The distribution of artificial intelligence–derived retinal cardiovascular risk scores and conventional risk factors in two Korean health screening cohorts: a descriptive study
Article information
Abstract
Background
Although retinal imaging–based artificial intelligence (AI) tools have recently been introduced for cardiovascular disease (CVD) risk assessment, little is known about the distribution of these AI-derived scores across the full age spectrum or their associations with traditional cardiometabolic risk factors at different ages.
Methods
We analyzed data from 138,745 participants who underwent routine health examinations at two health screening centers in Seoul, Korea. The AI-based retinal CVD risk score (Dr.Noon CVD), as well as anthropometric, hemodynamic, and metabolic indices and cardiometabolic disease status, were compared across ages 16 to 96 years. In a subgroup of 13,182 individuals who underwent coronary artery calcium scoring (CACS) by cardiac computed tomography, we evaluated the performance of the Dr.Noon CVD score in detecting CACS using receiver operating characteristic curve analysis.
Results
Mean Dr.Noon CVD scores rose steadily with age from 14.2±2.9 (<30 years) to 46.3±6.5 (≥70 years), closely mirroring the increase in traditional cardiovascular risk factors with age. Additional analysis using CACS demonstrated that the Dr.Noon CVD score achieved an area under the curve of 0.80 (95% confidence interval, 0.80–0.81) for detecting any coronary calcification, defined as CACS >0, and an area under the curve of 0.82 (95% confidence interval, 0.81–0.83) for identifying significant calcification burden, defined as CACS >100.
Conclusions
Dr.Noon CVD scores were consistently correlated with age, conventional risk factors, and CACS, suggesting a potential role in broad-based cardiovascular risk stratification and in guiding personalized prevention strategies.
INTRODUCTION
Cardiovascular disease (CVD) remains the leading cause of death worldwide and is a major contributor to disability, highlighting the urgent need for effective prevention strategies [1]. To reduce the impact of CVD, substantial efforts have focused on identifying at-risk individuals before clinical events occur, leading to widespread adoption of CVD risk calculators such as the Framingham Risk Score [2], the American Pooled Cohort Equations (PCE), and the European SCORE2 [3,4]. In Korea, CVD risk is also estimated using locally calibrated tools, such as the Korean Risk Prediction Model derived from domestic cohorts [5], and the “Health iN” calculator, developed by Korea’s National Health Insurance Service, for 10-year and lifetime CVD risk [6]. Although these models offer valuable estimates based on clinical measurements and lifestyle factors, they may underestimate CVD risk in individuals with subclinical disease or atypical risk profiles [7]. Additionally, traditional calculators are typically validated only in middle-aged and older adults (e.g., ≥40 years), limiting their applicability to younger populations. Therefore, alternative approaches to CVD risk assessment are being developed, including novel imaging–based and artificial intelligence (AI)-driven methods that aim to deliver a more direct and individualized evaluation of cardiovascular risk [8]. Among these, retinal imaging has emerged as a particularly promising modality, offering a noninvasive window into systemic vascular health across all age groups [9].
The retina is the only location in the human body where microvascular structures can be directly visualized noninvasively, providing a unique opportunity to assess systemic vascular health [10]. Retinal abnormalities—such as arteriolar narrowing, venular widening, arteriovenous nicking, retinal microaneurysms, hemorrhages, and increased vessel tortuosity—have long been associated with cardiovascular outcomes [11–13]. International guidelines endorse retinal examination for detecting target organ damage in hypertension and diabetes mellitus (DM) [14,15]. Building on this foundation, recent advances in AI have enabled automated extraction of cardiovascular risk markers directly from retinal images [15]. AI-based models have demonstrated the ability to predict clinical variables—including age, blood pressure, and lipid levels—with high accuracy [16], and have shown risk stratification performance comparable to established tools like PCE in large-scale cohorts [8]. Among these AI-driven tools, Dr.Noon CVD (Mediwhale Inc) has emerged as a leading model, trained on over 200,000 retinal images linked to coronary artery calcium scores (CACS) to generate personalized 5-year CVD risk estimates from retinal images [8,17].
The present study aims to provide a comprehensive descriptive overview of a large, age-diverse Korean health screening cohort, focusing on the distribution of Dr.Noon CVD scores and their relationship to traditional cardiometabolic risk factors across the lifespan. By presenting detailed age- and risk-stratified statistics, we seek to contextualize the AI-derived retinal biomarker within the broader landscape of conventional CVD risk factors and disease burden. In addition, we explore a secondary objective: evaluating the performance of Dr.Noon CVD scores in predicting coronary artery calcium in this population, highlighting the added value and clinical potential of integrating AI-based retinal risk assessment into routine health screening.
METHODS
Ethics statement
This study was granted exemption from full review by the Central Institutional Review Board via the e-IRB system (No. P01-202409-01-004). The study was conducted in accordance with the Declaration of Helsinki.
Study design and participants
This retrospective study included 138,745 subjects (aged 16–96 years) who visited two health screening centers in Seoul, Korea (GC I-MED Gangnam and GC I-MED Gangbuk), between July 2022 and September 2024. At each visit, up to three duplicate retinal images were obtained per eye, yielding a total of 304,950 retinal images. Among these participants, 3,756 underwent unilateral retinal imaging, while 135,212 participants underwent bilateral (or more repeated) imaging. Participants were excluded if they met any of the following criteria: (1) absence of retinal images; or (2) a prescription or diagnosis history of angina, myocardial infarction, or stroke according to the National Health Screening Survey.
Data collection and variables
All demographic, clinical, laboratory, and imaging data were obtained from the electronic medical record system as part of routine health screening examinations. Age, sex, height, and weight were recorded, and body mass index (BMI) was calculated as weight in kilograms divided by height in meters squared. Waist circumference, systolic and diastolic blood pressures, fasting glucose, glycosylated hemoglobin (HbA1c), total cholesterol, low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), and triglycerides were also retrieved. Diagnoses of hypertension, DM, and dyslipidemia were identified based on self-reported history in patient questionnaires or when laboratory or measurement results met predefined thresholds: hypertension as systolic blood pressure (SBP) ≥140 mmHg or diastolic blood pressure (DBP) ≥90 mmHg; DM as fasting glucose ≥126 mg/dL or HbA1c ≥6.5%; and dyslipidemia as total cholesterol ≥240 mg/dL, LDL-C ≥160 mg/dL, HDL-C <40 mg/dL, or triglycerides ≥200 mg/dL. Obesity was defined as a BMI of ≥25 kg/m2, and abdominal obesity as a waist circumference ≥90 cm in men or ≥85 cm in women.
Retinal image and CACS acquisition
Retinal images were captured using a fundus camera (CR-2 AF, Canon) and saved in DICOM format for subsequent analysis. CACS was performed on noncontrast chest computed tomography (CT) scans using a Philips Ingenuity 64-slice multidetector CT scanner (Philips Healthcare) and quantified by the Agatston method, employing the Coronary Calcium Scoring application on the Philips IntelliSpace Portal (Philips Healthcare).
Dr.Noon CVD score
Dr.Noon CVD (Mediwhale Inc) is an AI software-as-a-medical-device designed to estimate an individual’s 5-year risk of incident major adverse cardiovascular events (MACE; myocardial infarction, stroke, heart failure admission, or cardiovascular death). The system received approval from the Korean Ministry of Food and Drug Safety on August 1, 2022, after designation as an Innovative Medical Device on December 24, 2020. The algorithm is a convolutional neural network built on a 50-layer residual backbone pretrained on ImageNet. Training utilized 208,398 macula-centered, 45° color fundus photographs pair-linked to noncontrast coronary CT calcium scores (CACS) from health screening cohorts in Korea (71%), Singapore (23%), and the United Kingdom (6%). Images were randomly split (stratified by site) into sets of development (80%), tuning (10%), and locked internal-test (10%) (Fig. S1) [8,17]. The network outputs a probability—from 0 (no calcium) to 1 (calcium present; Agatston >0)—that is subsequently rescaled to the 0 to 100 Dr.Noon CVD score used clinically. For each participant, the model analyzes two nonmydriatic fundus photographs; the image-level probabilities are averaged, rescaled to a 0 to 100 index, and then mapped to three prespecified risk strata—low (<30), moderate (30–40), and high (>40)—corresponding to approximately 1%, 1% to 5%, and >5% predicted 5-year MACE risk, respectively. For each subject, two nonmydriatic retinal images are processed through successive convolutional layers to extract vascular and structural features, which are then aggregated into a single score ranging from 0 to 100 with three risk groups (low, <30; moderate, 30–40; and high, >40). These categories reflect 5-year CVD risk levels: a low-risk group with a predicted risk of 1%, a moderate-risk group with a risk between 1% to 5%, and a high-risk group with a risk greater than 5%. In the CMERC-HI cohort, direct comparison demonstrated Harrell C-indices for 5-year MACE of 0.751 (95% confidence interval [CI], 0.683–0.820) for the Dr.Noon CVD model, 0.741 (95% CI, 0.673–0.807) for CACS, 0.707 (95% CI, 0.623–0.791) for carotid intima–media thickness, and 0.710 (95% CI, 0.625–0.794) for brachial-ankle pulse wave velocity [8,17]. The system provides clinicians with an objective, image-based prediction of future cardiovascular events and supports earlier intervention and personalized risk management. In this study, for participants with two or more images, the mean Dr.Noon CVD score across all available images was used.
Subgroup analysis
To confirm the relationship between Dr.Noon CVD scores and CACS in this dataset, a separate subgroup analysis was performed for those who underwent CACS. CACS was summarized descriptively by sex and within each age group, then categorized into three tiers—absent (CACS, 0), mild (CACS, 1–99), and moderate to high (CACS, ≥100)—as well as binarized at two thresholds: any calcification (CACS >0) and significant burden (CACS >100).
Statistical analysis
The population was divided into 10 groups (<30, 30–34, 35–39, 40–44, 45–49, 50–54, 55–59, 60–64, 65–69, and >70 years). Continuous variables were compared using the Kruskal-Wallis test, while categorical variables were analyzed with the chi-square test. Statistical significance was defined as a P-value of <0.05. For variables showing statistical significance, age group differences were considered meaningful. For continuous measures, age-stratified boxplots of the cohort were generated. In the subgroup analysis, Pearson correlation with Dr.Noon CVD scores was calculated on both raw CACS and log-transformed CACS values [log10(CACS+1)] to account for the skewed distribution of calcium scores. Receiver operating characteristic (ROC) curves were generated for each binary CACS threshold against the Dr.Noon CVD score, reporting area under the curve (AUC) values with 95% CIs. The 95% CI for AUCs was estimated using the DeLong nonparametric method. All analyses were performed in R ver. 4.2.2 (R Foundation for Statistical Computing).
RESULTS
Baseline characteristics by sex
A total of 138,745 participants (67,840 women and 70,905 men) were included in the analysis (Table 1). On average, men were 2 years older than women (46.08±12.94 years vs. 44.40±13.44 years; P<0.001).
General characteristics of subjects according to sex (n=138,745)
Cardiovascular risk factors were more prevalent in men (SBP, 123.38±12.36 mmHg vs. 115.53±14.03 mmHg; DBP, 75.60±10.09 mmHg vs. 68.75±9.60 mmHg; fasting glucose, 100.11±19.07 mg/dL vs. 92.87±14.40 mg/dL; and HbA1c, 5.71%±0.73% vs. 5.54%±0.58%; all P<0.001). Men also exhibited less favorable lipid profiles, with higher total cholesterol (196.49±39.44 mg/dL vs. 194.60±36.01 mg/dL), LDL-C (114.57±35.75 mg/dL vs. 107.61±32.64 mg/dL), and triglycerides (136.27±98.82 mg/dL vs. 90.23±52.85 mg/dL), but lower HDL-C (55.24±13.65 mg/dL vs. 69.00±16.10 mg/dL; all P<0.001).
Prevalence rates of hypertension, DM, and dyslipidemia were also significantly higher in men than women (hypertension, 21.45% vs. 11.21%; DM, 11.41% vs. 5.87%; and dyslipidemia, 36.16% vs. 20.46%; all P<0.001). Finally, Dr.Noon CVD scores were higher in men than in women (29.40±11.24 vs. 22.78±10.42, P<0.001). When stratified by risk groups, 76.33% of women were classified as low risk, 14.32% as moderate risk, and 9.35% as high risk. In comparison, only 56.99% of men were low risk, while 21.97% were moderate risk and 21.04% high risk (all P<0.001) (Table 1).
Baseline characteristics by age group
Table 2 presents baseline characteristics, anthropometric measures (height, weight, BMI, waist circumference), hemodynamic variables (SBP, DBP), metabolic indices (fasting glucose, HbA1c, lipid panel), prevalence of cardiometabolic diseases (hypertension, DM, dyslipidemia, obesity, abdominal obesity), and the distribution of Dr.Noon CVD scores and risk categories across 10 age groups (<30 to ≥70 years). Overall sex distribution was balanced, but the proportion of men varied across the 5-year age groups (36.75% in <30 years to 57.62% in 45–49 years, P<0.001). Height and weight were recorded in each age group, and BMI was calculated accordingly, peaking in the ≥70 years group at 24.38±3.05 kg/m2. Waist circumference increased steadily with age, from 77.33±10.62 cm in those <30 years to 85.95±8.99 cm in participants ≥70 years (all P<0.001).
General characteristics of subjects according to age (n=138,745)
Mean SBP rose steadily with age, from 113.32±12.13 mmHg in the <30 years group to a maximum of 131.34±13.27 mmHg in the ≥70 years group. In contrast, DBP peaked earlier—in the 50–54 years group at 74.53±10.23 mmHg—before gradually declining in older strata (e.g., 71.53±9.46 mmHg in ≥70 years).
Glycemic markers showed a similar age-related trend: fasting glucose was lowest at 89.37±8.64 mg/dL in those <30 years and highest at 106.94±23.42 mg/dL in the ≥70 years group. HbA1c increased from 5.28%±0.32% (<30 years) to 6.16%±0.89% (≥70 years).
Lipid profiles demonstrated an early rise and subsequent plateau with age. Total cholesterol increased from 186.69±31.71 mg/dL (<30 years) to a peak of 203.25±40.69 mg/dL in the 50–54 years group, then declined to 173.47±38.86 mg/dL in those ≥70 years. LDL-C followed a similar pattern, peaking at 117.24±36.83 mg/dL in the 50–54 years group before falling to 94.14±34.93 mg/dL in ≥70 years. HDL-C decreased progressively from 67.18±16.46 mg/dL (<30 years) to 57.20±14.62 mg/dL (≥70 years). Triglycerides rose from 86.32±57.29 mg/dL (<30 years) to a maximum of 129.11±100.83 mg/dL in the 45–49 years group, then tapered modestly in older participants (111.00±58.25 mg/dL in ≥70 years) (all P<0.001).
The prevalence of cardiometabolic diseases also increased with age: hypertension rose from 1.90% in the <30 years group to 52.46% in the ≥70 years group, dyslipidemia from 10.88% to 35.87%, obesity from 20.64% to 40.18%, abdominal obesity from 10.60% to 42.64%, and diabetes from 0.49% to 31.41% (all P<0.001).
Consistent with these patterns, the mean Dr.Noon CVD score rose from 14.21±2.94 in participants <30 years to 46.26±6.52 in those ≥70 years (P<0.001), and the proportion classified as high risk increased sharply from 0.02% to 84.04% across these age groups. Age-specific Dr.Noon CVD score distributions are shown as boxplots in Fig. 1, and age- and sex-specific distributions are presented in Fig. 2.
Distribution of Dr.Noon CVD (Mediwhale Inc) scores across age groups. Boxplots illustrate the age-dependent increase in Dr.Noon CVD scores within a general screening population. The consistent upward shift in medians and interquartile ranges across age groups reflects progressive vascular aging detectable through retinal biomarkers.
Age-stratified distribution of Dr.Noon CVD (Mediwhale Inc) scores by sex. Median Dr.Noon CVD scores (dots) with interquartile ranges (error bars from Q1 to Q3) are plotted across age groups separately for (A) women and (B) men. The background color gradient indicates cardiovascular risk zones based on Dr.Noon CVD thresholds: green (<30, low risk), orange (30–40, moderate risk), and red (>40, high risk). A progressive rise in scores with age is observed in both sexes, with a steeper increase and earlier crossing into higher-risk zones among men, highlighting sex-specific trajectories of retinal vascular aging.
Subgroup analysis of CACS
CACS distributions by age
The mean CACS increased progressively with age, from 6.57±125.88 in participants <30 years to 276.73±573.04 in those ≥70 years. When CACS was categorized, the proportion with any coronary calcification (CACS >0) rose from 2.06% to 76.82%, mild burden (CACS, 1–99) increased from 2.06% to 35.2%, and moderate-to-high burden (CACS ≥100) rose from 0.26% to 41.65% from the youngest to oldest age groups (all P<0.001) (Table 3).
CACS of subjects according to age (n=12,742)
Association with Dr.Noon CVD score
We observed a moderate correlation between raw CACS and Dr.Noon CVD score (r=0.29, P<0.001), and a stronger association after log transformation (r=0.52, P<0.001) (Fig. S2). These findings indicate that CACS correlates well with Dr.Noon CVD, particularly when the skewed distribution of calcium scores is accounted for.
Performance of Dr.Noon CVD score in identifying coronary artery calcification
Fig. 3 presents ROC curves for two binary CACS cutoffs. For any calcification (CACS >0), the AUC was 0.80 (95% CI, 0.80–0.81) (Fig. 3A), and for significant burden (CACS ≥100), it was 0.82 (95% CI, 0.81–0.83) (Fig. 3B). These results demonstrate that the Dr.Noon CVD score discriminates well between individuals with and without coronary calcium, as well as those with clinically significant calcification.
Receiver operating characteristic (ROC) curves assessing the discriminative performance of Dr.Noon CVD (Mediwhale Inc) scores for coronary artery calcium burden. (A) Individuals with any detectable coronary artery calcification (coronary artery calcium score [CACS] >0). (B) Individuals with clinically significant calcification (CACS ≥100). The area under the curve (AUC) was 0.80 (95% confidence interval [CI], 0.80–0.81) for CACS >0 and 0.82 (95% CI, 0.81–0.83) for CACS ≥100, indicating strong predictive accuracy across both thresholds. The diagonal dashed line represents the line of no discrimination (AUC, 0.50).
DISCUSSION
Our analysis of over 138,000 participants from two health screening centers provides a comprehensive, age-stratified portrait of key anthropometric measures (height, weight, BMI, and waist circumference), traditional cardiovascular risk factors (blood pressure, glycemic markers, lipid panels), and cardiometabolic disease prevalence from early adulthood through the oldest age groups. Our cohort is healthier than the general Korean adult population: the age-standardized prevalence rates we observed were 21.45% in men and 11.21% in women for hypertension, 11.41% in men and 5.87% in women for DM, 36.16% in men and 20.46% in women for dyslipidemia , 46.22% in men and 19.18% in women for obesity, and 36.02% in men and 14.73% in women for abdominal obesity. In contrast, nationwide surveys generally report higher rates: 27.3% in men and 17.2% in women for hypertension [18], 15.2% in men and 11.0% in women for DM [19], 47.1% in men and 34.7% in women for dyslipidemia [20], 49.6% in men and 27.7% in women for obesity, and 31.3% in men and 18.0% in women for abdominal obesity [21]. This comparatively low cardiometabolic burden enhances the value of our age- and sex-specific percentile charts, allowing clinicians to benchmark an individual’s Dr.Noon CVD score against a reference group that approximates ideal cardiometabolic health across the adult life course.
Moreover, with 138,745 participants, our dataset is among the largest population-based cohorts assessed for age-stratified cardiovascular metrics. We also conducted additional analyses using CACS, showing that the Dr.Noon CVD score achieved an AUC of 0.82 (95% CI, 0.81–0.83) in identifying significant calcification burden across the full adult age range in a large screening population. By detailing how these parameters evolve from <30 to ≥70 years, our findings provide reference distributions that can inform both population-level surveillance and individualized risk assessment.
Across the age spectrum, we observed distinct, physiologically plausible trends in both anthropometric and cardiometabolic parameters. These age-related shifts largely recapitulate well-established epidemiologic patterns but also highlight features characteristic of East Asian populations. The gradual loss of height after the third decade and the midlife peak in BMI, followed by a plateau, mirror findings from both the National Health and Nutrition Examination Survey cohort [22] and the Korea National Health and Nutrition Examination Survey cohort [23]. Progressive increases in waist circumference underscore the prominence of central adiposity in Koreans, who tend to exhibit higher visceral fat at lower BMIs [24]. The steady rise in SBP alongside a midlife peak in DBP aligns with data on arterial stiffening with aging [25], while the steep increases in glucose and HbA1c after age 50 years reflect β cell decline and insulin resistance seen in older Koreans [26,27]. Lipid trajectories—midlife peaks in total and LDL-C, linear declines in HDL-C, and transient surges in triglycerides—correspond to age- and sex-specific shifts in lipid metabolism reported globally and in Korean surveys [28,29]; in older adults, the subsequent declines in total and LDL-C are also influenced by widespread statin use, which has led to population-level reductions in these lipids over time [30]. Finally, the sharply rising prevalence rates of hypertension, DM, and dyslipidemia underscore the cumulative cardiometabolic burden across our cohort’s lifespan [31].
Similarly, mean Dr.Noon CVD scores rose steadily with each successive age group, mirroring the increase in subclinical atherosclerotic burden as captured by CACS in both prior studies and our own subgroup analyses [32]. In that subgroup, Dr.Noon CVD scores correlated moderately with raw CACS (r=0.29, P<0.001), and more strongly after log transformation (r=0.52, P<0.001), and discriminated both any calcification (AUC, 0.80; 95% CI, 0.80–0.81) and clinically significant burden (AUC, 0.82; 95% CI, 0.81–0.83). These results confirm that Dr.Noon CVD not only tracks the aggregate rise in conventional risk markers with age, but also aligns with CACS, reinforcing its utility as a noninvasive, single-test biomarker of cumulative cardiovascular risk.
This study has several important limitations. First, the cross-sectional design precludes assessment of the Dr.Noon CVD score’s ability to predict hard cardiovascular outcomes such as myocardial infarction or stroke in this dataset. Although we confirmed strong correlations with established surrogate markers (e.g., CACS), these observations remain inherently correlational; prospective studies specifically examining the biomarker’s incremental prognostic value are already underway and will be required to establish causal and clinical relevance. Second, CACS was not measured in all participants, but rather in a subset likely selected for higher clinical risk or greater concern about cardiovascular health. This nonrandom selection introduces potential bias and limits the generalizability of CACS-related subgroup analyses. As such, observed associations between Dr.Noon CVD scores and CACS may overestimate their correlation in the general population. Third, although we report detailed descriptive and univariate analyses, the study adjusts for only a limited number of confounders, and residual confounding may influence the observed relationships. Thus, the conclusions should be interpreted as preliminary and hypothesis-generating, rather than as definitive evidence of clinical benefit. Fourth, our cohort consisted primarily of generally healthy adults undergoing routine health screening, which constrains our ability to define optimal target ranges for the Dr.Noon CVD score or specify actionable thresholds for intervention. Nevertheless, providing personalized CVD risk estimates gives these asymptomatic participants a concrete numerical reference to assess their cardiovascular health and motivate targeted preventive actions. Prospective studies are needed to establish score-change thresholds and to evaluate how Dr.Noon CVD monitoring might inform both primary screening and tailored secondary prevention in diverse risk populations.
In a large, age-diverse Korean screening cohort, we demonstrated that retinal AI-derived Dr.Noon CVD scores closely parallel the well-known trajectories of traditional cardiometabolic risk factors and subclinical atherosclerosis as measured by CACS. By providing a single, noninvasive metric that integrates complex microvascular changes, Dr.Noon CVD offers an additional tool that may enhance cardiovascular risk assessment, particularly for younger adults outside conventional screening age ranges and for populations with distinct metabolic phenotypes. Moreover, using these age- and sex-specific score percentiles (Fig. S3) to develop individualized educational materials could transform consultations, making primary prevention truly patient-centered and tailoring secondary risk management to each individual’s profile. Further studies should evaluate whether integrating retinal AI-derived scores such as Dr.Noon CVD into existing cardiovascular risk prediction models—including traditional clinical risk scores, CACS, or metabolic indices—can improve predictive accuracy for clinical events. Establishing the incremental value of these retinal biomarkers, particularly in diverse populations and across the risk spectrum, will be essential to determine their optimal role in guiding primary prevention and individualized cardiovascular risk management.
Notes
Author contributions
Conceptualization: JC, JS, JP, DN, THP, ST, THR, MJ; Data curation: JS, BC, MJ; Formal analysis: JC, JS; Investigation: JC, MJ; Methodology: JC, MJ; Resources: THR, BC; Software: THR; Supervision: JP, DN, THP, ST, THR, MJ; Visualization: JC; Writing–original draft: JC, JS, MJ; Writing–review & editing: all authors. All authors read and approved the final manuscript.
Conflicts of interest
Jungkyung Cho, Jaewon Seo, Junseok Park, Dongjin Nam, Taehyun Park, Sahil Thakur, Tyler Hyungtaek Rim, and Miso Jang have received honoraria from Mediwhale Inc, and Tyler Hyungtaek Rim owned stocks of Mediwhale Inc. The authors have no other conflicts of interest to declare.
Funding
The authors received no financial support for this study.
Supplementary materials
Fig. S1. Overview of the Dr.Noon CVD (Mediwhale Inc) deep learning pipeline.
cpp-2025-7-e14-Fig-S1.pdfFig. S2. Correlation between Dr.Noon CVD (Mediwhale Inc) scores and coronary artery calcium scores (CACS).
cpp-2025-7-e14-Fig-S2.pdfFig. S3. Age- and sex-specific Dr.Noon CVD score percentiles with cardiometabolic disease prevalence.
cpp-2025-7-e14-Fig-S3.pdfSupplementary materials are available from https://doi.org/10.36011/cpp.2025.7.e14.