AVSeg-XAI: Deep learning framework for A/V segmentation with vascular features reveals retinal oculomics as biomarker for cardiovascular disease Authors: Syed Abdullah Basit, Tanvir Alam Published: 2026 ABSTRACT BackgroundCardiovascular disease (CVD) is the leading cause of mortality across the globe. Conventional risk stratification often depends on invasive biomarkers or generic risk scores that may underperform in key demographics, particularly women and middle-aged adults. Retinal microvasculature provides a non-invasive window into systemic vascular health (oculomics), but its integration with routine clinical vitals for sex-specific risk assessment remains underexplored.Methods We developed AVSeg-XAI, an interpretable DINOv2-based deep-learning pipeline for (i) artery/vein (A/V) segmentation and (ii) geometric/topological feature extraction from retinal fundus images. We then implemented a multimodal approach by integrating interpretable retinal vascular biomarkers and non-invasive clinical vitals into a Random Forest classifier to stratify CVD risk in a Qatar Biobank (QBB) cohort. Performance was assessed using stratified 10-fold cross-validation with subgroup analyses by sex and age. Biological plausibility was evaluated through virtual phenotyping, which predicts systemic traits based solely on retinal features.ResultsFor artery/vein segmentation, the self-supervised AVSeg-XAI achieved a mean Dice score of 0.8368 on the benchmark dataset, exceeding the reported LUNet mean Dice score of 0.8327. Retinal feature extraction module yielded 34 refined vascular features encompassing caliber (CRAE, CRVE, AVR), fractal complexity, tortuosity, and branching topology; a parsimonious 7-feature interpretable subset was also defined for baseline comparison. Clinical inputs were restricted to 15 non-invasive variables obtainable without phlebotomy (e.g., age, BMI, blood pressure). The multimodal fusion model achieved an overall AUC of 0.81 (95% CI: 0.78–0.84). Stratified evaluation demonstrated superior performance in females (AUC 0.83, 95% CI: 0.79–0.87) compared with males (AUC 0.79, 95% CI: 0.75–0.83), with peak performance in the 45–55 year subgroup (AUC 0.90, 95% CI: 0.83–0.97).ConclusionWe present an explainable, non-invasive CVD screening approach that integrates retinal imaging biomarkers with basic clinical vitals. We believe this study advances AI-powered oculomics as a proof-of-concept for community health screening, subject to prospective external validation.Supplementary informationThe online version contains supplementary material available at https://doi.org/10.1186/s13040-026-00573-x. Background Cardiovascular disease (CVD) is the major cause of mortality worldwide, requiring the need for reliable early detection mechanisms. Traditional CVD risk calculators such as the Framingham Risk Score [1], or SCORE2 [2] rely on demographic factors, lifestyle and invasive blood biomarkers such as total cholesterol, HDL etc. Despite their effectiveness, there are limitations for this mechanism such as the need for phlebotomy, laboratory processing expense, and patient’s reluctance to invasive nature of sample collection. Moreover, prior work has highlighted that the that the accuracy of conventional risk scores can vary accross subgroups (including women and younger adults) depending on the population, endpoint definition and basline risk distribution [3]. Our study does not directly benchmark Framingham/SCORE2, rather motivated by the need for complementary non-invasive risk signals. Artificial Intelligence (AI) in medical imaging has created novel avenues for non-invasive risk assessment. One promising direction is “Oculomics,” which leverages the retinal microvasculature to expose the overall health status [4]. The retinal vasculature, including arteries and veins (A/V), shares embryologic and physiologic characteristics with the coronary and cerebral microcirculations, which supports the investigation of retinal vascular measurements as non-invasive biomarkers of systemic vascular health. Arteriolar narrowing, venular dilatation, and changes in fractal dimension are examples of pathological changes in retinal artery geometry that have been independently linked to hypertension, stroke, and incident CVD [5, 6]. Recently, it has been shown that fundus images contain sufficient information to be considered as an estimator for multiple cardio-metabolic biomarkers [7, 8]. However, the clinical translation of these AI-driven oculomic models faces four key barriers. Recent surveys have reviewed fundus-image processing pipelines for identifying ocular, neurological, and cardiovascular disorders, and summarized publicly available datasets and remaining challenges [9]. First, Accurate retinal vessel extraction is a key prerequisite for quantitative vascular biomarkers, yet remains technically challenging due to the fine caliber and variable appearance of retinal vessels; comprehensive taxonomies and challenges have been summarized in prior reviews [10]. Second, many end-to-end deep learning models operate as “black boxes”, producing embeddings with no clear biological meaning. This lack of transparency weakens the clinical trust and undermines regulatory approval because clinicians are unable to explain why a patient was flagged as high risk. Third, retinal imaging are being used in isolation by most studies, ignoring the potential of multimodal fusion using the combination of easily obtainable vitals such as blood pressure and BMI. Recent work has shown that retinal fundus image alone could predict cardiometabolic biomarkers [8], yet these approaches have not been systematically linked with clinical vitals in a multimodal framework. Fourth, generalizability remains challenging because many oculomics studies are evaluated in limited cohorts and acquisition settings; therefore, performance may not translate across imaging devices, protocols, and diverse populations. Considering the above backdrop, this study uses non-invasive features to develop an interpretable multimodal AI framework for CVD risk classification. Our framework is built upon three major tasks: (1) We develop an automated atery/vein (A/V) segmentation pipeline to extract geometric features of fundus image, i.e., vessel caliber, tortuosity, fractal dimension, etc. (2) We used multimodal fusion at the feature level, combining retinal vascular biomarkers with 15 non-invasive clinical parameters capturing both metabolic risk and structural vascular damage. (3) We conduct a comprehensive stratified analysis across gender and age subgroups to find “hidden cohorts” where AI-driven oculomics offer better results. We hypothesized that retinal vascular biomarkers capture microvascular structural changes (e.g., caliber alteration, rarefaction, tortuosity, branching complexity) that are complementary to non-invasive clinical vitals reflecting systemic and metabolic risk context (e.g., blood pressure, adiposity, age). Under this hypothesis, feature-level multimodal fusion should improve CVD discrimination relative to either modality alone while retaining interpretability. The contribution of this work can be summarized as follows:We developed a DINOv2-based two-stage self-supervised modeling approach for retinal A/V segmentation.We implemented an interpretable pipeline to extract retinal biomarkers and quantify caliber, geometric, and topological vascular features from A/V segmentation outputs.Integrating retinal biomarkers with common vitals, the AI model achieved an AUC of 0.81 for predicting CVD risk.Superior performance in females (AUC 0.83), 45–55 age group (AUC 0.90) highlights the significance of age- and gender-stratified modeling in a clinical setup. We developed a DINOv2-based two-stage self-supervised modeling approach for retinal A/V segmentation. We developed a DINOv2-based two-stage self-supervised modeling approach for retinal A/V segmentation. We implemented an interpretable pipeline to extract retinal biomarkers and quantify caliber, geometric, and topological vascular features from A/V segmentation outputs. We implemented an interpretable pipeline to extract retinal biomarkers and quantify caliber, geometric, and topological vascular features from A/V segmentation outputs. Integrating retinal biomarkers with common vitals, the AI model achieved an AUC of 0.81 for predicting CVD risk. Integrating retinal biomarkers with common vitals, the AI model achieved an AUC of 0.81 for predicting CVD risk. Superior performance in females (AUC 0.83), 45–55 age group (AUC 0.90) highlights the significance of age- and gender-stratified modeling in a clinical setup. Superior performance in females (AUC 0.83), 45–55 age group (AUC 0.90) highlights the significance of age- and gender-stratified modeling in a clinical setup. Methods Study design and population A retrospective case–control cohort was obtained from the Qatar Biobank (QBB) for this study. Qatar Biobank includes a substantially larger participant pool than used in this study; here we analyze a curated subset selected from QBB participants with available fundus imaging and required clinical metadata. The initial cohort pull targeted 1,000 participants (500 CVD cases and 500 controls), with age and sex-matched sampling to support balanced model development and subgroup analysis. After image-quality filtering and exclusion of participants with missing required metadata, the final analyzed cohort comprised 911 participants (453 CVD cases, 458 controls). Ethical approval was obtained from the Institutional Review Board (IRB) and all data were anonymized before analysis. In this study, the outcome label corresponds to prevalent CVD status in the QBB-derived records and was derived from a combination of questionnaire-based self-report and doctor-diagnosed confirmation fields available in the cohort metadata. The case group included participants with major cardiovascular disease (e.g., major heart disease events) and a hypertension-associated state but free from cancer, obesity and diabetes; while controls were defined as participants without a personal history of CVD and without current medications for hypertension or diabetes [11]. Because the outcome is derived from QBB questionnaire/metadata fields rather than centrally adjudicated clinical endpoints, detailed subtype definitions and event dates are not uniformly available within this study. Data acquisition Retinal fundus photography was captured using TopCon-TRC-NW camera at QBB premises. Images were captured centered on the optic disc (OD) and macula. We used OD-centered images to maximize the visibility of the major vascular network. Image quality was assessed after patient selection form QBB using an automated multi-metric pipeline (evaluating sharpness, contrast, blur, brightness, and border artefacts) [12]. Of the 1,850 OD-centered images initially available (from 926 candidate patients, two images per patient corresponding to left and right eyes), 69 images (3.7%) were flagged for insufficient quality and excluded. Quality filtering was applied at the image level: patients for whom at least one eye passed quality control were retained and their passing image(s) used for analysis. Patients for whom both eyes failed quality control (n = 15) were excluded entirely. The final study cohort thus comprised 911 patients contributing 1,781 OD-centered images (870 patients with both eyes, 41 with one eye) (Table 1).Table 1Cohort and image accounting after exclusions. OD-centered fundus images were screened using automated quality control (QC). Patients were retained if at least one eye passed QCStagePatients (n)OD-centered images (n)Initial candidate cohort (QBB extract)9261,850Flagged by automated QC (image-level exclusion)–69Patients with both eyes failing QC (excluded)15–Final analyzed cohort9111,781 Both eyes passed QC8701,740 Only one eye passed QC4141 Table 1Cohort and image accounting after exclusions. OD-centered fundus images were screened using automated quality control (QC). Patients were retained if at least one eye passed QCStagePatients (n)OD-centered images (n)Initial candidate cohort (QBB extract)9261,850Flagged by automated QC (image-level exclusion)–69Patients with both eyes failing QC (excluded)15–Final analyzed cohort9111,781 Both eyes passed QC8701,740 Only one eye passed QC4141 Cohort and image accounting after exclusions. OD-centered fundus images were screened using automated quality control (QC). Patients were retained if at least one eye passed QC Clinical variables Non-invasive clinical parameters were collected from the dataset collected from participants at QBB. To ensure applicability for community screening, we focus only to 15 non-invasive variables which does not require phlebotomy. These included demographic factors (i.e., age and gender), anthropometric measurements (i.e., height, weight, body mass index, waist and hip circumference, waist-to-hip ratio), vital signs (i.e., systolic blood pressure, diastolic blood pressure, and resting heart rate), physical function assessment (i.e., average grip strength), and habit (i.e., smoking status categorized as never/former/current, family history (i.e., coronary artery disease, and diabetes). AVSeg-XAI: AI-driven oculomic feature extraction pipeline We developed AVSeg-XAI, a two-stage AI pipeline for retinal artery/vein (A/V) segmentation from optic disc-centered fundus images, followed by automated extraction of interpretable retinal vascular features for downstream CVD risk stratification. Stage 1: Deep learning-based vessel segmentation and A/V classification In contrast to standard UNet [13], we developed a two-phase ConvNeXt-V2 architecture [14] transfer learning approach for A/V segmentation:Large-Scare Pretraining: We pretrained the encoder with DINOv2 self-distillation on more than > 90,000 fundus images drawn from EyePACS [15], APTOS 2019 [16] and Messidor-2 [17, 18]. This step was designed to encourage robust retinal representations across different imaging devices and disease states. Full detailed parameter details are provided in the Supplementary Material.Task-specific Fine-tuning: We fine-tuned the pretrained encoder for pixel-level artery and vein (A/V) segmentation using the UZLF (Leuven–Haifa) benchmark dataset [19, 20], which comprises 240 disc-centered fundus images from 224 patients split into 156/28/56 train/validation/test images with no patient overlap; dataset demographics and the predefined split are detailed in Supplementary Table S1. We followed the same test-time augmentation protocol as LUNet [21] for fair comparison. To assess whether differences in Dice scores between our method and LUNet are statistically significant, we performed paired testing on per-image Dice scores over the UZLF test set using a two-sided Wilcoxon signed-rank test (α = 0.05), and report 95% confidence intervals for the mean paired difference. The segmentation model followed a UNet–style design with a customized encoder and a plain decoder with deep supervision. We initialized the encoder from DINOv2 weights and trained the decoder from scratch. Figure 1 summarizes the architecture, with full hyperparameters reported in the Supplementary Material.Fig. 1Fine-tuning architecture for artery/vein segmentation Large-Scare Pretraining: We pretrained the encoder with DINOv2 self-distillation on more than > 90,000 fundus images drawn from EyePACS [15], APTOS 2019 [16] and Messidor-2 [17, 18]. This step was designed to encourage robust retinal representations across different imaging devices and disease states. Full detailed parameter details are provided in the Supplementary Material. Large-Scare Pretraining: We pretrained the encoder with DINOv2 self-distillation on more than > 90,000 fundus images drawn from EyePACS [15], APTOS 2019 [16] and Messidor-2 [17, 18]. This step was designed to encourage robust retinal representations across different imaging devices and disease states. Full detailed parameter details are provided in the Supplementary Material. Task-specific Fine-tuning: We fine-tuned the pretrained encoder for pixel-level artery and vein (A/V) segmentation using the UZLF (Leuven–Haifa) benchmark dataset [19, 20], which comprises 240 disc-centered fundus images from 224 patients split into 156/28/56 train/validation/test images with no patient overlap; dataset demographics and the predefined split are detailed in Supplementary Table S1. We followed the same test-time augmentation protocol as LUNet [21] for fair comparison. To assess whether differences in Dice scores between our method and LUNet are statistically significant, we performed paired testing on per-image Dice scores over the UZLF test set using a two-sided Wilcoxon signed-rank test (α = 0.05), and report 95% confidence intervals for the mean paired difference. The segmentation model followed a UNet–style design with a customized encoder and a plain decoder with deep supervision. We initialized the encoder from DINOv2 weights and trained the decoder from scratch. Figure 1 summarizes the architecture, with full hyperparameters reported in the Supplementary Material.Fig. 1Fine-tuning architecture for artery/vein segmentation Task-specific Fine-tuning: We fine-tuned the pretrained encoder for pixel-level artery and vein (A/V) segmentation using the UZLF (Leuven–Haifa) benchmark dataset [19, 20], which comprises 240 disc-centered fundus images from 224 patients split into 156/28/56 train/validation/test images with no patient overlap; dataset demographics and the predefined split are detailed in Supplementary Table S1. We followed the same test-time augmentation protocol as LUNet [21] for fair comparison. To assess whether differences in Dice scores between our method and LUNet are statistically significant, we performed paired testing on per-image Dice scores over the UZLF test set using a two-sided Wilcoxon signed-rank test (α = 0.05), and report 95% confidence intervals for the mean paired difference. The segmentation model followed a UNet–style design with a customized encoder and a plain decoder with deep supervision. We initialized the encoder from DINOv2 weights and trained the decoder from scratch. Figure 1 summarizes the architecture, with full hyperparameters reported in the Supplementary Material.Fig. 1Fine-tuning architecture for artery/vein segmentation Fine-tuning architecture for artery/vein segmentation Stage 2: Interpretable geometric feature extraction We automatically localized the optic disc region using PVBM [22] and formed a region of interest (ROI) centered on the disc in order to standarize vascular measurements and reduce inter-image variability. Caliber metrics (CRAE/CRVE) were calculated within Zone B according to established retinal vessel measurement protocols, which is defined as an annulus spanning 0.5–1.0 disc diameters from the optic disc margin. Figure 2 shows the optic disc localization and the Zone B annulus used for subsequent caliber estimation.Fig. 2Stages for automated retinal feature extraction pipeline. (A) Original cropped fundus image centered on the optic disc. (B) Automated optic disc detection (green circle) used to define the region of interest (ROI). (C) Artery–vein segmentation overlay. The yellow annulus defines Zone B (0.5–1.0 disc diameters from the optic disc margin), the standard region for calculating caliber metrics (CRAE/CRVE) to minimize variability Stages for automated retinal feature extraction pipeline. (A) Original cropped fundus image centered on the optic disc. (B) Automated optic disc detection (green circle) used to define the region of interest (ROI). (C) Artery–vein segmentation overlay. The yellow annulus defines Zone B (0.5–1.0 disc diameters from the optic disc margin), the standard region for calculating caliber metrics (CRAE/CRVE) to minimize variability Robustness considerations. Derived vascular features differ in sensitivity to imaging variability and segmentation noise. Caliber measures (CRAE/CRVE) and AVR are computed within Zone B (defined in disc-diameter units) to reduce inter-image magnification and acquisition variability by focusing on major peripapillary vessels. Global summary measures (e.g., artery/vein fractal metrics) are less affected by small local boundary perturbations, but may change if thin peripheral branches are systematically missed. In contrast, topology- and curvature-driven features (e.g., intersection/junction counts, branching angles, and tortuosity measures, particularly on thin vessels) are more sensitive to segmentation artifacts because small false positives/negatives can create or remove branches during skeletonization and topology estimation. Then we extracted a list of quantitative caliber metrics [6, 23] such as central retinal venular equivalent (CRVE), central retinal arteriolar equivalent (CRAE). We also calculated the arteriolar-to-venular ratio (AVR) which quantifies the narrowing or dilation associated with hypertensive remodeling in retina. Fractal dimension (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_f$$\end{document}) was also measured through the box-counting method (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_0$$\end{document}), and information dimension (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_1$$\end{document}) were employed for assessing vascular branching complexity, yielding a global index of microvascular density and perfusion efficiency [24]. Vascular twistiness was calculated by measuring the vessel curvature through tortuosity. For this purpose we used arc–chord ratio and integral curvature [25]. In order to measure the vascular tree’s structural design, branching ‘angles and junction density were also computed. In totality, a set of 34 geometric features were extracted. Additionally, a Basic subset of seven highly interpretable features (CRAE, CRVE, AVR, artery/vein fractal dimension, and artery/vein tortuosity) was used for baseline comparison. Feature selection rationale. We selected retinal vascular features to capture complementary and clinically interpretable axes of microvascular remodeling relevant to CVD: (i) caliber and AVR to quantify narrowing/dilation patterns associated with hypertensive remodeling; (ii) network density/complexity (e.g., fractal/capacity/entropy and intersection/junction measures) to characterize rarefaction and altered vascular organization; and (iii) tortuosity/branch geometry to reflect hemodynamic and endothelial stress. Features were computed within a standardized optic disc–centered ROI (Zone B for caliber) to reduce measurement variability. The selected feature families are further supported empirically in this study by significant case–control differences (Table 6) and by biological validation/interpretability analyses (virtual phenotyping and SHAP). Multimodal AI framework for CVD risk stratification Clinical data preprocessing Clinical features with missing values not exceeding 5% of total data points were imputed using median imputation; features with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${ > }5\%$$\end{document} missing values would have been excluded entirely. In practice, all 15 clinical features fell below this threshold (maximum observed missingness: 0.8% for grip strength and 0.2% for resting heart rate), so no feature was excluded. All continuous variables were standardized using z-score normalization. Categorical variables, including gender, smoking status, and family history, were encoded as integers. CVD detection model For the detection of CVD we fused explainable oculomics feature along with non-invasive clinical features. A Random Forest classifier was used as the machine learning model. Hyperparameters were selected via 5-fold cross-validated grid search over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n\_estimators \in \{100, 200, 300\}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathit{max\_depth} \in \{\mathrm{None}, 10, 20\}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathit{min\_samples\_split} \in \{2, 5, 10\}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathit{min\_samples\_leaf} \in \{1, 2, 4\}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathit{max\_features} \in \{\mathrm{sqrt}, \mathrm{log2}\}$$\end{document}, optimising AUC-ROC. The resulting best configuration (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n\_estimators{=}100$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathit{max\_depth}{=}\mathrm{None}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathit{min\_samples\_split}{=}2$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathit{min\_samples\_leaf}{=}1$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathit{max\_features}{=}\mathrm{sqrt}$$\end{document}) was adopted for all reported experiments. Class imbalance was addressed via class_weight=‘balanced’, which adjusts sample weights inversely proportional to class frequencies. As a sensitivity analysis, we also evaluated logistic regression and histogram-based gradient boosting classifiers under the same cross-validation protocol (Supplementary Table S5). To comprehensively assess the contribution of each feature set, we evaluated five configurations: (a) Oculomics-only (full, 34), using all 34 retinal features; (b) Oculomics-only (basic, 7), using the interpretable 7-feature retinal subset; (c) Clinical-only, using 15 non-invasive clinical variables; (d) Multimodal Fusion (interpretable), combining the 7-feature retinal subset with 15 clinical variables; and (e) Multimodal Fusion (full), combining all 34 retinal features with 15 clinical variables. The interpretable fusion (d) is the primary model reported; configurations (b) and (e) are included for completeness. Validation strategy Using 10-fold cross-validation (CV) approach to evaluate the model performance. The area under the Receiver Operating Characteristic Curve (AUC-ROC) was used as the primary metric for discriminative ability. Accuracy, Sensitivity, and Specificity for assessing clinical utility at the optimal threshold (Youden’s index) and 95% confidence intervals to quantify uncertainty, calculated via bootstrapping (1000 iterations). Because clinical deployment often targets high sensitivity, we additionally report specificity at a fixed sensitivity (0.80) as a screening-oriented operating point (Supplementary Table S6). Random Forest (RF) was pre-specified as the primary classifier for all discriminative tasks rather than being selected post-hoc. This choice was based on four principled criteria: (1) hyperparameter robustness, as RF performs well with default settings and reduces optimization bias on modest-sized cohorts; (2) interpretability alignment, owing to its native Gini importances and seamless compatibility with SHAP analysis; (3) strong precedent in the clinical prediction literature; and (4) suitability for sensitivity analysis against continuous parametric methods (like Logistic Regression) and sequential boosting algorithms (like Histogram-based Gradient Boosting), which are evaluated under identical conditions in Supplementary Table S5. Statistical analysis and model interpretation Stratified performance analysis Comprehensive stratified analysis was performed across (a) Gender: Male vs. Female, (b) Age Groups: < 45 years and 45–55 years, and (c) BMI categories: Normal weight ( < 25), overweight (25–30), and obese ( > 30). Using DeLong’s test, statistical significance was assessed between these subgroups for comparison of AUC values. To control the family-wise error rate across simultaneous comparisons, Benjamini–Hochberg (BH) false discovery rate (FDR) correction was applied to the following comparison families: (i) all 34 retinal features tested simultaneously in the univariate CVD vs. control analysis (Table 6 reports the 7 features that reach significance after correction); (ii) the 3 paired segmentation Dice comparisons (overall, artery, vein; Table 3); and (iii) the DeLong pairwise subgroup AUC comparisons. BH-adjusted q-values are reported alongside raw p-values in the respective tables and footnotes. The two primary model-configuration pairwise tests (multimodal vs. clinical-only; multimodal vs. oculomics-only) are reported with raw DeLong p-values given the reduced multiplicity in that family.Table 2Baseline characteristics of study participants. fasting glucose is included for cohort characterization only and was not used as a model input; all classifier inputs were restricted to the 15 non-invasive variables described in Sect. 2.3.CharacteristicControl (n=458)CVD Case (n=453)p-valueFemale gender221 (48.3%)215 (47.5%)0.863BMI (kg/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document})28.9±5.032.1±6.3 < 0.001Systolic BP (mmHg)110.3±10.6123.7±16.2 < 0.001Diastolic BP (mmHg)67.9±9.475.6±11.7 < 0.001Fasting glucose (mmol/L)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\dagger}$$\end{document}4.9±0.65.8±2.7 < 0.001Total cholesterol (mmol/L)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\dagger}$$\end{document}5.0±0.95.1±1.00.051Current smoker107 (23.4%)123 (27.2%)0.215Values are presented as mean ± SD for continuous variables and n (%) for categorical variables. p-values were calculated using an independent t-test for continuous variables and a chi-square test for categorical variables. †Reported for cohort characterization only; not used as a model inputTable 3Comparison with state-of-the-art methods on the UZLF test set (Dice coefficient). values in parentheses for Ours and LUNet are bootstrap 95% confidence intervals (10,000 iterations, per-image UZLF test data, n = 56). BFN, VC-Net, and junior annotator CIs are from [21]ModelArtery DiceVein DiceMean DiceOurs (AVSeg-XAI)0.826 (0.816–0.835)0.848 (0.839–0.856)0.837 (0.828–0.845)LUNet [21]0.820 (0.809–0.830)0.845 (0.837–0.853)0.832 (0.824–0.841)Junior Annotator [21]0.815 (0.780–0.844)0.837 (0.809–0.863)0.826VC-Net [26]0.783 (0.768–0.797)0.819 (0.810–0.829)0.801BFN [27]0.781 (0.763–0.797)0.808 (0.795–0.819)0.795Ours vs. LUNet: mean paired Dice difference +0.004 overall (artery +0.006, vein +0.002); two-sided Wilcoxon signed-rank test p = 0.0004 (overall), p = 0.0002 (artery), p = 0.027 (vein); 95% CI for mean paired difference: overall [+0.002, +0.007], artery [+0.003, +0.009], vein [+0.000, +0.005]. Benjamini–Hochberg FDR correction applied across the 3 simultaneous Dice comparisons: adjusted q = 0.0006 (overall), q = 0.0006 (artery), q = 0.027 (vein); all survive FDR correction at α = 0.05. VC-Net and BFN 95% CIs from [21] bootstrap resampling; paired significance testing not possible as per-image scores are not publicly available for those methods Table 2Baseline characteristics of study participants. fasting glucose is included for cohort characterization only and was not used as a model input; all classifier inputs were restricted to the 15 non-invasive variables described in Sect. 2.3.CharacteristicControl (n=458)CVD Case (n=453)p-valueFemale gender221 (48.3%)215 (47.5%)0.863BMI (kg/\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^2$$\end{document})28.9±5.032.1±6.3 < 0.001Systolic BP (mmHg)110.3±10.6123.7±16.2 < 0.001Diastolic BP (mmHg)67.9±9.475.6±11.7 < 0.001Fasting glucose (mmol/L)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\dagger}$$\end{document}4.9±0.65.8±2.7 < 0.001Total cholesterol (mmol/L)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{\dagger}$$\end{document}5.0±0.95.1±1.00.051Current smoker107 (23.4%)123 (27.2%)0.215Values are presented as mean ± SD for continuous variables and n (%) for categorical variables. p-values were calculated using an independent t-test for continuous variables and a chi-square test for categorical variables. †Reported for cohort characterization only; not used as a model input Baseline characteristics of study participants. fasting glucose is included for cohort characterization only and was not used as a model input; all classifier inputs were restricted to the 15 non-invasive variables described in Sect. 2.3. Values are presented as mean ± SD for continuous variables and n (%) for categorical variables. p-values were calculated using an independent t-test for continuous variables and a chi-square test for categorical variables. †Reported for cohort characterization only; not used as a model input Table 3Comparison with state-of-the-art methods on the UZLF test set (Dice coefficient). values in parentheses for Ours and LUNet are bootstrap 95% confidence intervals (10,000 iterations, per-image UZLF test data, n = 56). BFN, VC-Net, and junior annotator CIs are from [21]ModelArtery DiceVein DiceMean DiceOurs (AVSeg-XAI)0.826 (0.816–0.835)0.848 (0.839–0.856)0.837 (0.828–0.845)LUNet [21]0.820 (0.809–0.830)0.845 (0.837–0.853)0.832 (0.824–0.841)Junior Annotator [21]0.815 (0.780–0.844)0.837 (0.809–0.863)0.826VC-Net [26]0.783 (0.768–0.797)0.819 (0.810–0.829)0.801BFN [27]0.781 (0.763–0.797)0.808 (0.795–0.819)0.795Ours vs. LUNet: mean paired Dice difference +0.004 overall (artery +0.006, vein +0.002); two-sided Wilcoxon signed-rank test p = 0.0004 (overall), p = 0.0002 (artery), p = 0.027 (vein); 95% CI for mean paired difference: overall [+0.002, +0.007], artery [+0.003, +0.009], vein [+0.000, +0.005]. Benjamini–Hochberg FDR correction applied across the 3 simultaneous Dice comparisons: adjusted q = 0.0006 (overall), q = 0.0006 (artery), q = 0.027 (vein); all survive FDR correction at α = 0.05. VC-Net and BFN 95% CIs from [21] bootstrap resampling; paired significance testing not possible as per-image scores are not publicly available for those methods Comparison with state-of-the-art methods on the UZLF test set (Dice coefficient). values in parentheses for Ours and LUNet are bootstrap 95% confidence intervals (10,000 iterations, per-image UZLF test data, n = 56). BFN, VC-Net, and junior annotator CIs are from [21] Ours vs. LUNet: mean paired Dice difference +0.004 overall (artery +0.006, vein +0.002); two-sided Wilcoxon signed-rank test p = 0.0004 (overall), p = 0.0002 (artery), p = 0.027 (vein); 95% CI for mean paired difference: overall [+0.002, +0.007], artery [+0.003, +0.009], vein [+0.000, +0.005]. Benjamini–Hochberg FDR correction applied across the 3 simultaneous Dice comparisons: adjusted q = 0.0006 (overall), q = 0.0006 (artery), q = 0.027 (vein); all survive FDR correction at α = 0.05. VC-Net and BFN 95% CIs from [21] bootstrap resampling; paired significance testing not possible as per-image scores are not publicly available for those methods Virtual phenotyping Distinguishing genuine physiological signals from spurious correlations is a difficult challenge in AI-driven biomarker discovery. Hence, to provide supporting evidence that the retinal features are encoding real systemic vascular information “virtual phenotyping” experiment was performed where using only retinal features as input a random forest regressor was trained specifically to predict systemic biomarkers (Systolic blood pressure, Diastolic blood pressure, BMI, Smoking status). Cofficient of determination (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document}) was used to quantify the strength of the retinal-systemic link. Strong predictive strength for blood pressure would demonstrate that the model learns patterns associated with hypertensive retinopathy; yet, a limited prediction of BMI would imply orthogonality between retinal and metabolic variables, supporting the interpretation that retinal and clinical variables provide complementary information in the fusion model. Explainable AI by SHAP analysis A game-theoretic approach, SHAP (SHapley Additive exPlanations), was employed to enhance clinical interpretability, which quantifies the contribution of each feature to individual predictions, enabling both global feature ranking and patient-specific explanations [28]. Global feature importance by ranking features using the mean absolute SHAP value across all patients, and Dependence plots for visualizing non-linear relationships between feature values and risk predictions with color-coding to reveal feature interactions were generated. Results Baseline characteristics The study population was based on 1,000 participants, balanced between CVD cases (n=500) and healthy controls (n=500). After image-quality filtering and exclusion of missing required metadata, 911 participants remained for modeling (453 cases / 458 controls). CVD group exhibited significantly higher BMI, blood pressure, and fasting glucose levels compared to controls (p < 0.001). The age for Controls and CVD were 39.2 ± 3.0 and 40.8 ± 4.3, respectively. Baseline demographic and clinical characteristics of cases and controls are summarized in Table 4. A/V segmentation benchmarking on the UZLF dataset Table 3 presents a multi-model comparison on the UZLF test set. Our AVSeg-XAI model achieved a mean Dice coefficient of 0.8368, outperforming all compared methods. Against the current state-of-the-art LUNet [21] (mean Dice 0.8327), the improvement is statistically significant (paired Wilcoxon signed-rank test, two-sided): overall mean paired Dice +0.004 (95% CI: [+0.002, +0.007], p=0.0004), artery +0.006 ([+0.003, +0.009], p=0.0002), vein +0.002 ([+0.000, +0.005], p=0.027). Relative to earlier baselines evaluated on the same benchmark by [21], our model surpasses VC-Net [26] (mean Dice 0.801) and BFN [27] (mean Dice 0.795) by larger margins, and also exceeds the junior human annotator inter-rater Dice of 0.826. All segmentation metrics are reported on the labeled UZLF benchmark; because the QBB cohort does not provide manual A/V segmentation labels, we cannot directly quantify segmentation accuracy on the target cohort. Table 4 further confirms that DINOv2 self-supervised pretraining provides statistically significant improvements over both ImageNet-supervised (mean Dice: 0.8336, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p=9.3\times10^{-3}$$\end{document}) and random initialization (mean Dice: 0.8284, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p=6.4\times10^{-8}$$\end{document}), demonstrating a clear monotonic benefit from pretraining quality. Full per-metric results for AVSeg-XAI on the UZLF test set (IoU, Precision, Recall, and Specificity in addition to Dice) are reported in Table 5.Table 4Ablation study: effect of encoder pretraining on A/V segmentation performance (UZLF test set). all models share the same architecture, decoder, loss function, augmentation, and TTA protocol; only the encoder initialization differs. Statistical significance assessed via two-sided Wilcoxon signed-rank test on per-image Dice (n=56, α=0.05)Encoder Init.Artery DiceVein DiceMean DiceMean IoUp-value (vs. Ours)Random (no pretraining)0.8155 ± 0.0430.8413 ± 0.0320.82840.7090\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.4 \times 10^{-8}$$\end{document}ImageNet-22k (FCMAE)0.8212 ± 0.0390.8459 ± 0.0320.83360.7164\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.3 \times 10^{-3}$$\end{document}DINOv2 (Ours)0.8258 ± 0.0380.8478 ± 0.0320.83680.7210—Table 5Performance metrics on the UZLF test setMetricArtery (Mean ± Std)Vein (Mean ± Std)Overall MeanDice Coefficient0.8258 ± 0.0380.8478 ± 0.0320.8368IoU0.7050 ± 0.0530.7371 ± 0.0470.7210Precision0.8312 ± 0.0650.8697 ± 0.0540.8505Recall0.8253 ± 0.0520.8316 ± 0.0540.8285Specificity0.9938 ± 0.0020.9946 ± 0.0020.9942 Table 4Ablation study: effect of encoder pretraining on A/V segmentation performance (UZLF test set). all models share the same architecture, decoder, loss function, augmentation, and TTA protocol; only the encoder initialization differs. Statistical significance assessed via two-sided Wilcoxon signed-rank test on per-image Dice (n=56, α=0.05)Encoder Init.Artery DiceVein DiceMean DiceMean IoUp-value (vs. Ours)Random (no pretraining)0.8155 ± 0.0430.8413 ± 0.0320.82840.7090\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$6.4 \times 10^{-8}$$\end{document}ImageNet-22k (FCMAE)0.8212 ± 0.0390.8459 ± 0.0320.83360.7164\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$9.3 \times 10^{-3}$$\end{document}DINOv2 (Ours)0.8258 ± 0.0380.8478 ± 0.0320.83680.7210— Ablation study: effect of encoder pretraining on A/V segmentation performance (UZLF test set). all models share the same architecture, decoder, loss function, augmentation, and TTA protocol; only the encoder initialization differs. Statistical significance assessed via two-sided Wilcoxon signed-rank test on per-image Dice (n=56, α=0.05) Table 5Performance metrics on the UZLF test setMetricArtery (Mean ± Std)Vein (Mean ± Std)Overall MeanDice Coefficient0.8258 ± 0.0380.8478 ± 0.0320.8368IoU0.7050 ± 0.0530.7371 ± 0.0470.7210Precision0.8312 ± 0.0650.8697 ± 0.0540.8505Recall0.8253 ± 0.0520.8316 ± 0.0540.8285Specificity0.9938 ± 0.0020.9946 ± 0.0020.9942 Performance metrics on the UZLF test set Retinal biomarker analysis A univariate statistical analysis was performed to determine whether specific retinal features differed significantly in the CVD vs. healthy controls comparison. As depicted in Table 6, caliber metrics (i.e., CRAE, CRVE, and AVR) exhibit significant differences (p < 0.01) . This result is in line with generalized vascular narrowing in the CVD group reported in [29]. Multiple geometric features (i.e., count of artery intersection, fractal capacity of artery, entropy) also differed significantly in this univariate comparison, suggesting rarefaction or altered vascular organization in the CVD group compared to the control.Table 6Statistical analysis of retinal biomarkers (CVD vs. control)Feature CategoryFeatureCVD (Mean±SD)Control (Mean±SD)P-valueBHq-valueCaliberCRAE (Arteriolar Equivalent)130.2±17.8136.6±15.7 < 0.001 < 0.001CRVE (Venular Equivalent)172.5±23.6177.2±23.00.0030.011AVR (Artery–Vein Ratio)0.76±0.090.78±0.07 < 0.0010.004Advanced GeometryArtery Intersections (Count)65.6±21.572.1±21.2 < 0.001 < 0.001Artery\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_0$$\end{document}Capacity (Fractal)1.360±0.0531.371±0.041 < 0.0010.001Vein Tortuosity Index1.102±0.0141.099±0.0100.0020.008Artery\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_1$$\end{document}Entropy1.340±0.0521.352±0.039 < 0.0010.001Only features with statistically significant differences (p < 0.05) are shown; these represent a subset of the 34 retinal features tested. Benjamini–Hochberg FDR correction was applied across all 34 retinal features tested simultaneously; BH-adjusted q-values are reported in the final column. All 7 listed features survive FDR correction (q ≤ 0.011, the worst case being CRVE at q = 0.011). Reduced artery intersections and D0 capacity in CVD cases suggest rarefaction or reduced vascular complexity Table 6Statistical analysis of retinal biomarkers (CVD vs. control)Feature CategoryFeatureCVD (Mean±SD)Control (Mean±SD)P-valueBHq-valueCaliberCRAE (Arteriolar Equivalent)130.2±17.8136.6±15.7 < 0.001 < 0.001CRVE (Venular Equivalent)172.5±23.6177.2±23.00.0030.011AVR (Artery–Vein Ratio)0.76±0.090.78±0.07 < 0.0010.004Advanced GeometryArtery Intersections (Count)65.6±21.572.1±21.2 < 0.001 < 0.001Artery\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_0$$\end{document}Capacity (Fractal)1.360±0.0531.371±0.041 < 0.0010.001Vein Tortuosity Index1.102±0.0141.099±0.0100.0020.008Artery\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_1$$\end{document}Entropy1.340±0.0521.352±0.039 < 0.0010.001Only features with statistically significant differences (p < 0.05) are shown; these represent a subset of the 34 retinal features tested. Benjamini–Hochberg FDR correction was applied across all 34 retinal features tested simultaneously; BH-adjusted q-values are reported in the final column. All 7 listed features survive FDR correction (q ≤ 0.011, the worst case being CRVE at q = 0.011). Reduced artery intersections and D0 capacity in CVD cases suggest rarefaction or reduced vascular complexity Statistical analysis of retinal biomarkers (CVD vs. control) Only features with statistically significant differences (p < 0.05) are shown; these represent a subset of the 34 retinal features tested. Benjamini–Hochberg FDR correction was applied across all 34 retinal features tested simultaneously; BH-adjusted q-values are reported in the final column. All 7 listed features survive FDR correction (q ≤ 0.011, the worst case being CRVE at q = 0.011). Reduced artery intersections and D0 capacity in CVD cases suggest rarefaction or reduced vascular complexity CVD model performance Table 7 summarizes the comparative performance of all model configurations. The primary multimodal fusion model (7 interpretable retinal features + 15 clinical variables) achieved an AUC of 0.81 (95% CI: 0.78–0.84), outperforming the clinical-only baseline (AUC 0.80, 95% CI: 0.77–0.83). A DeLong test on the out-of-fold predictions yielded a difference in AUC of +0.013 (95% CI for the difference: -0.002 to +0.027; z=1.73, p=0.084). This difference did not reach conventional statistical significance (α=0.05), and the overlapping confidence intervals confirm the modest and uncertain magnitude of the gain. The retinal features therefore provide an incremental rather than a definitively superior improvement in discrimination; the clinical relevance of this increment is further contextualized in the Discussion and in the operating-point analysis (Supplementary Table S5), where the multimodal model yields a specificity gain of +10.5 percentage points over the clinical-only baseline at a fixed sensitivity of 0.80. Sensitivity analysis using alternative classifiers (logistic regression and histogram-based gradient boosting) showed consistent trends, with multimodal fusion outperforming the clinical-only baseline (Supplementary Table S5). For completeness, we additionally report performance on the basic 7-feature oculomics subset alone and on the full 34 + 15 feature multimodal configuration. The oculomics-only models (both 34-feature and 7-feature subsets) performed similarly poorly (AUC 0.58), confirming that retinal features alone are insufficient for accurate CVD prediction. This performance should be interpreted in the context of study design choices. Unlike many retina-based CVD studies that use end-to-end deep models on the full fundus image (often leveraging additional non-vascular cues and very large training cohorts) [7, 8], our oculomics-only baseline is restricted to 34 interpretable artery/vein geometric biomarkers derived from segmentation masks, and targets a broad prevalent CVD label in a modest-sized curated cohort. These factors likely reduce standalone discriminative performance, motivating our primary emphasis on multimodal fusion where retinal microvascular features act as complementary risk modifiers rather than a replacement for established clinical predictors. The full multimodal model (34+15) achieved AUC 0.80 (0.77–0.82), marginally lower than the interpretable fusion (0.81), indicating that the compact 7-feature subset captures the clinically relevant retinal signal without overfitting. As shown in Fig. 3, the ROC curves illustrate the superior discriminative performance of the multimodal fusion model relative to the unimodal baselines. A statistically significant but modest performance increase of 2.5% in AUC is observed upon fusion (0.79 → 0.81). At the Youden operating point, sensitivity and specificity for the primary fusion model were 74.2% and 77.8%, respectively.Table 7Comparative performance of unimodal and multimodal AI models. results are reported for features extracted by our AVSeg-XAI segmentation model and, as an ablation, for features extracted from LUNet [21] segmentations of the same images. All experiments use 10-fold stratified cross-validation; metrics at youden operating pointModel ConfigurationFeaturesAUC (95% CI)AccuracySensitivitySpecificityAVSeg-XAI segmentation (ours)Multimodal Fusion (interpretable)Retinal (7) + Clinical (15)0.81 (0.78–0.84)76.1%74.2%77.8%Multimodal Fusion (full)Retinal (34) + Clinical (15)0.80 (0.77–0.83)74.9%65.3%84.3%Clinical Only15 Non-Invasive Vitals0.80 (0.77–0.83)76.5%69.5%83.4%Oculomics Only (full, 34)34 Retinal Features0.58 (0.54–0.62)56.5%54.1%58.8%Oculomics Only (basic, 7)7 Interpretable Retinal Features0.58 (0.54–0.61)58.4%32.9%83.6%LUNet segmentation [21] (ablation)Multimodal Fusion (interpretable)Retinal (7) + Clinical (15)0.79 (0.76–0.82)76.2%67.5%84.7%Multimodal Fusion (full)Retinal (34) + Clinical (15)0.79 (0.76–0.82)74.4%76.8%72.1%Oculomics Only (basic, 7)7 Interpretable Retinal Features0.58 (0.54–0.62)57.1%60.0%54.1%Fig. 3Receiver operating characteristic (ROC) curves demonstrating the superior performance of the multimodal fusion model (red) compared to clinical-only (blue) and oculomics-only (green) baselines Table 7Comparative performance of unimodal and multimodal AI models. results are reported for features extracted by our AVSeg-XAI segmentation model and, as an ablation, for features extracted from LUNet [21] segmentations of the same images. All experiments use 10-fold stratified cross-validation; metrics at youden operating pointModel ConfigurationFeaturesAUC (95% CI)AccuracySensitivitySpecificityAVSeg-XAI segmentation (ours)Multimodal Fusion (interpretable)Retinal (7) + Clinical (15)0.81 (0.78–0.84)76.1%74.2%77.8%Multimodal Fusion (full)Retinal (34) + Clinical (15)0.80 (0.77–0.83)74.9%65.3%84.3%Clinical Only15 Non-Invasive Vitals0.80 (0.77–0.83)76.5%69.5%83.4%Oculomics Only (full, 34)34 Retinal Features0.58 (0.54–0.62)56.5%54.1%58.8%Oculomics Only (basic, 7)7 Interpretable Retinal Features0.58 (0.54–0.61)58.4%32.9%83.6%LUNet segmentation [21] (ablation)Multimodal Fusion (interpretable)Retinal (7) + Clinical (15)0.79 (0.76–0.82)76.2%67.5%84.7%Multimodal Fusion (full)Retinal (34) + Clinical (15)0.79 (0.76–0.82)74.4%76.8%72.1%Oculomics Only (basic, 7)7 Interpretable Retinal Features0.58 (0.54–0.62)57.1%60.0%54.1% Comparative performance of unimodal and multimodal AI models. results are reported for features extracted by our AVSeg-XAI segmentation model and, as an ablation, for features extracted from LUNet [21] segmentations of the same images. All experiments use 10-fold stratified cross-validation; metrics at youden operating point Receiver operating characteristic (ROC) curves demonstrating the superior performance of the multimodal fusion model (red) compared to clinical-only (blue) and oculomics-only (green) baselines Stratified analysis: uncovering the “hidden cohorts” We also performed stratified analysis for the fusion model on gender, age and BMI based subgroups. Results on stratified analaysis is highlighted in Table 8. Figure 4 shows a visual summary of results, highlighting the highest performance in the 45–55-year age group and females. As this strata of age group is relatively small (e.g., age 45–55, N = 147), results from this should be interpreted carefully.Table 8Stratified model performance for the multimodal fusion model. values in parentheses are 95% confidence intervalsStratification GroupSubgroupNAUC (95% CI)Accuracy (95% CI)InterpretationGenderFemale4360.83 (0.79–0.87)79.6% (75.6–83.1%)Stronger vascular signal in womenMale4750.79 (0.75–0.83)72.4% (68.2–76.2%)—Age < 45years7640.77 (0.73–0.80)73.6% (70.4–76.6%)Early disease stage45–55 years1470.90 (0.83–0.97)87.8% (81.5–92.1%)Optimal screening windowBMINormal weight1350.75 (0.65–0.84)79.3% (71.7–85.3%)—Overweight (25–30)3550.77 (0.71–0.81)75.2% (70.5–79.4%)—Obese ( > 30)4210.82 (0.78–0.86)75.3% (71.0–79.2%)Metabolic syndrome signatureFig. 4Stratified performance analysis revealing the “mid-life peak” in the 45–55 age group and the performance advantage in female patients Table 8Stratified model performance for the multimodal fusion model. values in parentheses are 95% confidence intervalsStratification GroupSubgroupNAUC (95% CI)Accuracy (95% CI)InterpretationGenderFemale4360.83 (0.79–0.87)79.6% (75.6–83.1%)Stronger vascular signal in womenMale4750.79 (0.75–0.83)72.4% (68.2–76.2%)—Age < 45years7640.77 (0.73–0.80)73.6% (70.4–76.6%)Early disease stage45–55 years1470.90 (0.83–0.97)87.8% (81.5–92.1%)Optimal screening windowBMINormal weight1350.75 (0.65–0.84)79.3% (71.7–85.3%)—Overweight (25–30)3550.77 (0.71–0.81)75.2% (70.5–79.4%)—Obese ( > 30)4210.82 (0.78–0.86)75.3% (71.0–79.2%)Metabolic syndrome signature Stratified model performance for the multimodal fusion model. values in parentheses are 95% confidence intervals Stratified performance analysis revealing the “mid-life peak” in the 45–55 age group and the performance advantage in female patients The “mid-life peak”: a critical window for AI screening The multimodal fusion model achieved the best performance in the 45–55 age bracket (AUC 0.90, accuracy 87.8%). This high level of accuracy showcases a critical time for oculomic screening. During the midlife transition period corresponding to the emergence of subclinical atherosclerosis and, in women, the menopausal transition, retinal microvascular damage aligns more closely with systemic risk. The disease signal in younger patients ( < 45) remains too subtle (AUC 0.77), whereas in older patients, age-related microvascular changes introduce confounding factors that dilute the signal. The performance for this age-specific cohort suggests that AI-driven oculomics is optimally deployed as a “gatekeeper” test in the midlife cohort, identifying individuals at high risk for preventive interventions before catastrophic events occur. Gender-specific AI performance: bridging the gender gap A higher performance was seen in females (AUC 0.83) as compared to males (AUC 0.79, p < 0.05). This performance difference is clinically significant and consistent with established pathophysiology: women’s cardiovascular disease often presents as coronary microvascular dysfunction instead of the obstructive coronary artery disease that is common in males. Conventional risk assessments, such as Framingham, frequently perform poorly for women and are significantly biased toward male-pattern risk variables. Our findings suggest that the retina, a microvascular bed in its own right, may be a particularly sensitive indicator of female-pattern vascular aging. This presents AI-driven oculomics as a precision medicine approach to close the gender gap in cardiovascular diagnoses, rather than merely a generic screening tool. Virtual phenotyping: supporting biological plausibility of AI-Learned features Distinguishing genuine physiological signals from spurious correlations or shortcut learning is a critical challenge in medical AI. To provide supporting evidence that our retinal features encode real systemic vascular information, we performed “virtual phenotyping” by training separate models to predict systemic biomarkers using only retinal features as input. Blood pressure: the hypertensive retinopathy signal Retinal features alone predicted systolic blood pressure with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2=0.20$$\end{document} (RMSE: 13.61 mmHg) and diastolic blood pressure with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2=0.14$$\end{document} (RMSE: 10.46 mmHg). While modest, these correlations were statistically significant (p < 0.001). Moreover, feature importance analysis indicated that arteriolar fractal dimension (lower complexity associated with higher BP) and arteriolar tortuosity (higher curvature associated with higher BP) were the important features for this prediction, which is consistent with known hypertensive retinopathy biomarkers. However, an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} of 0.20 explains only 20% of the variance in SBP, which is insufficient for clinically useful blood pressure prediction. This analysis is presented solely as exploratory evidence of biological plausibility—the retinal–BP association is compatible with known hypertensive remodeling—and should not be interpreted as clinical validation of any feature or the pipeline. As shown in Fig. 5, predicted versus observed systolic blood pressure demonstrates a statistically meaningful association based solely on retinal vascular features.Fig. 5Virtual phenotyping results showing the relationship between measured systolic blood pressure and values predicted solely from retinal vascular features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(R^2=0.20)$$\end{document} Virtual phenotyping results showing the relationship between measured systolic blood pressure and values predicted solely from retinal vascular features \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(R^2=0.20)$$\end{document} Smoking status: detecting “smoker’s vasculopathy” The model distinguished smokers from non-smokers based solely on retinal vasculature with an AUC of 0.69 (95% CI: 0.64–0.74), demonstrating a detectable “smoker’s vasculopathy” signature. Current smokers exhibited increased venular tortuosity and reduced arteriolar fractal dimension compared to never-smokers, consistent with oxidative stress-induced endothelial dysfunction. Metabolic factors: confirming feature orthogonality Retinal features showed poor predictive power for BMI (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2 \approx 0.02$$\end{document}) and age (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2 \approx 0.03$$\end{document}), despite these being strong clinical predictors of CVD. These negative results indicate that the fusion model works because retinal and clinical inputs are orthogonal (non-redundant). Clinical variables, such as obesity and age, provide a metabolic risk context, whereas the retina offers a direct readout of vascular consequences (e.g., hypertension and endothelial stress). This relationship explains why multimodal fusion outperforms either modality in isolation. Explainable AI: understanding the model’s decision logic SHAP analysis was employed to characterize the model’s decision-making and to enable clinical trust and actionable insights. Global feature importance SHAP analysis was conducted to understand the importance of features (Fig. 6). We observed Systolic Blood Pressure (mean |SHAP:| 0.076) was the most dominant feature, exceeding Age (0.047) and BMI (0.037), which indicates hypertension as the primary dominant factor for this CVD cohort. Among oculomics biomarkers, CRAE (0.014) was the most important global feature. Notably, the beeswarm plot shows that lower arteriolar caliber (blue dots on the right side of the axis) is strongly associated with higher CVD risk, confirming the generalized narrowing hypothesis. Moreover, the Vein Tortuosity Index (0.012) ranked as the top topological feature, indicating that venous hemodynamic stress substantiates distinct prognostic information complementary to standard caliber measurements.Fig. 6SHAP summary plot illustrating global feature importance. Each dot represents a patient; red indicates a high feature value and blue indicates a low feature value. Features are ranked by their impact on model output SHAP summary plot illustrating global feature importance. Each dot represents a patient; red indicates a high feature value and blue indicates a low feature value. Features are ranked by their impact on model output Retinal features as “risk modifiers” SHAP analysis showed retinal vascular structure acted as a critical risk modifier for established clinical risk factors. More specifically, the analysis suggests that in the presence of the same blood pressure, patients may have different probabilities of CVD risk due to variations in microvascular density. For patients with high systolic blood pressure, those with vascular networks characterized by a low arteriolar intersection count and, therefore, sparse networks, were assigned significantly higher risk scores than those with denser and more complex networks. This illustrates the clinical value proposition: oculomics provides the detailed risk stratification necessary to identify high-risk phenotypes within the broad clinical categories, as shown in Fig. 7.Fig. 7SHAP dependence plot for systolic blood pressure. The vertical dispersion at high blood pressure levels demonstrates an interaction effect: color indicates Arteriolar number of Intersections, showing that patients with vascular rarefaction (blue dots, low intersections) carry a higher predicted risk than those with preserved vascular complexity (red dots) for the same blood pressure SHAP dependence plot for systolic blood pressure. The vertical dispersion at high blood pressure levels demonstrates an interaction effect: color indicates Arteriolar number of Intersections, showing that patients with vascular rarefaction (blue dots, low intersections) carry a higher predicted risk than those with preserved vascular complexity (red dots) for the same blood pressure Discussion Principal findings In this article we proposed AVSeg-XAI a explainable multimodal AI framework which fuses retinal vascular biomarkers and non-invasive clinical vitals for CVD risk stratification. AVSeg-XAI leverages segmentation (mean Dice: 0.8368) for A/V segmentation of retina and then applied geometric feature extraction from vascular network. For the CVD prediction model, we achieved AUC of 0.58 using oculomics-only features. For A/V segmentation we used UZLF dataset of only 156 images for training. As this is relatively a smaller number compared to the amount of data required for deep learning models, this smaller size may impact the quality of A/V segmentation ultimately affecting the capability of 34 oculomics-only features to distinguish CVD patients from normal QBB participants. Moreover, we worked on CVD patients, and the manifestation of CVD over retinal changes may take some time to progress. That’s why in the final proposed model, we incorporated additional 15 clinical features along with oculomics-only features to enhance the performance of our prediction model. The final model achieved AUC of 0.81 which shows that signal from the retina vascular network and vital signs jointly can contribute in the CVD risk stratification. Non-invasive vitals provide a representation of metabolic manifestation (i.e., factors driven by obesity and age), while retinal vasculature provide evidence for vascular changes (i.e., hypertensive vascular remodeling and endothelial dysfunction). Their combination exhibits greater predictive power for the AI model at a statistically significant level. Gender-specific performance The AI model performed better for women (AUC 0.83) than for men (AUC 0.79). This gender-specific advantage addresses a critical gap in CVD medicine where women with CVD are frequently present with coronary microvascular dysfunction rather than obstructive coronary artery disease, which is the phenotype for which traditional risk scores were optimized [3]. However, coronary microvascular dysfunction is not directly measured in QBB, so this pathophysiological explanation should be interpreted as hypothesis-generating rather than mechanistically established in the current study. Traditional CVD risk calculators such as Framingham and SCORE2 were developed in specific cohorts and may not generalize uniformly across sex-specific phenotypes or populations [30]. Our findings should therefore be interpreted as evidence of subgroup-dependent performance within this dataset, rather than as a direct comparative evaluation of existing risk scores. The retina, as an alternative, appears particularly sensitive to female-pattern vascular aging that precedes over CVD events. Our virtual phenotyping experiments support this by demonstrating that retinal features encode systemic blood pressure (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2 = 0.20$$\end{document}), a key driver of microvascular remodeling. These findings suggest that AI-driven oculomics may provide complementary information for sex-stratified cardiovascular risk assessment; however, this interpretation should be considered hypothesis-generating and requires external and prospective validation. The “Mid-Life window”: optimal timing for AI screening The best performance of the proposed model was observed in the middle aged 45–55 group (AUC 0.90, 95% CI 0.83–0.97; accuracy 87.8%, 95% CI 81.5–92.1%) determines a critical window of opportunity for preventive intervention. However, this subgroup contains a limited number of participants (N = 147), so the observed peak should be interpreted cautiously and validated in larger cohorts. This age range corresponds to the onset of subclinical atherosclerosis and, in women, the menopausal transition that accelerates vascular aging. Menopausal status and reproductive hormone measures were not available in the analyzed cohort records; therefore, this interpretation is based on age-range association rather than direct measurement which needs to be validated with cohort having sufficient relevant clinical information. In younger patients ( < 45 years), cardiovascular pathology remains too subtle for reliable detection (AUC 0.77, 95% CI 0.73–0.80), whereas in older patients, age-related microvascular changes introduce confounding factors that may dilute the signal. The clinical implications of this age-specific performance profile are significant. Rather than employing universal screening across all ages, midlife adults may represent the most appropriate target age cohort for AI-driven oculomics being used optimally as a targeted “gatekeeper” test. Individuals identified as high-risk during this window can receive aggressive preventive interventions (e.g., lifestyle modification and pharmacotherapy) decades before catastrophic events occur, maximizing the potential for disease modification. It is important to interpret the peak performance in the 45–55 age subgroup (AUC 0.90) with appropriate caution. The bootstrap 95% CI for this stratum is 0.83–0.97, reflecting a wide confidence interval (range 0.14) due to the small effective sample size (n=147). In particular, there is a scarcity of controls in this subgroup (122 cases vs. 25 controls), resulting in a local CVD prevalence of 83% that is substantially higher than both the overall modeled cohort (50%) and any realistic community screening population. Consequently, the apparent discriminative advantage may partially reflect the specific demographic composition and prevalence of this single-cohort subgroup rather than solely a genuine interaction between age and model performance. This subgroup finding should be considered strictly exploratory; it requires prospective replication in an independent, prevalence-representative cohort before it can be relied upon as a definitive clinical screening window. Biological plausibility supported by virtual phenotyping In this work, we use virtual phenotyping as a supportive analysis to assess whether the extracted retinal vascular features capture physiologically meaningful information, rather than as a validation of clinical prediction performance or causal mechanisms. We observe that retinal features alone can predict systolic blood pressure (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2=0.20$$\end{document}) and smoking status (AUC 0.69), which is consistent with prior evidence linking retinal microvascular changes to hypertension- and smoking-related vascular injury [5, 6]. These results are also in line with recent reports suggesting that retinal imaging can reflect systemic cardiometabolic status [8]. In addition, our univariate analysis indicates that several advanced geometric/topological features (e.g., Artery Intersections, D0 Capacity) are lower in participants with prevalent CVD (p < 0.001). We interpret these associations cautiously: they are compatible with a pattern of microvascular remodeling and possible rarefaction-related changes, but they do not establish mechanism, temporality, or specificity, and may be influenced by cohort characteristics and measurement noise. Overall, the virtual phenotyping and univariate findings provide internal consistency supporting biological plausibility of the derived features. However, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2=0.20$$\end{document} for systolic BP prediction is modest and insufficient for clinically useful blood pressure prediction; the virtual phenotyping analyses should not be interpreted as establishing clinical utility of any individual feature or the pipeline. External validation and longitudinal studies are required to confirm clinical utility and mechanistic interpretation. The negative result is equally important in CVD, where BMI is a strong risk factor for CVD, and it is poorly predicted by retinal features (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} approximately 0.02). This confirms the orthogonality between the features, as both retina and clinical inputs measure different dimensions of cardiovascular risk. Pathways of biological interest, such as hyperlipidemia and post-revascularization recovery mechanisms [31], help contextualize the cardiovascular risk features of the learned AI features. This orthogonality supports the interpretation that multimodal fusion provides complementary information, yielding a modest improvement over either modality alone. Clinical translation and implementation The current assessment of CVD risk requires people to visit clinics for blood extraction, which is a barrier to accessibility. Moreover, the phlebotomy tests are expensive. Our model, however, involves fundus photographs (which can be taken in optometry settings) and basic vitals (which can be obtained in pharmacies or through home monitoring devices). Due to its non-invasive nature, AVSeg-XAI has potential to be adopted for community-based CVD screening. While the observed AUC gain is modest, clinical utility is determined by operating thresholds. As an illustrative example, at a fixed sensitivity of 0.80 the multimodal model yields higher specificity than the clinical-only baseline in this cohort (Supplementary Table S6), which could reduce false-positive referrals at a screening-oriented sensitivity. Moreover low-cost of portable fundus cameras and smartphone based AI model integration can be considered as a viable option for retinal imaging and screening for diabetes and cardiovascular disease screening at the community which lacks proper clinical resource [32]. However, population-based validation and calibration in representative prevalence settings are required before translating these results into screening thresholds. Using Random Forest classifiers in comparison against complex deep neural networks is also beneficial due to their lower computational demands, which allow deployment on more basic hardware, faster prediction times, and built-in measurement of feature importance, which can be useful for clinical interpretation (including SHAP). To ensure these findings were classifier-independent, a DeLong test on the 911 matched out-of-fold predictions confirmed that Random Forest and Histogram-based Gradient Boosting (HistGB) yielded statistically indistinguishable performance for the primary multimodal model (p=0.886, 95% CI [-0.019, +0.016]), indicating that the classifier choice has no material impact on the conclusions. Additionally, our evaluation revealed that the full 34 + 15 feature set underperformed the parsimonious 7 + 15 interpretable subset when using Random Forest (RF DeLong p=0.003, significant), yet this degradation vanished when using HistGB (p=0.60, not significant). This classifier-specific phenomenon arises because Random Forest samples a fixed number of candidate features at each split node; the high proportion of mathematically correlated or uninformative geometric features in the full 34-feature set dilutes the candidate pool, preventing trees from consistently selecting the most discriminative physiological signals. HistGB, conversely, builds trees sequentially on residuals and naturally suppresses uninformative features. Rather than undermining the 34-feature extraction pipeline—which remains valuable for exploratory biomarker discovery—this mechanism validates the selection of the compact 7-feature interpretable subset as the optimal configuration for predictive modeling. Segmentation module performance and role in the pipeline The proposed segmentation model achieved a mean Dice coefficient of 0.8368 on UZLF, which is slightly higher than the reported LUNet benchmark (0.8327). Given the small numerical difference, this result should be viewed cautiously. Within the scope of this work, the main utility of the segmentation module is to support artery/vein-aware retinal biomarker extraction for the downstream interpretable oculomics framework. Limitations and future directions This study has several limitations that should be considered when interpreting the results. First, the retrospective case–control design precludes causal inference and does not establish whether retinal vascular changes predict future incident CVD events. Standardized ophthalmic grading labels (e.g., hypertensive retinopathy or diabetic retinopathy severity grades) were not available in the QBB records used here, preventing direct validation of individual retinal features against expert ophthalmic grading in this study. Second, the cohort was intentionally balanced for case/control representation during model development; therefore, discrimination metrics from this curated dataset may not directly translate to real-world screening settings with different CVD prevalence, calibration properties, or clinical decision thresholds. Third, the cohort was drawn from a single biobank (Qatar Biobank) and a single population context, so external validation across imaging devices, acquisition protocols, and ethnic groups is required. In addition, the A/V segmentation module was quantitatively evaluated only on the UZLF benchmark, and because QBB does not provide manual A/V labels, segmentation generalizability to the target cohort cannot be directly assessed in the current study. Additionally, some advanced geometric/topological biomarkers are sensitive to segmentation errors and image resolution, and future work will quantify feature robustness under controlled perturbations and across multi-device labeled datasets. Fourth, subgroup analyses (particularly smaller strata, such as the 45–55 age subgroup [AUC: 0.90, 95% CI: 0.83–0.97]) may be unstable and susceptible to optimistic estimates. Given the prevalence discrepancy in this stratum (83% cases versus a community screening context), this result must be validated in an independent prevalence-representative cohort before being relied upon as a key finding. Moreover, given the limited cohort size and the use of non-linear models with higher-dimensional feature sets, some degree of overfitting and optimistic cross-validated performance estimates cannot be excluded. Fifth, while the pipeline is more interpretable than end-to-end image-only models, residual confounding and unmeasured factors may influence both retinal features and CVD status, examples include differences in medication use, comorbidity burden, and unmeasured behavioral or socioeconomic factors that may affect both retinal microvasculature and CVD status, and SHAP-based explanations do not imply causality. Additionally, proposed biological mechanisms (e.g., menopausal transition, coronary microvascular dysfunction, or endothelial dysfunction) are not directly measured in this dataset, and require prospective studies with adjudicated phenotyping and longitudinal event follow-up for confirmation. Selection bias may arise from biobank participation which allows only Qatari national or long term residents in Qatar over 15 years, availability of retinal imaging/clinical metadata, and automated image-quality exclusions. Because the study uses prevalent CVD labels rather than incident longitudinal outcomes, survivorship/prevalence bias is also possible. These factors may limit generalizability and may contribute to uncertainty in subgroup estimates. Accordingly, the label does not provide consistent adjudicated subtype granularity (e.g., specific CVD subtypes) or event timing for all participants. Sixth, although Benjamini–Hochberg FDR correction was applied to the primary comparison families (retinal biomarker univariate tests, segmentation Dice comparisons, and subgroup AUC comparisons), the manuscript reports results across multiple tables and analysis dimensions; a comprehensive family-wide correction across all reported statistics was not performed and residual inflation of the family-wise error rate cannot be excluded. Seventh, the case–control design employed strict phenotypic eligibility criteria: cases were required to present with both a major CVD event and a hypertension-associated state while being free from cancer, obesity, and diabetes; controls were required to have no personal CVD history and no current antihypertensive or antidiabetic medications. While these criteria sharpen the phenotypic contrast between groups and reduce confounding from overlapping comorbidities, they produce a highly selected comparison that does not reflect the full clinical spectrum encountered in routine cardiovascular screening. In practice, a substantial proportion of CVD patients present concurrently with diabetes, obesity, or treated hypertension; excluding these groups means discriminative performance may be optimistic relative to mixed-comorbidity populations. Conversely, restricting controls to medication-naïve individuals removes the large fraction of the general population on antihypertensive therapy who remain event-free, potentially artificially widening the retinal phenotypic contrast between groups. These forms of selection bias represent a constraint on external validity, and independent validation in cohorts inclusive of common comorbid presentations is required before reported performance estimates can be generalised to routine screening populations. Future work should focus on prospective validation, multi-center external validation, and integration with existing clinical workflows. Extending the multimodal framework to incorporate additional non-invasive modalities (e.g., electrocardiography and carotid ultrasound) may further enhance predictive performance. Conclusion We present an interpretable multimodal oculomics framework for discriminating prevalent CVD status in a curated Qatar Biobank case–control cohort. By combining artery/vein-aware retinal vascular features with non-invasive clinical vitals, the proposed model achieves an overall AUC of 0.81 in cross-validation; stratified results (e.g., higher AUC in women and in the 45–55 subgroup) are exploratory and should be interpreted cautiously given subgroup sample sizes. Importantly, this study is intended as a proof-of-concept demonstration of feasibility rather than a deployment-ready screening tool. Prospective validation in prevalence-representative cohorts, external multi-center evaluation across imaging devices and populations, and calibration for clinical operating thresholds are required before translation into routine screening workflows. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1