Abstract
Background
Coronary heart disease (CHD) is a major cause of mortality worldwide. This study aimed to develop and validate a multimodal deep learning algorithm using retinal imaging to assist in CHD risk assessment.
Methods
In this retrospective study, we developed a deep learning algorithm that integrates retinal fundus photographs and optical coherence tomography (OCT) images. A clinical nomogram was also developed by combining the imaging-based predictions with clinical risk factors. Model performance was evaluated on internal validation and test cohort using receiver operating characteristic (ROC) analysis and calibration curves.
Results
The algorithm was developed and validated using a dataset of 505 patients, which included 282 with CHD. On the training cohort, the model achieved an area under the curve (AUC) of 0.9954 (95% CI: 0.9904–1.0000). On the independent validation and test cohorts, the model achieved AUCs of 0.9834 (95% CI: 0.9556–1.0000) and 0.9138 (95% CI: 0.8404–0.9871), respectively. The nomogram demonstrated an AUC of 0.9963 (95% CI, 0.9923–1.0000) on the training cohort, and AUCs of 0.9423 (95% CI, 0.8626–1.0000) and 0.9153 (95% CI, 0.8289–1.0000) on the internal validation and test cohort, respectively.
Conclusions
This study demonstrates the feasibility of using a deep learning algorithm based on retinal imaging for CHD risk stratification. Future prospective, multicenter studies are needed to validate these findings and evaluate their potential clinical utility.
Supplementary Information
The online version contains supplementary material available at 10.1186/s12880-026-02551-5.
Key words: Multimodal model, Deep learning, Coronary heart disease, Retinal fundus photograph, Optical coherence tomography
Background
Coronary heart disease (CHD), a cause of global mortality, imposes a substantial health burden and reduces the quality of life for affected individuals [1]. The diagnosis of CHD relies on coronary angiography and computed tomography (CT), which are technically demanding and costly [2]. In recent years, the extensive development of artificial intelligence (AI), including deep learning algorithms, has facilitated the precise diagnosis of cancer [3, 4], infective diseases [5], and cardiovascular diseases (CVD) [6, 7] such as hypertension [8], arrhythmia [9], and CHD [10]. However, in prior research [11], conventional AI models for CHD diagnosis are based on coronary angiography, coronary CT angiography, intravascular imaging, and cardiac magnetic resonance. These methods are both resource-dependent and costlier than non-invasive ocular imaging, with their effectiveness often tied to the clinical setting and capabilities of the treating hospital. Therefore, this study sought to develop and validate a risk stratification algorithm based on non-invasive, cost-effective retinal imaging, to predict coronary heart disease risk and thereby facilitate the identification of high-risk individuals for targeted preventive interventions.
Structural and functional abnormalities in retinal microcirculation are associated with the incidence and adverse outcomes of CVD [12]. The structural similarity between the ocular and systemic vasculature enables the retinal vessels to serve as a window for CVD risk assessment [13]. Consequently, deep-learning models that automatically quantify retinal vessel features [6] from non-invasive retinal fundus photographs and OCT images hold potential for screening and identifying individuals at high risk of CVD.
Our present study developed a deep-learning algorithm to analyze retinal fundus photographs and OCT images from 505 patients for coronary heart disease risk stratification, which facilitates large-scale screening with the ultimate goal of preventing adverse cardiovascular events and mitigating the associated economic burden in high-risk populations.
Methods
Patients
This retrospective study enrolled adult patients from Anzhen Hospital between January 2022 and December 2024. The inclusion criteria were: ①Patients diagnosed with CHD by coronary angiography or coronary CT angiography; ②Patients who underwent retinal fundus photography and OCT, with binocular fundus photos and OCT images acquired according to standard protocols; ③Patients with complete medical records and laboratory examination data. The exclusion criteria were: ①No record of coronary angiography or coronary CT angiography; ②Patients with severe cataract, vitreous hemorrhage, or other ocular diseases that could affect retinal imaging or require close clinical monitoring (e.g., any stage of age-related macular degeneration, moderate or worse diabetic retinopathy, retinal vein/artery occlusion, epiretinal membrane, or macular hole); ③Patients with intraocular inflammation, or a history of laser therapy, intraocular surgery, or intravitreal medication; ④Incomplete medical records. Informed consent was obtained from all participants and included a statement on the clinicopathological data for scientific research. All participants also completed a baseline questionnaire (Supplementary Table S1). This study was approved by the Ethics Committee of Anzhen Hospital, Capital Medical University (approval number: 2022194x).
Case records
Comprehensive case records for patients included in this study were collected. Data on hypertension (HTN) and diabetes mellitus (DM) diagnoses were obtained from questionnaires and physical examinations. A history of HTN was self-reported and cross-verified against available medical records and medication lists. DM was defined as a fasting blood glucose level of ≥7.0 mmol/L, a 2-hour postprandial blood glucose level of ≥11.1 mmol/L, or a documented history of antidiabetic medication use (including insulin or oral agents). Biochemical parameters measured from blood samples included total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL), low-density lipoprotein cholesterol (LDL), and glucose.
Retinal imaging collection
Retinal fundus photographs of both eyes were acquired during hospital admission using a non-mydriatic fundus camera (CRO Plus, Micro Clear, China). For each eye, 60° non-stereo fundus images centered on the macula were obtained, encompassing the optic disc and most of the temporal vascular arcade. All color images were digitally stored in JPEG format at a resolution of 2176 × 2296 pixels (Fig. 1A).
Fig. 1.
Representative images for retinal fundus photography and OCT. (A) Representative images for retinal fundus photograph. (B) Representative images for OCT
OCT scans were acquired using a Cirrus HD-OCT 5000 device (Carl Zeiss Meditec, Inc., USA). For each participant, a single macular scan centered on the fovea was obtained from each eye. Each scan was saved in Audio Video Interleaved format (Fig. 1B).
Model training and implementation details
This study used a multimodal retinal image-assisted classification network model for diagnosing CHD.
Preprocessing and Data Augmentation: We used Maximum Intensity Projection (MIP) along the depth (axial) axis for three-dimensional (3D) OCT image processing. This technique selectively preserved the structures with the highest signal across all depth layers into a single two-dimensional (2D) en face image, while preserving the original three-color (red, green and blue) channel attributes of the OCT volume data. The obtained 2D OCT images and retinal fundus photographs were then scaled to 299 × 299 pixels. The retinal fundus photographs and 2D-OCT images of the corresponding eyes were utilized as samples to establish the dataset for the classification model in this study. We applied online data augmentation to the training cohort to increase data diversity, using random horizontal and vertical flipping, brightness adjustment, and contrast adjustment.
Model Construction: Subsequently, we designed a CHD classification network model based on multimodal retinal images to integrate the information from fundus photographs and OCT images. The model consisted of two parallel feature extraction branches, a feature fusion module based on the attention mechanism and a classifier based on fully connected layers (Fig. 2A). The image feature extraction branch refers to the InceptionV4 [14] network architecture (Fig. 2B). To effectively integrate the features extracted from different modalities of images, based on feature concatenation, this study used convolution block attention module (CBAM) to perform a weighted fusion of multimodal image features. Finally, the weighted multimodal feature map was input into a classifier composed of fully connected layers.
Fig. 2.
Schematic overview of the multimodal approaches for AI deep learning. (A) Overall architecture diagram of classification network. (B) Structure diagram of Inception
We conducted training on binocular images and selected the eyes with a higher probability of coronary heart disease predicted by the model as the final output to obtain the probability for diagnosing CHD. Following the generation of initial predictive probabilities by the AI system, a post-processing calibration step was implemented to enhance the reliability and monotonicity of the output scores. This was achieved by applying the isotonic regression method, a non-parametric technique designed to fit a piecewise constant, non-decreasing function to the data (R package “isotone”, Version 1.1–2). The procedure effectively maps the raw, potentially miscalibrated probabilities onto a calibrated scale while preserving their original rank order. The calibration model was fitted using the observed binary outcomes against the AI-predicted probabilities, ensuring that the final scores are both data-driven and optimally aligned with the empirical event rates. This transformation yields a final scoring system where higher scores correspond strictly to higher estimated probabilities, thereby improving interpretability and clinical utility.
(3) Implementation Details and Computational Environment: The training was carried out on a single NVIDIA GeForce RTX 4090 (24 GB VRAM) GPU. The complete training process (including validation) took approximately 1.5 hours. All code was developed and executed on a system running Ubuntu 20.04 LTS, utilizing PyTorch 2.5.1 with CUDA 12.1 support.
Weighted Cross-Entropy Loss was adopted during training and the final hyperparameters are selected based on on-site practice and preliminary experiments in the internal validation cohort. We employed Adam optimizer with an initial learning rate of 0.0001, mini - batch size of 16 to train the network. The weight factors of auxiliary loss in Eq. 1 were all set to an equal value of 1. The training process was terminated after 300 epochs as the network had converged.
Statistical analysis
Baseline characteristics are presented as percentages for categorical variables and as mean ± standard deviation or median [interquartile range] for continuous variables. Group comparisons were performed using the chi-square test for categorical variables. For continuous variables, the Student’s t-test or the Mann-Whitney U test was applied based on data distribution. Receiver operating characteristic (ROC) analysis was conducted using the “pROC” package (version 1.17.0.1) [15] to evaluate model discrimination, with the optimal cutoff value determined by maximizing Youden’s index. A nomogram was constructed using the “rms” package (version 6.1), incorporating predictors identified by multivariate logistic regression. The length of each variable’s line in the nomogram is proportional to its contribution to the outcome. Individual risk scores were derived using the “nomogramFormula” package (version 1.2). The AI algorithm and nomogram’s calibration were assessed using calibration curves, and its clinical utility was evaluated via decision curve analysis with the “rmda” package (version 1.6). For the calibration curves presented in our study, we applied bias correction using bootstrapping (also referred to as overfitting-correction or optimism correction) via the rms:calibrate function in R software. All statistical analyses were conducted using R software (version 4.0.3). A two-sided p-value < 0.05 was considered statistically significant.
Results
Characteristics of the study population
A total of 505 patients were included in this study, comprising 265 males (52.5%) and 240 females (47.5%), with an age range of 29–104 years. We performed cross-verification of self-reported hypertension and diabetes history using available medical records—specifically, hospital electronic health records and medication lists. Among the participants, 342 (67.7%) had hypertension, 218 (43.2%) had diabetes, and 282 (55.8%) were diagnosed with CHD according to the gold standard of diagnosis. Baseline characteristics of the cohort are presented in Table 1. Patients with CHD were significantly older and had a higher prevalence of hypertension and diabetes than controls (p < 0.001 for all). Notably, TC and LDL levels were significantly lower in the CHD group (p < 0.001), which may be attributable to the widespread use of statin therapy in this population.
Table 1.
Baseline characteristics for CHD patients in Anzhen Hospital cohort
| Variables | Overall (N = 505) |
Group | p Value | ||
|---|---|---|---|---|---|
| Without CHD (N = 223) |
With CHD (N = 282) |
||||
| Age (Years, mean ± SD) | 69.08 ± 11.78 | 65.81 ± 12.74 | 71.66 ± 10.27 | <0.001 | |
| Sex (%) | Female | 240 (47.5) | 113 (50.7) | 127 (45.0) | 0.242 |
| Male | 265 (52.5) | 110 (49.3) | 155 (55.0) | ||
| TC (mmol/L, median [IQR]) |
4.48 [3.76, 5.30] |
4.85 [4.14, 5.50] |
4.22 [3.54, 5.06] |
<0.001 | |
| TG (mmol/L, median [IQR]) |
1.37 [0.99, 1.91] |
1.43 [1.04, 2.10] |
1.32 [0.97, 1.87] |
0.129 | |
| LDL (mmol/L, median [IQR]) |
2.53 [1.87, 3.28] |
2.93 [2.27, 3.55] |
2.22 [1.66, 2.95] |
<0.001 | |
| HDL (mmol/L, median [IQR]) |
1.25 [1.04, 1.49] |
1.24 [1.06, 1.47] |
1.27 [1.03, 1.50] |
0.9 | |
| Glucose (mmol/L, median [IQR]) |
5.90 [5.30, 6.82] |
5.68 [5.21, 6.71] |
6.04 [5.39, 6.91] |
0.004 | |
| Hypertension (%) | No | 163 (32.3) | 109 (48.9) | 54 (19.1) | <0.001 |
| Yes | 342 (67.7) | 114 (51.1) | 228 (80.9) | ||
| Diabetes (%) | No | 287 (56.8) | 147 (65.9) | 140 (49.6) | <0.001 |
| Yes | 218 (43.2) | 76 (34.1) | 142 (50.4) | ||
| Statin (%) | No | 203 (40.2) | 136 (61.0) | 67 (23.8) | <0.001 |
| Yes | 302 (59.8) | 87 (39.0) | 215 (76.2) | ||
The differences between the 2 groups of categorical variables were evaluated by Pearson’s chi-square test. The differences between the 2 groups of normally distributed continuous variables (Age) were calculated using the t test. The differences between the 2 groups of non-normally distributed continuous variables (TC, TG, LDL, HDL and Glucose) were calculated using the Wilcoxon rank-sum test. IQR: interquartile range. SD: standard deviation, CHD: coronary heart disease, TC: total cholesterol, TG: triglyceride, HDL: high-density lipoprotein, LDL: low-density lipoprotein
Model interpretation
Our multimodal network was interpreted using Gradient-weighted Class Activation Mapping (Grad-CAM) (Fig. 3) to visualize the model’s focus areas. The resulting heat maps indicate that the model consistently attended to vascular-rich regions and the macula in the fundus photographs. These findings are consistent with the known pathophysiological connections between retinal microvascular changes [16] and systemic cardiovascular health [17] and enhance the model’s transparency and explanatory power.
Fig. 3.
Representative feature heatmaps for CHD patients with Grad-CAM
Predictive value of AI-based image analysis for diagnosing coronary heart disease
All data were randomly partitioned at the patient level into training, internal validation and internal test cohort in an 8:1:1 ratio, ensuring that all images from the same patient were allocated to the same subset. This rigorous partitioning approach preserved the integrity of the evaluation. The model outputs a continuous risk score for CHD diagnosis. The predictive performance of the multimodal model (fundus photography + OCT) was evaluated using receiver operating characteristic (ROC) analysis. In the training cohort, the model achieved an area under the curve (AUC) of 0.9954 (95% CI: 0.9904–1.0000; Fig. 4A). At an optimal cut-off value of 1.661 for the AI risk score, the sensitivity was 0.994 and the specificity was 0.978 (Fig. 4A). We also assessed a simplified model and a fundus-only model (without OCT). The simplified model attained an AUC of 0.9920 (95% CI: 0.9852–0.9987), and the fundus-only model achieved an AUC of 0.9847 (95% CI: 0.9749–0.9945) (Fig. 4A). Delong’s test for comparison of ROC curves indicated that the multimodal model had significantly better discriminative ability for identifying high-risk patients than the fundus-only model (p = 0.0124; Fig. 4B). In both the validation and test cohorts, the multimodal AI model demonstrated higher AUC values compared with the simplified model and the fundus-only model (Fig. 4C–D). In the validation cohort, the multimodal model yielded a sensitivity of 92.31%, specificity of 95.45%, positive predictive value (PPV) of 96.00%, negative predictive value (NPV) of 91.30%, and accuracy of 93.75%. Corresponding metrics in the test cohort were 82.76% sensitivity, 86.96% specificity, 88.89% PPV, 80.00% NPV, and 84.62% accuracy.
Fig. 4.
Predictive performance of AI-based image analysis for diagnosing coronary heart disease. (A) Receiver operating characteristic (ROC) curves of three AI deep-learning models for predicting CHD in the training cohort: the multimodal model (fundus photography + OCT), the simplified model, and the fundus photography‑only model (without OCT). (B) Statistical comparison of the ROC curves among the three AI models using DeLong’s test (p‑values shown in table). (C-D) ROC curves of the three AI models for predicting CHD in the validation cohort (C) and the test cohort (D). (E-F) Calibration curves of the multimodal model for predicting CHD in the validation cohort (E) and the test cohort (F). (G) Decision curve analysis comparing the clinical utility of the multimodal model and the fundus photography‑only model in the validation and test cohorts
Model calibration was evaluated using calibration curves and the Hosmer–Lemeshow test (p > 0.05), which indicated consistency between predicted probabilities and observed outcomes (Fig. 4E–F). Finally, decision curve analysis revealed that the multimodal model provided greater clinical benefit and utility than the fundus photography-only model across the validation and test cohorts (Fig. 4G).
Identification of clinical features for the CHD predictive model
Univariable and subsequent multivariable logistic regression analyses were performed to identify independent clinical predictors of CHD in the training cohort. Based on the initial univariable screening, LDL cholesterol level, glucose level, hypertension, and diabetes were selected as candidate predictors and were included in the multivariable model. Multivariable analysis identified age (OR = 1.043, 95% CI: 1.023–1.063; p < 0.001), LDL (OR = 0.577, 95% CI: 0.459–0.726; p < 0.001), hypertension (OR = 3.080, 95% CI: 1.907–4.976; p < 0.001), and diabetes (OR = 1.672, 95% CI: 1.060–2.637; p = 0.027) were retained as significant independent predictors (Table 2). Notably, LDL cholesterol exhibited an inverse association with CHD (i.e., lower levels were associated with higher risk), a finding likely reflecting the prevalent use of lipid-lowering therapy in this clinical population.
Table 2.
Multivariate regression analysis of the risk factors for CHD
| Variables | Univariable Regression | Multivariable Regression | |||||
|---|---|---|---|---|---|---|---|
| OR | 95%CI | p Value | OR | 95%CI | p Value | ||
| Gender | Female | 1.000 | 0.261 | ||||
| Male | 1.253 | 0.846–1.857 | |||||
| Age | 1.041 | 1.023–1.060 | <0.001 | 1.043 | 1.023–1.063 | <0.001 | |
| TC (mmol/L) | 1.000 | 0.999–1.000 | 0.854 | ||||
| TG (mmol/L) | 0.901 | 0.785–1.032 | 0.133 | ||||
| LDL (mmol/L) | 0.531 | 0.428–0.660 | <0.001 | 0.577 | 0.459–0.726 | <0.001 | |
| HDL (mmol/L) | 1.121 | 0.664–1.894 | 0.669 | ||||
| Hypertension | No | 1.000 | <0.001 | 1.000 | <0.001 | ||
| Yes | 4.000 | 2.567–6.232 | 3.080 | 1.907–4.976 | |||
| Diabetes | No | 1.000 | 0.001 | 1.000 | 0.027 | ||
| Yes | 1.969 | 1.314–2.951 | 1.672 | 1.060–2.637 | |||
TC: total cholesterol, TG: triglyceride, HDL: high-density lipoproteincholesterol, LDL: low-density lipoprotein cholesterol, OR: odds ratio, CI:confidence interval
Nomogram construction and validation
We developed a nomogram incorporating age, hypertension, diabetes, and the AI-derived risk stratification score (cut‑off: 1.661) (Fig. 5A). A reference nomogram incorporating only the clinical predictors (age, hypertension, diabetes) was also developed for comparison (Figure S1A). The decision to exclude LDL cholesterol from the final model was based on prior multivariable analysis, which indicated an inverse association with CHD risk—a finding likely due to high statin usage within the CHD cohort or other confounders. Model comparison demonstrated that the exclusion of LDL cholesterol yielded a lower Akaike Information Criterion (AIC = 47.47) compared with the model retaining it (AIC = 51.11), indicating improved model fit and supporting its omission.
Fig. 5.
Nomogram for predicting coronary heart disease. (A) Nomogram integrating clinical predictors (age, hypertension, diabetes) with the AI-derived risk stratification (cut‑off: 1.661). (B) Receiver operating characteristic (ROC) curve of the nomogram with the AI-derived risk stratification for predicting CHD in the training cohort. (C-D) ROC curves comparing the nomogram with AI risk stratification versus the nomogram without AI in the validation (C) and test (D) cohorts. (E-F) Calibration curves of the nomogram with AI risk stratification in the validation (E) and test (F) cohorts. (G-H) Decision curve analysis evaluating the clinical utility of the nomogram with and without AI risk stratification in the validation (G) and test cohorts (H)
The predictive performance of both nomograms was evaluated using ROC analysis. In the training cohort, the AI-integrated nomogram demonstrated significantly higher discriminative ability, with an AUC of 0.9963 (95% CI: 0.9923–1.0000), compared to the clinical-only model (AUC = 0.7244, 95% CI: 0.6744–0.7743) (Fig. 5B and S1B). This superior performance was consistently observed in the validation and test cohorts, which was statistically confirmed by Delong’s test (p < 0.05) (Fig. 5C–D). In the validation cohort, the AI-integrated nomogram achieved a sensitivity of 92.31%, specificity of 95.45%, PPV of 96.00%, NPV of 91.30%, and accuracy of 93.75%. Corresponding metrics in the test cohort were 82.76% sensitivity, 86.96% specificity, 88.89% PPV, 80.00% NPV, and 84.62% accuracy. Calibration was assessed using calibration curves and the Hosmer–Lemeshow test. Both nomograms showed consistency between predicted and observed probabilities (p > 0.05) in the training and validation cohorts (Fig. 5E–F, Fig. S1C–D). Finally, decision curve analysis demonstrated that the AI-integrated nomogram provided greater clinical benefit across a range of probability thresholds than the clinical-only model in both the validation and test cohorts (Fig. 5G-H).
Discussion
Current definitive diagnostic modalities for CHD, such as coronary CT angiography and coronary angiogram, entail substantial costs, exposure to ionizing radiation, and significant resource utilization, thereby confining their use primarily to specialized clinical settings and hindering their feasibility for widespread screening. A deep-learning approach based on retinal fundus photographs and OCT is designed to automatically extract and analyze discriminative visual patterns from non-invasive, routinely obtained ocular images. Specifically, the algorithm analyzes geometric characteristics of the retinal vasculature—such as vessel diameter, curvature tortuosity, and branching angles—features that have been previously linked to cardiovascular diseases [18, 19]. This approach offers a potentially accessible and scalable tool for CHD risk assessment.
Published research on artificial intelligence in cardiovascular risk assessment has largely focused on analyzing retinal fundus photographs. In a landmark proof-of-concept study, Poplin et al. [16] applied a deep-learning model to fundus images for predicting future adverse cardiovascular events, achieving an AUC of 0.70 (95% CI: 0.648–0.740). Predictive performance improved to an AUC of 0.73 (95% CI, 0.69–0.77) after integrating traditional risk factors such as age, blood pressure, and body mass index. Notably, their model was developed to stratify future risk within a generally healthy population. In contrast, the objective of the present study is to diagnose existing CHD, which represents a distinct clinical application with different implications for patient management. Qu et al. [20] developed an automated retinal image analysis system (ARIA-CHD) incorporating a deep-learning model to estimate overall CHD risk. When applied specifically to patients with metabolic disorders, their model achieved an AUC of 0.96, demonstrating the potential utility of retinal imaging as a risk assessment tool within this high-risk subgroup. Son et al. [21] developed a deep-learning algorithm to identify individuals at high or low risk of coronary artery calcium (CAC) using retinal fundus photographs. Their model achieved AUCs of 0.823 (95% CI: 0.795–0.850) and 0.832 (95% CI: 0.802–0.863) using unilateral and bilateral images, respectively. This study highlights the potential of retinal imaging for large-scale cardiovascular risk screening, particularly for identifying individuals with high coronary artery calcium burden. Building upon this approach, the present study utilize a more comprehensive multimodal ocular dataset to develop a tool aimed directly at assisting in the diagnosis of existing CHD. Diaz-Pinto et al. [22] also reported a deep-learning system for CVD prediction using retinal fundus photography. In their retrospective study of 476 participants, the system yielded an AUC of 0.80, demonstrating a sensitivity of 0.74 and a specificity of 0.71 for predicting CVD events. Collectively, these studies relied solely on fundus photographs to predict CHD or CVD risk and have substantiated the association between retinal features and cardiovascular disease. However, the predictive performance of these models remains limited.
OCT is a non-invasive imaging modality that visualizes pathophysiological changes that are not discernible on fundus photography, including specific metrics such as choroidal thickness and the integrity of individual retinal layers. Therefore, the combined analysis of fundus photography and OCT can provide complementary information, mitigating analytical limitations associated with the poor quality of any single image. Recently, artificial intelligence techniques have been applied to analyze OCT imaging to explore its potential in diagnosing cardiovascular diseases. Long et al. [23] identified retinal ischemic peripheral lesions (RILPs) using OCT. Their study revealed an increased number of RILPs in patients with CVD and further demonstrated an association between the number of RILPs and CVD risk. Drakopoulos et al. [17] reported using AI to identify RILPs in OCT images. Retinal and choroidal thickness measurements obtained via OCT are often investigated as potential biomarkers for CVD risk. Chen et al. [24] reported that retinal nerve fiber layer thinning was independently associated with increased cardiovascular risk and improved risk reclassification. Kim et al. [25] revealed that reduced subfoveal choroidal thickness could serve as a useful biomarker for predicting CVD.
Previous studies have utilized multimodal approaches for cardiovascular assessment. A multimodal deep-learning model developed by Samsung Medical Center was used to identify CVD and achieved AUCs of 0.781 (95% CI: 0.766–0.798) and 0.872 (95% CI: 0.857–0.886) on internal validation and external validation, respectively [26]. Another deep-learning model based on fundus photography combined with dual-energy X-ray absorptiometry (DXA) to predict CHD reached an accuracy of 77.4% [27]. In contrast, our multimodal deep-learning model, which integrates retinal fundus photography with OCT, demonstrated superior diagnostic performance compared to these prior approaches and outperformed models relying solely on fundus photographs. Beyond imaging modalities, established clinical features also hold significant predictive value for CHD. Factors such as age, sex [28], hypertension [29], and diabetes [4] are well-established risk factors for CVD and are routinely used to stratify cardiovascular risk. Furthermore, TC, LDL, and HDL are established biochemical indicators for CHD risk assessment [30, 31]. The integration of these clinical features including age, hypertension and diabetes along with the AI stratification into a nomogram demonstrated strong predictive utility for CHD risk, yielding an AUC of 0.9963.
Future work will focus on several key developments. First, our algorithm will undergo further technical refinement. A multicenter, large-scale, high-quality, and standardized validation dataset will be assembled to validate the stability and generalizability of the model. We plan to develop an integrated, automated image quality assessment module to filter or annotate images prior to analysis, ensuring input data reliability. A software interface will be developed to connect with ophthalmic imaging devices, facilitating convenient image upload, automated analysis, and integration with other clinical data (e.g., laboratory results, medical history) to generate a comprehensive risk score. Ultimately, the AI system is intended to undergo rigorous clinical validation and cost-effectiveness analysis, and secure regulatory approval prior to clinical deployment.
However, this study had several limitations. Firstly, this study adopted the processing method of projecting 3D OCT data into 2D representations. Although the key intensity information was retained as much as possible through the MIP method, some tomographic information along the depth axis was inevitably lost. Future studies should perform a comparative analysis of the performance differences between MIP and other projection methods, while also exploring more advanced architectures, such as 3D-2D fusion networks or lightweight 3D convolutional neural networks, to make fuller use of volumetric data. Secondly, hypertension history was self-reported by a subset of participants, potentially introducing recall bias or misclassification bias. Although cross-verification was performed using available hospital records and medication lists, this potential measurement error must be acknowledged, and its implications for generalizability considered. Thirdly, since all patients underwent coronary angiography or coronary CT angiography, there was a significant selection bias, which limits the model’s generalizability to other clinical settings. Finally, the current model performance remains insufficient for direct clinical deployment and warrants further improvement and rigorous external validation.
Conclusion
Our study developed an AI algorithm based on retinal fundus photographs and OCT images for identifying individuals at high risk for CHD. Further technical refinement and external validation are required before this algorithm can be translated into a clinical software tool. Such a tool would potentially facilitate large-scale screening, ultimately aiming to prevent adverse cardiovascular events in high-risk populations.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Acknowledgements
We thank Yuwei Ling phD. MD. from Department of Hepatobiliary Surgery, Xijing Hospital, for the advice on processing raw data. We appreciate Jie Gong phD. and Professor Mei Shi phD. MD. from Department of Radiation Oncology, Xijing Hospital, for the professional advice.
Author contributions
Xinxiao Gao and Zongqing Ma guided the research direction, reviewed and revised the paper, and was responsible for the submission of the manuscript. Ran Yan was responsible for the study design, data collection and analysis, and was a major contributor in writing the manuscript. Xiaoxiao Guo, Yan Zhu and Tingting Hong performed the data collection. Jiang Zhu and Weijie Zhang was responsible for specific experimental manipulations and assisted in data analysis. All authors read and approved the final manuscript.
Funding
This work was supported by grants from China Health Promotion Foundation (20250088) and Beijing Anzhen Hospital (2025AZD4004).
Data availability
The datasets are not publicly available due to patient privacy and confidentiality but are available from the corresponding author on reasonable request.
Declarations
Ethics approval and consent to participate
The study was conducted in accordance with the rules of the Declaration of Helsinki and was approved by ethics committee of Anzhen Hospital, Capital Medical University. (Approval No. 2022194x).
Consent for publication
Each patient signed an informed consent form before starting the study.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Ran Yan and Xiaoxiao Guo contributed equally to this work.
Contributor Information
Zongqing Ma, Email: zqma@bistu.edu.cn.
Xinxiao Gao, Email: drxinxiaogao@hotmail.com.
References
- 1.Roth GA, Mensah GA, Johnson CO, Addolorato G, Ammirati E, Baddour LM, et al. C. D. W. Global burden of cardiovascular diseases and risk factors, 1990-2019. Update from the GBD, 2019 study. J Am Coll Cardiol. 2020;76(25):2982–3021. 10.1016/j.jacc.2020.11.010. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Liu CY, Tang CX, Zhang XL, Chen S, Xie Y, Zhang XY, et al. Deep learning powered coronary CT angiography for detecting obstructive coronary artery disease: the effect of reader experience, calcification and image quality. Eur J Radiol. 2021;142:109835. 10.1016/j.ejrad.2021.109835. [DOI] [PubMed] [Google Scholar]
- 3.Tran KA, Kondrashova O, Bradley A, Williams ED, Pearson JV, Waddell N. Deep learning in cancer diagnosis, prognosis and treatment selection. Genome Med. 2021;13(1):152. 10.1186/s13073-021-00968-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Coles B, Zaccardi F, Ling S, Davies MJ, Samani NJ, Khunti K. Cardiovascular events and mortality in people with and without type 2 diabetes: an observational study in a contemporary multi-ethnic population. J Diabetes Invest. 2021;12(7):1175–82. 10.1111/jdi.13464. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Theodosiou AA, Read RC. Artificial intelligence, machine learning and deep learning: potential resources for the infection clinician. J Infect. 2023;87(4):287–94. 10.1016/j.jinf.2023.07.006. [DOI] [PubMed] [Google Scholar]
- 6.Cheung CY, Xu D, Cheng CY, Sabanayagam C, Tham YC, Yu M, et al. A deep-learning system for the assessment of cardiovascular disease risk via the measurement of retinal-vessel calibre. Nat Biomed Eng. 2020;5(6):498–508. 10.1038/s41551-020-00626-4. [DOI] [PubMed] [Google Scholar]
- 7.Krittanawong C, Zhang H, Wang Z, Aydar M, Kitai T. Artificial intelligence in precision cardiovascular medicine. J Am Coll Cardiol. 2017;69(21):2657–64. 10.1016/j.jacc.2017.03.571. [DOI] [PubMed] [Google Scholar]
- 8.Loh EW, Huang W, Sultana R, Kong SC, Ng CJ, Tan SY, et al. Association of blood pressure and left ventricular mass with subclinical coronary atherosclerosis. Open Heart. 2025;12(1):e002791. 10.1136/openhrt-2024-002791. [DOI] [PMC free article] [PubMed]
- 9.Trayanova NA, Popescu DM, Shade JK. Machine learning in arrhythmia and electrophysiology. Circ Res. 2021;128(4):544–66. 10.1161/CIRCRESAHA.120.317872. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Lin A, Manral N, McElhinney P, Killekar A, Matsumoto H, Kwiecinski J, et al. Deep learning-enabled coronary CT angiography for plaque and stenosis quantification and cardiac risk prediction: an international multicentre study. Lancet Digit Health. 2022;4(4):e256–65. 10.1016/S2589-7500(22)00022-X. [DOI] [PMC free article] [PubMed]
- 11.Lu H, Yao Y, Wang L, Yan J, Tu S, Xie Y, et al. Research progress of machine learning and deep learning in intelligent diagnosis of the coronary atherosclerotic heart disease. Comput Math Method M. 2022;2022:1–14. 10.1155/2022/3016532. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Guo S, Yin S, Tse G, Li G, Su L, Liu T. Association between caliber of retinal vessels and cardiovascular disease: a systematic review and meta-analysis. Curr Atheroscler Rep. 2020;22(4). 10.1007/s11883-020-0834-2. [DOI] [PubMed]
- 13.London A, Benhar I, Schwartz M. The retina as a window to the brain—from eye research to CNS disorders. Nat Rev Neurol. 2013;9(1):44–53. 10.1038/nrneurol.2012.227. [DOI] [PubMed] [Google Scholar]
- 14.Szegedy C, Ioffe S, Vanhoucke V, Alemi A. Inception-v4, inception-ResNet and the impact of residual connections on learning. Proc AAAI Conf Artif Intel. 2017;31(1):4278–84. 10.1609/aaai.v31i1.11231. [Google Scholar]
- 15.Robin X, Turck N, Hainard A, Tiberti N, Lisacek F, Sanchez JC, et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves. BMC Bioinf. 2011;12(1). 10.1186/1471-2105-12-77. [DOI] [PMC free article] [PubMed]
- 16.Poplin R, Varadarajan AV, Blumer K, Liu Y, McConnell MV, Corrado GS, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng. 2018;2(3):158–64. 10.1038/s41551-018-0195-0. [DOI] [PubMed] [Google Scholar]
- 17.Drakopoulos M, Hooshmand D, Machlab LA, Bryar PJ, Hammond KJ, Mirza RG. Machine teaching allows for rapid development of automated systems for retinal lesion detection from small image datasets. Ophthalmic Surg Lasers Imag Retina. 2024;55(8):475–78. 10.3928/23258160-20240410-01. [DOI] [PubMed] [Google Scholar]
- 18.Wang SB, Mitchell P, Liew G, Wong TY, Phan K, Thiagalingam A, et al. A spectrum of retinal vasculature measures and coronary artery disease. Atherosclerosis. 2018;268:215–24. 10.1016/j.atherosclerosis.2017.10.008. [DOI] [PubMed] [Google Scholar]
- 19.Wong YL, Yu M, Chong C, Yang D, Xu D, Lee ML, et al. Association between deep learning measured retinal vessel calibre and incident myocardial infarction in a retrospective cohort from the UK biobank. BMJ Open. 2024;14(3):e079311. 10.1136/bmjopen-2023-079311. [DOI] [PMC free article] [PubMed]
- 20.Qu Y, Lee JJ, Zhuo Y, Liu S, Thomas RL, Owens DR, et al. Risk assessment of CHD using retinal images with machine learning approaches for people with cardiometabolic disorders. JCM. 2022;11(10):2687. 10.3390/jcm11102687. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Son J, Shin JY, Chun EJ, Jung KH, Park KH, Park SJ. Predicting high coronary artery calcium score from retinal fundus images with deep learning algorithms. Trans Vis Sci Tech. 2020;9(2):28. 10.1167/tvst.9.2.28. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Diaz-Pinto A, Ravikumar N, Attar R, Suinesiaputra A, Zhao Y, Levelt E, et al. Predicting myocardial infarction through retinal scans and minimal personal information. Nat Mach Intell. 2022;4(1):55–61. 10.1038/s42256-021-00427-7. [Google Scholar]
- 23.Long CP, Chan AX, Bakhoum CY, Toomey CB, Madala S, Garg AK, et al. Prevalence of subclinical retinal ischemia in patients with cardiovascular disease – a hypothesis driven study. EClinicalMedicine. 2021;33:100775. 10.1016/j.eclinm.2021.100775. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Chen Y, Yuan Y, Zhang S, Yang S, Zhang J, Guo X, et al. Retinal nerve fiber layer thinning as a novel fingerprint for cardiovascular events: results from the prospective cohorts in UK and China. BMC Med. 2023;21(1):24. 10.1186/s12916-023-02728-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Kim JH, Kim SE, Kim SH, Choi BW, Rim TH, Byeon SH, et al. Relationship between coronary artery calcification and central chorioretinal thickness in patients with subclinical atherosclerosis. Ophthalmologica. 2021;244(1):18–26. 10.1159/000506056. [DOI] [PubMed] [Google Scholar]
- 26.Lee YC, Cha J, Shim I, Park WY, Kang SW, Lim DH, et al. Multimodal deep learning of fundus abnormalities and traditional risk factors for cardiovascular risk prediction. NPJ Digit Med. 2023;6(1):14. 10.1038/s41746-023-00748-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Al-Absi HRH, Islam MT, Refaee MA, Chowdhury MEH, Alam T. Cardiovascular disease diagnosis from DXA scan and retinal images using deep learning. Sensors (Basel). 2022;22(12):4310. 10.3390/s22124310. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Kane AE, Howlett SE. Differences in cardiovascular aging in men and women. Adv Exp Med Biol. 2018;1065:389–411. 10.1007/978-3-319-77932-4_25. [DOI] [PubMed] [Google Scholar]
- 29.Zhu JZ, Manne-Goehler J, Agarwal A, Bahendeka SK, Damasceno A, Marcus ME, et al. Population project on access to care for cardiometabolic, D. Medication use for cardiovascular disease prevention in 40 low- and middle-income countries. J Am Coll Cardiol. 2023;81(6):620–22. 10.1016/j.jacc.2022.12.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Stamler J, Daviglus ML, Garside DB, Dyer AR, Greenland P, Neaton JD. Relationship of baseline serum cholesterol levels in 3 large cohorts of younger men to long-term coronary, cardiovascular, and all-cause mortality and to longevity. JAMA. 2000;284(3):311–18. 10.1001/jama.284.3.311. [DOI] [PubMed] [Google Scholar]
- 31.Casula M, Colpani O, Xie S, Catapano AL, Baragetti A. HDL in Atherosclerotic cardiovascular disease: in search of a role. Cells. 2021;10(8):1869. 10.3390/cells10081869. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The datasets are not publicly available due to patient privacy and confidentiality but are available from the corresponding author on reasonable request.