very similar findings to those reported above for both the UK Biobank and EyePACS with a few differences (Supplementary Tables 4 and 5). In the UK Biobank, the increased prevalence of kidney disease was only significant for males. The increased prevalence of diabetes and obesity in the top quartile did not reach statistical significance when males and females were examined separately. In EyePACS, differences in hypertension, kidney disease, suboptimally controlled diabetes, and DR remained significant for both males and females. The difference in obesity was only significant for males.

The higher prevalence of hypertension in both datasets between the top vs. bottom BioAgeGap quartiles remained significant after analyzing by hypertension medication status (Supplementary Tables 6 and 7). There remained no difference in elevated cholesterol prevalence in both datasets after analyzing by cholesterol medication status. Interestingly, UK Biobank individuals in Q4 were more likely to be on hypertensive and cholesterol-lowering medications than those in Q1 (Supplementary Table 1). In EyePACS, the proportion on hypertension and cholesterol medications was similar, but individuals in Q4 were twice as likely to be on insulin than those in Q1.

Discussion

Our AI analysis of retinal BioAge in UK and US cohorts found that individuals in the top vs. bottom quartiles of retinal BioAgeGap had a significantly worse biomarker profile and a higher prevalence of multiple indicators of CKM syndrome. Specifically, those in UK Biobank had a higher SBP, DBP, arterial wall stiffness, BMI, waist circumference, and HbA1c and were more likely to have hypertension, kidney disease, obesity, and diabetes. Those in EyePACS had a higher SBP, HbA1c, a lower eGFR, and were more likely to have hypertension, kidney disease, suboptimally controlled diabetes, and DR. Thus, retinal BioAgeGap appears promising as a marker of greater burden of important cardiovascular, kidney, and metabolic disease indicators.

Prior work has associated biological aging with a wide range of cardiovascular and other diseases, with a recent series of review articles published in the Journal of the American College of Cardiology²⁶. Notably, AI models that estimate biological age from retinal images have shown they can also predict all-cause, cardiovascular, and cancer mortality and the risk of ASCVD events^30,31,42. We have shown previously that our retinal BioAge model, when tested only in the UK Biobank dataset, correlated with shortened leukocyte telomere length and elevated ASCVD biomarkers²⁹. In the current study, we extend the analysis to a more diverse, higher-risk, US-based population and evaluated retinal BioAgeGap with both the levels of key CKM biomarkers and the prevalence of actionable indicators of adverse CKM health.

The two cohorts examined in this study differ substantially in their characteristics and clinical context. UK Biobank represents a predominantly middle-aged to older general population cohort, primarily of non-Hispanic White participants, with relatively low prevalence of diabetes and chronic kidney disease. In contrast, EyePACS is a diabetic retinopathy screening population in the United States, where all participants have diabetes and many have advanced disease. These differences in demographics, underlying disease burden, clinical context (research assessment center vs. community diabetic screening), and biomarker measurement protocols likely contribute to observed differences in biomarker patterns across the two cohorts, including the unusual inverse relationship between BioAgeGap and obesity in EyePACS.

That the biomarker profiles of those individuals with a higher retinal BioAgeGap was worse than those in the lower group indicates that retinal BioAge is capable of stratifying individuals into higher- and lower-risk cohorts. Importantly, the worse biomarker profiles translated to higher prevalence of clinical indicators of the CKM syndrome, and this was observed in both datasets. This is consistent with other studies that have reported DL models trained on the retina can stratify an individual’s risk for cardiovascular disease, kidney failure, metabolic disease, and diabetic retinopathy^{29,30,42–45}.

With hypertension the top modifiable risk factor for global cardiovascular mortality, and often unrecognized or uncontrolled, the association of retinal BioAge with higher SBP, arterial stiffness, and hypertension prevalence provides a valuable opportunity for detecting at-risk patients^1,2,35. Kidney disease is a key component of the CKM syndrome and is often underrecognized and undertreated, contributing to adverse cardiometabolic outcomes and potentially kidney failure^3,46. In persons living with diabetes, approximately 40% have chronic kidney disease per the KDIGO risk classification, which is based on having either decreased eGFR or increased UACR^37,47. Prior retinal AI studies have focused primarily on predicting chronic kidney disease based on eGFR only^20–22. Here we show that retinal BioAge correlated with higher prevalence of chronic kidney disease using the KDIGO criteria (i.e., eGFR and UACR) in both UK Biobank and EyePACS.

Excess or dysfunctional adipose tissue is a key component of the CKM syndrome and contributes to hyperglycemia, hypertension, and dyslipidemia^6,48. In UK Biobank, we found an elevated retinal BioAgeGap correlated with a higher BMI, waist circumference, and obesity prevalence. Only BMI data were available in the EyePACS dataset and a similar relationship between adiposity and BioAgeGap was not observed in this diabetic population. Importantly, retinal BioAge was associated with a higher HbA1c in both UK Biobank and EyePACS datasets, plus a higher prevalence of diabetes in the former and suboptimally controlled diabetes in the latter. Also in EyePACS, the top quartile of retinal BioAgeGap had a 6-fold higher prevalence of any DR and a significantly higher prevalence of more-than-mild DR (45.8% vs. 3.3%, p < 0.001). Modjtahedi et al. have demonstrated that not only is DR associated with future risk of ASCVD events and death, but the magnitude of this risk correlates with the severity of retinopathy⁴⁹. These authors hypothesize that DR could be a surrogate marker of wider vascular disease and conclude that future calculators that are designed to stratify ASCVD risk in diabetic individuals could incorporate detailed retinopathy input information to model this risk more accurately. It is important to acknowledge that in the EyePACS cohort, BioAgeGap likely reflects a composite of DR-related changes, other microvascular features, and more global retinal aging. The very strong association between BioAgeGap and DR prevalence (57.8% vs. 9.4% in top vs. bottom quartiles) raises the question of whether BioAgeGap is simply capturing DR severity rather than a distinct biological aging signal. From a biological standpoint, DR and other microvascular changes are integral to the pathway by which retinal aging manifests in diabetic populations, making it difficult to cleanly separate these constructs. We did not formally test whether BioAgeGap contributes independently of DR grade in predicting other CKM indicators; future work in cohorts with detailed DR grading and a broader mix of diabetic and non-diabetic participants will be needed to disentangle these relationships.

The finding that higher BioAgeGap was associated with lower BMI and obesity prevalence in EyePACS, opposite to the pattern observed in UK Biobank, warrants specific discussion. Several plausible explanations exist for this discordant result. First, EyePACS is a diabetic retinopathy screening cohort rather than a general population; individuals with long-standing, poorly controlled diabetes may experience unintentional weight loss (due to disease progression, medication effects, or cachexia) yet still carry substantial CKM burden and therefore a high BioAgeGap⁵⁰. Second, there may be an ‘obesity paradox’ effect, wherein lower body weight among individuals with established cardiometabolic disease can paradoxically be associated with more advanced pathology and higher mortality risk⁵¹. Third, differences in age distribution, treatment patterns, and unmeasured comorbidities between the two cohorts may contribute. We emphasize that this finding is hypothesis-generating, that BMI is only one component of CKM risk, and that our primary conclusions are based on the broader pattern of associations across multiple biomarkers and disease indicators in both cohorts.

The primary strengths of this study include: (1) demonstrating that accelerated retinal BioAge from AI analysis of retinal images is associated with a significantly worse CKM biomarker profile and a higher prevalence of important clinical indicators of CKM syndrome, (2) demonstrating these findings in both the general UK Biobank population and the diverse, diabetic US EyePACS cohort.

Hypertension, increased adiposity, chronic kidney disease, and diabetes are all clinical indicators of CKM syndrome and call for both lifestyle modification and guideline-directed medical therapy. CKM staging is advocated to personalize risk assessment and guide management strategies⁵. It is important to note that the current study examines cross-sectional associations, and BioAgeGap should be considered a potential marker of underlying CKM syndrome burden rather than a validated predictor of incident outcomes. Future longitudinal work could quantify the incremental predictive value of BioAgeGap beyond standard clinical risk factors using metrics such as net reclassification improvement and integrated discrimination improvement. With over 100 million eye exams performed annually in the US alone, AI-based point-of-care retinal image analysis may provide an important opportunity to help detect indicators for adverse cardiovascular, kidney, and metabolic health.

A limitation that is in common with all DL models is the generalizability of the current retinal BioAge model as it was trained primarily on the UK Biobank dataset with limited diversity and low rates of chronic kidney disease and diabetes. Encouragingly, the model performed well on the diverse and higher-disease-burden, US-based EyePACS cohort. Further validation on additional external datasets is warranted. The retinal BioAge DL model also required training on cases with chronological ages above and below the age range for analysis, so a limited number of EyePACS cases with ages ≤ 40 and > 70 years were used for training only (> 95% of the training dataset was from UK Biobank), with the subsequent validation testing restricted to ages 41–70. Notably, the information about medication use was quite different between the two datasets, with UK Biobank relying on self-report and with insulin listed as the only diabetes medication. Furthermore, the EyePACS dataset lacks several variables that were available in UK Biobank, including complete lipid profiles in all participants, waist circumference measurements, and arterial stiffness indices. This limited our ability to fully characterize CKM biomarker associations in the EyePACS cohort and may contribute to differences in observed patterns between the two datasets.

Our analytical framework of comparing BioAgeGap quartiles within age strata provides a transparent description of associations but does not fully adjust for potential confounders. Residual confounding by age (within strata), sex, race/ethnicity, medication use, socioeconomic factors, and other unmeasured variables may remain. The consistent direction and magnitude of associations across multiple CKM indicators and in both cohorts supports the robustness of the signal, but more detailed multivariable modeling would be valuable in future studies designed specifically for regression-based risk modeling and prediction. Additionally, the cross-sectional design of this study precludes inferences about the temporal relationship between retinal BioAgeGap and CKM outcomes. We cannot determine whether elevated BioAgeGap precedes or merely accompanies CKM disease indicators. Future work leveraging longitudinal follow-up data in UK Biobank and similar cohorts will be necessary to test whether baseline BioAgeGap predicts incident CKM events such as myocardial infarction, stroke, heart failure, and kidney disease progression, using appropriate survival models and competing risk approaches.

Our analyses cannot definitively separate general biological aging signals from features that may act as proxies for race, ethnicity, or other unmeasured demographic variables. The retinal BioAge model was trained primarily on a UK Biobank cohort that is predominantly non-Hispanic White participants, yet performed consistently when applied to the predominantly Hispanic EyePACS population. While this consistency across diverse populations is reassuring, it does not exclude the possibility that the model learns demographic shortcuts or domain-specific features rather than universally applicable biological signals. We did not perform a formal bias audit or unsupervised domain adaptation analyses in this study. Such fairness-oriented analyses would be highly informative but require substantial additional modeling work that lies beyond the scope of this descriptive association study; we frame them as an important priority for future research.

Another important direction for future research is the application of explainable AI methods, such as SHAP values or Grad-CAM saliency maps, to characterize which retinal regions are most influential in determining BioAgeGap predictions. Prior work on retinal biological age has suggested that vascular and optic disc features are particularly informative, but a detailed analysis specific to our model would help determine whether it captures established retinal biomarkers of aging and CKM risk or potentially identifies novel subclinical retinal phenotypes. Such interpretability analyses constitute a substantial methodological undertaking and are therefore planned as a dedicated follow-up study.

In conclusion, these results suggest that retinal BioAgeGap derived from deep learning analysis of retinal images is associated with key indicators of CKM syndrome in cross-sectional analyses. Rather than serving as a standalone risk predictor, retinal BioAge may provide opportunistic, point-of-care screening to help identify individuals who warrant formal CKM assessment, including associated tools such as the new PREVENT risk calculator⁵². As retinal photographs can be captured rapidly and noninvasively, these algorithms could be widely deployed in eye care and primary care settings without significant additional investment. Moreover, the ease of use makes these technologies particularly relevant to lower-resource settings and thus facilitates the broader detection of at-risk populations. If these results can be replicated, retinal DL models may offer the potential to significantly increase detection and awareness of actionable CKM syndrome indicators and ultimately help improve patient outcomes.

Tables.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

D.S., E.V., and M.V.M. contributed to the design of the work, the analysis and interpretation of data, drafting of the manuscript, and reviewing it for important intellectual content. C.N., S.A., S.M., S.Y., L.X., and A.R. contributed to the acquisition, analysis, and interpretation of data, and reviewing the manuscript for important intellectual content. M.K.D., H.H., and R.N.W. contributed to interpretation of data and reviewing the manuscript for important intellectual content. All authors read and approved the final manuscript.

Data availability

UK Biobank and EyePACS datasets are available for research per the policies of those organizations. UK Biobank—https://www.ukbiobank.ac.uk/about-our-data/. EyePACS—https://www.eyepacs.com/data-analysis. Study-specific data are available from the corresponding author upon reasonable request.

Declarations

Competing interests

DS, CN, SA, SM, SY, LX, AR, EV and MVM report employment by Toku, Inc. MVM reports compensation by Porter Health for consultant services. MKD and HH report employment by Topcon Healthcare. RNW reports compensation by Toku, Inc. for consultant and board of directors services and by Topcon Healthcare for consultant services, as well as research instruments from Topcon, Visionix, Centervue, and Konan. We sought to mitigate the risk of bias by fully documenting our methods so they can be scrutinized and replicated by independent groups. The CKM-specific analysis plan was prespecified before the CKM associations with retinal BioAgeGap were examined. All data extraction and quality control procedures adhered to standard operating procedures established in our prior peer-reviewed retinal BioAge studies.