Recent advances in AI have significantly improved CVD risk prediction using CFP. Evidence indicates that CFP captures biomarkers predictive of future CVD risk, with DL-derived risk scores independently associated with CVD events (43–45). Retinal features—including ischemic perivascular lesions, subretinal drusen-like deposits, and microvascular metrics such as vessel density, caliber, tortuosity, and fractal dimension—have been established as clinically useful indicators for systemic disease assessment (46, 47). For example, a U-Net-based DL model demonstrated a significant association between retinal microvascular density and fractal dimensions with congestive heart failure, reinforcing the retina's role as a window into cardiovascular health (47).

Several studies have developed specific DL models for CVD prediction. A model based on the Inception-v3 architecture predicted high coronary artery calcium scores (CACS >100) from CFP using vascular and macular features, achieving an area under the receiver operating characteristic curve (AUROC) of 83.2%—outperforming single clinical parameters and comparable to age-based indicators (28). Another model developed to estimate 10-year ischemic cardiovascular disease (ICVD) risk showed strong performance (AUC 0.85–0.97) and was trained on a large dataset of nearly 400,000 participants, enhancing its generalizability (29). Wang et al. introduced a DL-based retinal aging score (Reti-aging score) that efficiently predicts vascular aging and CVD risk from CFP, with an AUC of 0.779–0.826, outperforming both single clinical parameters (e.g., age, hypertension) and specialist physicians (30). The same model also predicted new-onset hypertension and carotid artery plaque with AUC values of 0.703 and 0.705, respectively (30).

Several advanced AI models have been developed to predict hypertension-related CVD risk using retinal imaging. Srilakshmi et al. introduced a deep neuro-fuzzy network combined with a fractional calculus-based optimization algorithm, achieving an accuracy of 91.6%, sensitivity of 92.3%, and specificity of 91.9%, outperforming conventional models (31). Similarly, the Singapore I Vessel Assessment-Deep Learning System (SIVA-DLS) automates retinal vessel diameter measurement, showing high agreement with expert graders (intraclass correlation coefficient: 0.82–0.95) (26). Other approaches include a hybrid computer vision and DL model with 87.5% accuracy (32), a convolutional neural network that integrates multi-feature fusion to capture microvascular abnormalities (33), and an osprey gannet optimization-based model reaching 92.1% accuracy (33). Additional high-performing frameworks include a fractional chef-based optimization algorithm combined with SpinalNet (accuracy: 91.3%) (34), a Dense Spiking Forward Fractional Network exceeding 90% accuracy in multi-center datasets (35), and a hybrid Inception V3-VGG16 model achieving 99.5% accuracy by leveraging both global and fine-grained retinal features (36). The latter also offers rapid processing, making it suitable for elderly screening in resource-limited settings (36).

In summary, most CFP-based AI models demonstrate high predictive accuracy (>90%), often with screening times under one minute. Integration of retinal microvascular features with clinical data further improves performance. These non-invasive, cost-effective tools show particular promise for early CVD risk screening in underserved populations. Nevertheless, many models rely on single-center, limited-sample datasets, constraining generalizability, and further validation in diverse cohorts is needed before broad clinical adoption.

CVD risk stratification is crucial for the prevention and management of CVD (48). AI models using CFP can not only predict CVD risk, but also stratify CVD risk. A study found that only 25.8% of participants had undergone CVD risk assessment (37). This method not only identifies high-risk and borderline-risk groups but also facilitates more precise early intervention, thereby improving patient survival rates. Retinal vessels, due to their anatomical specificity, have emerged as a potential tool for CVD risk stratification (48). Consequently, an increasing number of research and development efforts are focused on developing AI models for CVD risk stratification using CFP. Research has found that AI models based on CFP [Residual Neural Network 18 (ResNet 18)/Multiple Instance Learning] can be used for screening and grading diabetic cardiac autonomic neuropathy (49). Several AI models demonstrate the potential of CFP for CVD risk stratification. One rpCVD model achieved an AUC of 0.672 in predicting 10-year CVD risk, performing comparably to the World Health Organization (WHO) risk score (AUC = 0.693) (37). Risk classifications (low/moderate/high) were consistent with the WHO score in 63.4% of participants, and both patients (92.5%) and general practitioners (87.5%) reported high satisfaction with the tool (37). This model can be incorporated into primary care workflows to help address the limited coverage of conventional risk assessment methods (37). An ordered regression DL model has also been developed to predict and stratify 10-year CVD risk based on CFP (38). By integrating significance analysis and imaging decomposition techniques—such as vessel or macula removal experiments—the model provides interpretable predictions by identifying key retinal features influencing risk stratification (38). The Retinal Imaging-Based CVD (Reti-CVD) tool identified a 10-year CVD risk of 13.1% in its high-risk group (40). Compared to QRISK3, the current UK standard, Reti-CVD improved stratification of borderline-risk individuals and detected high-risk patients missed by QRISK3 (e.g., 8.6% of the high-risk group had a QRISK3 score of only 0%–5%) (40). These findings support its use as a non-invasive aid for clinical decision-making, such as initiating statin therapy (40).

In summary, AI-enhanced retinal imaging allows more precise and comprehensive CVD risk stratification, which may facilitate personalized early intervention and slow disease progression. However, few CFP-based stratification models currently exist, most are derived from single-center studies, and clinical practicality remains limited.

In addition to risk prediction, AI has also demonstrated diagnostic potential for specific CVD subtypes. The DL-funduscopic atherosclerosis score (DL-FAS) model enhances the predictive capability of cardiovascular mortality risk based on the Framingham Risk Score (FRS), demonstrating significant value particularly in risk stratification among individuals with moderate risk (FRS 10%–19%) (41). This model holds promise as an adjunct tool for CVD risk stratification (41). Additionally, by quantifying atherosclerotic features in carotid angiography, the model achieved an AUROC of 0.713 for predicting carotid atherosclerosis, along with an accuracy of 0.583 and a sensitivity of 0.891 (41). A multicenter, large-scale study found that retinal imaging of coronary artery calcium (RetiCAC) is comparable to computed tomography-measured coronary artery calcium (CAC) in terms of risk stratification performance (39). Furthermore, it significantly improves the predictive ability of the pooled cohort equation (PCE) for intermediate-risk and borderline-risk groups, serving as a complementary tool to PCE to optimize decision-making for intermediate-risk groups (e.g., 7.5%–20% 10-year risk) (39). Besides, this model predicts CAC achieved an AUC of 0.742, outperforming single clinical parameters such as age and blood glucose levels (39). This aids in the early assisted diagnosis and management of CAC. Furthermore, an attention-based multi-instance learning model has been validated for its feasibility in the early assisted diagnosis of PAD (42). By leveraging microvascular changes in the optic disc and temporal vascular arcade, this model offers a novel perspective for the systematic assessment of atherosclerosis (42). Currently, there is limited literature on the assisted diagnosis of CVD using CFP, and it lacks clinical practicality. In the future, more attempts can be made to develop models for diagnosing CVD by combining clinical data, population information, and other factors.

OCT is a high-resolution imaging technique capable of capturing the layered structure of the retina. Studies have shown that changes in retinal thickness and structure are associated with CVD. The risk monitoring and healthcare assistance system model identified certain retinal fundus features in OCT retinal imaging that were significantly associated with the development of aortic aneurysms and adverse aortic events (57). This suggests that OCT may be a promising tool for the early detection and intervention of aortic aneurysms and adverse aortic events (57). A model based on interactive ML technology has, for the first time, enabled automated detection of reticular intraretinal perivascular lesions (RIPLs) and subretinal drusenoid deposits (SDDs) using a small-scale OCT dataset (58). Its efficiency (6-hour training) and high specificity (>90%) provide a new tool for studying retinal biomarkers of CVD (58). OCTA is a high-resolution vascular imaging technique. By enhancing the visualization of deep capillary networks and the choroid, it facilitates the identification of different types of retinal capillary networks and provides detailed imaging of retinal microvascular structures (59–63). OCTA can detect early changes in retinal microvasculature, which may be early biomarkers of CVD (50, 51, 53, 64). A prospective cross-sectional study found that the density of the superficial retinal capillary plexus measured by OCT-A correlates with CVD risk profiles and impaired left ventricular ejection fraction in patients with high-risk CVD status (65). Therefore, quantitative retinal microvascular data can be considered a valid surrogate for CVD risk profiles and may enhance CVD risk assessment. This represents the first evidence linking OCT-A metrics to CVD risk (65). Studies have shown a significant correlation between retinal capillary density and markers of adverse cardiac remodeling, providing a new perspective for predicting the risk of systemic diseases such as CVD (66). A prospective observational cohort study demonstrated that preoperative retinal hypoperfusion independently predicted an elevated risk of perioperative adverse cardiovascular events in patients with coronary heart disease, highlighting retinal microcirculation assessment as a non-invasive biomarker to inform cardiac surgical prognostication and guide personalized therapeutic strategies (67). Supervised ML models combining CFP and OCTA data demonstrated high accuracy (up to 96.53%) in predicting CVD risk (51). The retinal foundation model (RETFound) model achieved AUROC values of 0.737 and 0.794 when using CFP and OCT imaging to predict MI and heart failure, respectively (52).

The AUC of the CNN model for predicting congestive heart failure risk under 3 × 3 mm high-resolution OCTA imaging scanning was 0.61 (53). The scanning performance significantly decreased under 6 × 6 mm resolution OCTA images and 8 × 8 mm resolution OCTA images (AUC ≤ 0.5) (53). Based on retinal OCT-A patient images from the open-source RASTA dataset, a study evaluated the accuracy of ML and DL algorithms in predicting CHA₂DS₂-VASC neurocardiovascular risk scores (50). The findings revealed that EfficientNetV2-B3 is a suitable DL model for retinal OCT-A imaging, correctly predicting risk in 68% of cases (50). This contributes to the assessment of future neurocardiovascular characteristics (50). A study combining traditional cardiovascular risk factors employed multiple ML algorithms (such as k-nearest neighbors, naive bayes classifier, support vector machines, etc.) to predict the risk of coronary artery disease. The classification accuracy of the ML algorithms (up to 86%) was significantly higher than that of traditional logistic regression models (78.7%) (54). As a non-invasive diagnostic method, OCTA holds promise as a new biomarker for CVD (54).

These findings underscore the potential of OCT and OCTA as powerful, non-invasive tools for detecting early retinal microstructural and microvascular changes linked to CVD, offering insights beyond traditional risk assessment methods. The integration of AI-driven analysis with high-resolution retinal imaging enhances predictive accuracy, enabling more precise CVD risk stratification and early intervention strategies. Future research should focus on standardizing imaging protocols and expanding multicenter datasets to further validate these innovative biomarkers for clinical adoption.

OCTA serves as a critical tool for assisted diagnosis, progression tracking, and guiding therapeutic decision-making for a wide range of systemic health conditions. Retinal OCTA-derived biomarkers enable non-invasive quantification of microvascular pathology, offering actionable insights to refine clinical decision-making (68). At the same time, Ultra Wide-Field OCTA has been shown to provide reliable performance for detecting neovascularization and intraretinal microvascular abnormalities achieving similar accuracy to fluorescein angiography (69). Traditional assessment methods rely on clinical indicators and patient history, but may suffer from a lack of accuracy and timeliness, especially in areas with relatively scarce resources. Therefore, it is important to explore new, non-invasive assessment methods. Studies have shown that a reduction in the thickness of inner segment/outer segment junction of the retinal pigment epithelium and inner nuclear layer to external limiting membrane is significantly associated with heart failure (55). This study analyzed retinal OCT features using ML models (XGBoost, etc.) and achieved the first non-invasive classification of heart failure subtypes with the best performance (AUC 0.70) (55).

It has been shown that increased CVD risk is inversely related to leukocyte telomere length, whether assessed by traditional biomarkers, CVD risk scores, or our DL heart biological age (BioAge) CVD risk model (56). Shortened leukocyte telomere length serves as an alternative biomarker for increased CVD risk. This model reliably captures this CVD risk biomarker and accurately distinguishes between high and low risk individuals, consistent with known CVD risk trends. Therefore, it can facilitate rapid and accurate screening for CVD risk (56). The ASI-Net model study demonstrates the applicability of AI-enhanced OCTA imaging analysis in the detection of ischemic stroke and its subtype classification (68). Collectively, these advances position OCTA and AI-enhanced retinal analysis as transformative approaches for systemic disease management, bridging critical gaps in early detection, risk stratification, and personalized therapeutic planning across cardiovascular conditions.

As research on AI models predicting CVD through retinal imaging continues to deepen, more and more models are combining clinical data, genomic data, metabolomic data, and other information to comprehensively predict CVD risk and improve model efficiency (89, 90). The study used DL models and metabolomics technology to reveal the molecular link between retinal aging and CVD, establishing a novel biomarker called the metabolomic signature of retinal aging (MSRA) (70). It also verified that MSRA has statistical significance in predicting CVD (p < 0.05) (70). A multimodal DL model combining CNN and deep neural network (DNN) significantly improved the predictive performance of CVD by combining CFP and traditional risk factors (such as age, blood pressure, cholesterol, etc.) (71). The model performed well in both internal and external validation and was able to identify high-risk patients for future CVD events (71). The efficientnet-base model variant 3 (EfficientNet-B3) multimodal model achieves high-precision prediction of CVD risk (AUC-ROC 96.3%) by integrating CFP with clinical data, significantly outperforming traditional models such as residual neural network with 50 layers (ResNet-50) and VGG16 (72). Its core advantage lies in combining clinical data to overcome the limitations of single data sources, thereby enhancing the comprehensiveness of predictions (72). On the other hand, it uses gradient-weighted class activation mapping (Grad-CAM) to generate heatmaps, visualizing the model's focus on critical retinal regions (such as areas with vascular abnormalities), thereby enhancing the model's credibility (72). A study achieved high-precision prediction of CVD risk based on CFP (AUC-ROC 90.41%) through the fusion of multimodal data with the siamese squeeze-and-excitation resnext (Siamese SE-ResNeXt) model (73). The dataset used for this model is “China-Fundus-Carotid Intima-Media Thickness (CIMT) dataset”. This dataset is an integration of retinal and carotid intima-media thickness data to provide a reference for the creation of new datasets in the future. Moreover, it is also an important resource for the development and validation of AI-based early CVD screening models using retinal imaging (73). Research has found that by combining traditional risk factors and retinal characteristics, AI models demonstrate high accuracy in predicting the risk of coronary artery disease in human immunodeficiency virus (HIV)-infected individuals (AUC close to 0.99) (74).

A study used a multimodal AI model [Electrocardiogram (ECG) + CFP] to fuse spatiotemporal features using fast fourier transform + earth mover's distance, achieving an 84% accuracy rate in predicting CVD risk, with a particular strength in identifying early microvascular lesions (75). A study combining ML and OCT technology confirmed that reduced choroidal thickness in patients with type 1 diabetes is significantly associated with carotid plaque (76). Its AUC for predicting carotid plaque was 0.82–0.85 (76). Research has found that DL models based on efficientnet-base model variant 2 (EfficientNet-B2) can efficiently predict the 10-year major adverse cardiovascular events (MACE) risk of type 2 diabetes patients through CFP, with performance comparable to traditional clinical scores (PCE) (77).

A new AI model combines retinal imaging, clinical data, and genomic data to achieve performance comparable to or slightly better than the traditional FRS in predicting the risk of MI (78). The combination of CFP and DL technology has enabled 5-year risk prediction for MACE, with an AUC of 0.7 (79). The AUC is comparable to the AUC of 0.72 for the European SCORE risk calculator (79). Multimodal AI models, combining CFP, cardiac magnetic resonance (CMR) images, and demographic data, perform as well as traditional CVD risk models (such as the Framingham score) in predicting MI risk, while being less expensive and more accessible (80). The AI-driven photoreceptor metabolic window model integrates OCT, metabolomics, and DL. This model significantly improves the prediction of MI risk and reveals the metabolic basis of the relationship between photoreceptor layer thickness and the risk of multisystem diseases (81).

In summary, currently, there are relatively more studies combining clinical data and genomics with AI models based on CFP, and their predictive performance has improved to a certain extent. However, there are fewer multi-modal AI models based on OCT and OCTA, but their performance is better than that of single-modal models. Multi-modal models combining OCT, CFP, OCTA and other clinical data have great potential in predicting CVD risk and are worth further research. This shows that in the future we can try to integrate OCT, OCTA, CFP and other prediction methods. This will not only improve the prediction rate of CVD risk, but also refine the risk stratification more. This helps with early management.

Early detection and risk stratification of CVD are crucial for prevention and treatment. The accuracy of predicting CVD risk and stratification using single imaging or assisted examinations needs to be improved, while multimodal fusion technology can significantly improve predictive accuracy. A multimodal AI model integrating retinal, genomic, and clinical data can predict the risk of MACE in patients with type 2 diabetes and effectively distinguish between high-risk and low-risk patients (82). Compared with DL models, the features selected by Lasso regression in this model have clear clinical significance, enhancing the interpretability of the model (82). By analyzing vascular parameters in CFP and combining them with genomic data, the AI model vascular mapping and perfusion imaging reconstruction (VAMPIRE) can effectively predict and stratify the 10-year MACE risk in patients with type 2 diabetes (83). The model's AUC for predicting MACE is 0.663, comparable to the traditional PCE risk score (AUC 0.658) (83). When combined with retinal parameters and a polygenic risk score, the AUC improves to 0.686, significantly outperforming the PCE risk score (83).

The AI algorithm based on DL (Reti-CVD) combines CFP, clinical data, and other biomarkers to predict 5-year CVD risk and stratify it. Its predictive accuracy is 0.751, and the risk ratio for the estimated three CVD risk groups (low, medium, and high risk) is 2.02 (84). This model is equivalent to CAC scoring, but is less costly and easier to operate (84). The SIVA-DLS model combines retinal vascular parameters and clinical data to achieve an AUC of 0.760 for predicting CVD risk (compared to an AUC of 0.720 for traditional risk factors) (85). The study also supports retinal microvascular assessment as a low-cost, non-invasive CVD risk stratification tool, particularly suitable for resource-limited areas (85). The CLAiR model uses CFP and limited demographic data to predict and stratify 10-year atherosclerotic CVD risk (86). The AUROC for predicting atherosclerotic CVD (ASCVD) risk is 0.89–0.9. Its risk stratification ability is consistent with traditional ASCVD risk assessment (86).

In summary, multimodal AI models that integrate retinal imaging, genomics, and clinical data demonstrate superior predictive performance for CVD risk stratification compared to traditional methods like PCE risk score, while also offering cost-efficiency and scalability. These advancements highlight the potential of AI-driven, interpretable risk assessment tools to enhance early CVD detection and personalized prevention strategies in high-risk populations, such as type 2 diabetes patients.

The AI model combining VGG16 and transfer learning integrates CFP and clinical data to determine CAC in diabetic patients and stratify CAC risk (with an accuracy rate of 72%) (18). It also innovatively combines clinical data to optimize predictions (with an accuracy rate of 91%) (18). The M2AI-CVD system combines CFP with clinical data and improves model performance through entropy-optimized segmentation and genetic algorithm feature selection (87). It achieves high-precision (95.89%) CVD risk prediction (87). This provides a promising solution for the accurate and early detection of CVD (87). This study is the first to propose a multimodal DL model combining CFP and dual-energy x-ray absorptiometry (DXA) data for non-invasive CVD assisted diagnosis, with an accuracy rate of 78.3% (88). Despite data limitations and generalization challenges, its non-invasive, efficient, and interpretable nature offers new insights for early CVD screening (88). In the future, by expanding the dataset and conducting clinical validation, it is expected to become a powerful tool for rapid CVD screening in primary care.

Existing models are primarily developed for specific populations (e.g., diabetic patients), and their generalization ability across populations and races has not been sufficiently validated. Moreover, the diversity and quality of datasets vary significantly, affecting the stability of model performance.

On the one hand, microvascular changes observed through retinal imaging may lack specificity for certain diseases. On the other hand, differences in imaging quality and parameters among different retinal imaging devices lead to inconsistent model performance across devices. Additionally, the lack of a unified database of normal values and imaging acquisition protocols hinders the comparability of results. Furthermore, motion artifacts or signal attenuation may also interfere with the accuracy of analysis.

Due to the lack of clear AI decision-making processes and quantitative metrics or standards, AI is often viewed as a “black box” (98). This can lead to skepticism among doctors and patients regarding the results. Additionally, the data-driven nature of the model makes it susceptible to biases in the training data. This further exacerbates public distrust of AI.

Compared with single-modal models, multimodal models have better predictive performance. On the one hand, they can integrate technologies for identifying various types of retinal imaging. The synergistic application of multiple retinal imaging technologies, such as CFP, OCT, and OCTA, can be explored and combined with computer vision technology and self-supervised learning models (32, 99) to improve predictive accuracy. On the other hand, cross-modal data fusion can be conducted. By integrating genomic and metabolomic data (e.g., the photoreceptor metabolic window model), the specificity and clinical value of the model can be enhanced. Additionally, the model can be deployed in clinical settings for validation. Referencing models combining ECG with retinal imaging (75) and studies combining OCTA parameters with carotid artery stenosis (100), the validation of multimodal models in real clinical scenarios can be advanced.

The lack of interpretability in AI models significantly impacts the trust doctors and patients have in them. On the one hand, Visualization techniques should be used as much as possible to identify retinal imaging. For example, gradient-weighted class activation mapping (Grad-CAM) heatmaps (28, 32, 49) and shapley additive explanations (SHAP) models (101) can be used to visually demonstrate the key retinal regions the model focuses on (such as blood vessels, the macula, and the optic disc). Analyze the anatomical structures relied upon by the model using attention weights to enhance clinical credibility (80). On the other hand, mathematical modeling methods can be explored to improve interpretability. Develop interpretable algorithms such as adaptive elliptical templates (102) to maintain robustness under conditions of abundant lesions or low contrast, thereby addressing the “black box” challenge of AI. Research suggests that explainable AI (XAI) technology can reveal black-box ML models, enhancing model credibility and reliability (103). One study incorporated SHAP analysis into AI models to achieve better model understanding (101, 104). Future AI models may incorporate XAI technology and SHAP analysis to improve model credibility (101, 103, 104). Additionally, the model combines multi-scale feature extraction and fusion techniques with a dual attention mechanism, which promotes the extraction of multi-scale vascular features and may help improve the model's interpretability (105). In addition, a pilot study evaluating HbA1c demonstrated the potential and considerations required to develop reliable AI in the oculomics pilot, which contributes to the transparency of AI models (106).

Current AI models for CVD screening are predominantly trained on limited, single-center datasets, which restricts their generalizability and clinical applicability. There is a pressing need to develop large-scale, multi-center, and multi-ethnic datasets to enhance model robustness and performance across diverse populations. For instance, the Retinal OCT Angiography and Cardiovascular Status (RASTA) dataset includes retinal microvascular imaging from 499 patients, featuring 814 vascular cuboids and 2,005 facial images, and represents the only publicly available resource with imaging data from both healthy individuals and high-risk CVD populations (107). This dataset is expected to facilitate the development of universal screening models using OCT-A imaging (107). Similarly, the mBRSET dataset—the first publicly available diabetic retinopathy resource—contains 5,164 retinal images from 1,291 ethnically diverse patients, all acquired using handheld cameras, thereby addressing data scarcity in low- and middle-income settings (108). A related study outlines steps for constructing a large-scale online retinal imaging database in India, offering a replicable framework for cost-effective, AI-based diagnostic tool development (109).

To improve model generalizability, future efforts should prioritize standardized, multi-institutional datasets that incorporate varied imaging devices (e.g., portable cameras) and population characteristics. External validation across diverse datasets—such as the UK Biobank and Eye Picture Archive Communication System (PACS), as performed for the CLAiR model—is also essential. One study reported a DL model capable of predicting glaucoma progression with varying accuracy across ethnic groups (76.9% in Caucasians, 14.6% in African Americans, and 8.5% in Asians), highlighting the importance of ethnically balanced training data (22). Insights from such studies can inform the design of more generalizable AI systems (13, 22). Finally, establishing unified data annotation standards will be critical to supporting reproducible and scalable model development.

Due to the non-invasive, efficient, and convenient nature of retinal imaging, they are particularly suitable for primary care settings and resource-limited regions. Therefore, future efforts could focus on designing portable devices to enhance clinical practicality. A study comparing automated and semi-automated methods for measuring retinal microvascular biomarkers found good correlation between the two approaches in assessing vascular complexity and vessel diameter measurements, with consistent clinical relevance (110). However, the automated model exhibited a higher rejection rate within the dataset (110). Before transitioning from semi-automated to automated algorithms in retinal microvascular biomarker analysis, further comparative research is warranted (110). Additionally, dynamic risk assessment models could be developed to integrate long-term follow-up data on retinal changes and CVD progression for personalized management.