Artificial intelligence-based prediction of neurocardiovascular risk score from retinal swept-source microvascular imaging: the RASTA dataset

Artificial intelligence-based prediction of neurocardiovascular risk score from retinal swept-source microvascular imaging: the RASTA dataset | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Artificial intelligence-based prediction of neurocardiovascular risk score from retinal swept-source microvascular imaging: the RASTA dataset Clement Germanese, Atif Anwer, Petra Eid, Laure-Anne Steinberg, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4326028/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 07 Nov, 2024 Read the published version in Scientific Reports → Version 1 posted 10 You are reading this latest preprint version Abstract The recent rise of artificial intelligence represents a revolutionary way of improving current medical practices, including cardiovascular (CV) assessment scores. Retinal vascular alterations may reflect systemic processes such as the presence of CV risk factors. The value of swept-source retinal optical coherence tomography–angiography (SS OCT-A) imaging is significantly enhanced by image analysis tools that provide rapid and accurate quantification of vascular features. We report on the interest of using machine-learning (ML) and deep-learning (DL) models for CV assessment from SS OCT-A microvasculature imaging. We assessed the accuracy of ML and DL algorithms in predicting the CHA 2 DS 2 -VASc neurocardiovascular score based on SS OCT-A retinal images of patients from the open-source RASTA dataset. The ML and DL models were trained on data from 491 patients. The ML models tested here achieved good performance with area under the curve (AUC) values ranging from 0.71 to 0.96. According to a classification into two or three CV risk groups, the EfficientNetV2-B3 tool predicted risk correctly in 39% and 68% of cases, respectively, with a mean absolute error (MAE) of approximately 0.697. Our models enable a confident prediction of the CHA 2 DS 2 -VASc score from SS OCT-A imaging, which could be a useful tool contributing to the assessment of neurocardiovascular profiles in the future. Health sciences/Biomarkers Health sciences/Cardiology Health sciences/Neurology Health sciences/Risk factors OCT-A CHA2DS2-VASc machine learning deep learning neurocardiovascular prediction Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction With approximately 17.9 million deaths every year 1 , cardiovascular diseases (CVD) are the leading cause of death worldwide, and the assessment of cardiovascular (CV) risk plays a major role in their prevention. Today, CV risk profiles can be estimated with numerous validated score models such as the Framingham Risk Score (FRS), the American Heart Association risk score (AHA risk score), or the SCORE2 2–4 . Nevertheless, these clinical scores do not take into account personalized vascular conditions, and recurrent neurovascular and cardiovascular events are still difficult to predict 5 , 6 . We previously demonstrated through a conventional regression approach in post-myocardial infarction patients in the EYE-MI study that retinal optical coherence tomography–angiography (OCT-A) could be an effective tool for predicting cardiovascular risk scores 7 . Swept-source OCT–angiography (SS OCT-A) technology, with more thorough retinal microvasculature description and better resolution, could be used to refine the EYE-MI pilot study results in patients with broader phenotypes 8 – 13 . Moreover, the CHA 2 DS 2 -VASc score, a cardioembolic risk score 14 , may better reflect the connection between the retinal, neurovascular, and cardiovascular systems. The role of advanced imaging analysis is becoming increasingly important in daily clinical practice and biomedical research due to the recent advancements in oculomics 15 and other artificial intelligence (AI)-based image analysis methods 16 . In the framework of open-source datasets in ophthalmology 17 , we previously published the Retinal oct-Angiography and cardiovascular STAtus (RASTA) dataset 18 . It combined SS OCT-A retinal imaging and CHA 2 DS 2 -VASc score calculation in patients with various conditions (e.g., diabetes mellitus, giant cell arthritis, dyslipidemia, and ischemic CVD) and accurate phenotype. Recent studies using deep-learning (DL) convolutional neural networks (CNNs) have associated retinal photographs with several CVD risk factors, including diabetes mellitus, blood pressure, body mass index (BMI), smoking, and glycated-hemoglobin level 19 – 22 . Nevertheless, the literature is sparse regarding the interest of SS OCT-A for artificial intelligence-based prediction of the neurocardiovascular risk score. The purpose of this study was to train and validate machine-learning- (ML) and DL-based models capable of predicting neurocardiovascular risk scores (CHA 2 DS 2 -VASc score) using the detection method from SS OCT-A acquisitions in the RASTA dataset. Results Dataset Overall, 491 patients were registered in this study. Of the 491 patients, 225 (45.8%) patients had a low neurocardiovascular risk and 266 (54.2%) had an intermediate–high neurocardiovascular risk. The mean age of the patients was 52.4 ± 18.4 years and 51.5% were women. All the cardiovascular clinical data and quantitative SS OCT-A variables were significantly different between the two neurocardiovascular risk categories. A comparison of the clinical and demographic variables is available in the supplementary material ( supplementary Table S1 ). Machine-learning model results The SVM model performed better than the others in predicting neurocardiovascular risk categories (AUC 0.98 ± 0.03 versus 0.96 ± 0.02 and 0.91 ± 0.04 and 0.78 ± 0.12 for logistic regression, RF, and decision tree, respectively). The results are shown in Table 1 . Table 1 Machine-learning model results Decision tree Random forest SVM Logistic regression AUC Accuracy (%) AUC Accuracy (%) AUC Accuracy (%) AUC Accuracy (%) k folds (k = 10) 0.78 ± 0.12 77.2 ± 3.4 0.91 ± 0.04 81.2 ± 2.9 0.98 ± 0.03 85.1 ± 5.9 0.96 ± 0.02 84.9 ± 5.6 k folds (k = 5) 0.77 ± 0.06 75.8 ± 3.0 0.90 ± 0.06 80.8 ± 2.3 0.98 ± 0.02 84.2 ± 5.1 0.96 ± 0.02 84.4 ± 5.2 SVM: support vector machines. AUC: area under the curve. Descriptive results for models are presented as mean ± standard deviation Deep-learning model results For classifying the two categories, the network was able to achieve a balanced accuracy of 68% as compared to 61% and 54% for RF and RF-FO, respectively. In evaluating the performance of the networks on the test set, the EfficientNet model and both variants of the RF model (RF and RF-FO) exhibited similar results. All the models demonstrated the same mean absolute error (MAE) of approximately 0.697. The R 2 scores for the two models were identical at − 0.9446. In terms of balanced accuracy, the EfficientNet model outperformed the RF models, achieving a balanced accuracy of 39%, compared to 33% for both RF variants. Main results are shown in Table 2 . The learning curves are presented in the supplementary material ( supplementary Figure S1 ). Classification results details are shown in the supplementary material ( supplementary Figure S2 ). Table 2 Performance of deep-learning-based model EfficientNetB3-V2 Random forest Random forest features only Balanced accuracy 0.68 0.61 0.54 Accuracy 0.67 0.58 0.52 R 2 score -0.3444 -0.7111 -0.9556 Discussion In this study, we found good results for neurocardiovascular risk category prediction based on retinal SS OCT-A, with a strong predictive accuracy of up to 98% for ML models and 68% for the EfficientNetV2-B3 backbone model. To our knowledge, this is the first study to investigate artificial intelligence-based prediction of neurocardiovascular risk score with SS OCT-A. Our results highlight the potential value of SS OCT-A as a biomarker of global neurocardiovascular status. Fundus photography (FP) was the first imaging modality in ophthalmology to prove its value in automatic CV risk assessment. Most hospital departments and practitioners have FP equipment, which has led to the creation of many rich databases such as the UK Biobank and MESSIDOR 31 . These databases, containing hundreds or thousands of fundus images, have been used in several studies to assess and predict CVD. Our results on artificial intelligence-based prediction of neurocardiovascular risk score with SS OCT-A were in line with previous studies based on FP. Poplin et al. used FPs to demonstrate the contribution of retinal vasculature to the automatic detection of CVD and CVD risk factors, where AUC values of DL models were greater than 0.70 and demonstrated their effectiveness in predicting some CVD risk factor and the occurrence of major adverse cardiovascular events (MACE) over a 5-year period 19 . Cheung et al. published their work on the assessment of CVD risk via automatic measurement of retinal-vessel caliber (RVC) 32 . They developed and tested a DL model to specifically measure RVC from more than 70,000 FPs. They assessed the agreement of the RVC measurement between the DL model and a human expert. The DL models predicted CVD risk factors significantly better than the human-based models or were at least comparable 32 . More recently, Zhang et al. demonstrated the capability of DL models to identify chronic kidney disease and diabetes mellitus using 115,344 FPs alone or in combination with clinical metadata (i.e., age, sex, BMI, and blood pressure), with AUCs ranging from 0.85 to 0.93. Additionally, the models could predict glomerular filtration rates and blood glucose levels, yielding MAEs of 11.1–13.4 mL/min per 1.73 m 2 and 0.65–1.1 mmol/L, respectively 33 . However, the exploration of microvasculature at the micrometer level in various plexuses and vascular networks made possible by SS OCT-A could offer hope for an even more accurate and earlier assessment of CV risk compared to FP. Based on previous ophthalmological research, Hassan et al. conducted an evaluation using three-dimensional CNNs to predict the individual age and sex directly from 3D retinal OCT scans, using a large dataset comprising 66,767 participants from the UK Biobank dataset. Model results showed accurate predictions for age (MAE = 3.30 years, R ²=0.89) and for sex (AUC = 0.86) 34 . In the same vein, Munk et al. focused on evaluating the performance of DL models in predicting patient age or sex using FPs and OCT scans. Their dataset comprised 135,667 FPs and 85,536 volumetric OCT scans. For sex prediction, the DL models achieved AUC values of 0.80 for FPs, 0.84 for OCT cross sections, and 0.90 for OCT volumes. In terms of age prediction, the input OCT volume models were better than OCT cross sections and better than FPs (MAE = 4.541 years, 5.625 years, and 6.328 years, respectively) 35 . These findings showed the varying predictive capabilities between FPs and OCT scans, where OCT seems to yield a better prediction of CV risk factors. Considering the favorable measurability and the wealth of retinal information offered by OCT, it is expected that OCT-A studies could expand the research on the potential of CV risk assessment. Initial work was undertaken to estimate the CV risk score (American Hospital Association [AHA] risk score, Syntax risk, and SCORE risk score) with ML models based on retinal vascular quantitative parameters measured with FPs and OCT-A scans through a multimodal approach 36 . Using OCT-A data, the K-nearest neighbor (KNN) and the naïve Bayes (NB) approaches more accurately predicted the three CV risk scores, with prediction rates ranging from 76.09 ± 3.08 to 96.13 ± 1.08 for KNN and 76.19 ± 5.30 to 96.23 ± 1.88 for NB. With FP-based vascular parameters, these two ML models also performed better than the others in CV risk assessment, with prediction rates ranging from 70.54 ± 8.56 to 95.83 ± 1.19 for KNN and 74.36 ± 6.17 to 96.28 ± 1.21 for NB. When combining both FP and OCT-A quantitative data, NB was the best fitted model, with an accuracy ranging from 75.64 ± 5.96 to 96.53 ± 1.25 36 . Concurrently, Zhong et al. investigated the prediction of coronary artery disease using a combination of clinical, electrocardiographic (ECG), and OCT-A data. The model trained on the combined clinical, ECG, and OCT-A data was presented as the individual prediction nomogram, exhibiting good discrimination (AUC = 0.897 [95% CI, 0.861–0.933]). Notably, the OCT-A model outperformed the ECG model in predicting individuals with coronary heart disease (AUC = 0.730 [95% CI, 0.673–0.788]) 37 . Research in CV risk assessment using SS OCT-A was held back by the lack of data. In fact, only a few datasets such as the OCTAGON and FOCTAIR datasets 38 , the Retinal OCTA SEgmentation dataset (ROSE) 39 , and OCTA-500 40 were publicly available. However, none of these datasets combined the OCT-A scans with the CV data of the patients included. We therefore chose to use the RASTA dataset, the first open-source dataset that combined clinical CV data and SS OCT-A scans. To the best of our knowledge, using this database enabled us to be the first research team to focus on predicting the neurocardiovascular risk category from SS OCT-A. This retinal imaging modality provides precise quantitative measurements of vascular density and blood perfusion, unlike FPs where assessment of vascular flow is not possible. Moreover, SS OCT-A has made it possible to explore the chorioretinal vasculature in much greater depth than PFs, enabling visualization of the anatomical vascular layers, i.e., the superficial and deep plexuses. The other major advantage is that this approach could monitor the evolution of lesions by allowing for repeated examinations, thereby assessing the effectiveness of treatments and the progression of the CVD. Although there were several ways to quantify the density of the retinal vasculature, the 15 quantitative datasets used to train our models could be considered a comprehensive representation of retinal microvasculature complexity because they included an analysis of the foveal avascular zone (FAZ), perfusion, and vascular density in a central area of 3 × 3 and 6 × 6 mm. In our study we found that the ML models performed better than the CNN model; however, our CNN model only had SS OCT-A scans as input, unlike the ML models that combined raw data of 11 CV risk factors. Therefore, we could have obtained better DL results in our study if we had included other CV risk factors in the algorithm. The models exhibited a consistent absolute prediction error (MAE = 0.697) and strong negative correlation ( R 2 = − 0.9446). A major limitation of our study was the small sample size compared to FP-based algorithms and consisted solely of European individuals. The generalizability of our models beyond the RASTA dataset requires further validation on larger external datasets with different ethnic groups. Second, our dataset was based on a specific SS OCT-A device (PLEX Elite 9000®, Carl Zeiss Meditec Inc., Dublin, OH, USA), which could limit our results with other manufacturers. Furthermore, longitudinal follow-up of neurocardiovascular events could strengthen our results. The ML and DL models described in this study accurately predicted the CHA 2 DS 2 -VASc score. The models were validated on a public dataset that registers patients with different CV risk factors. The models achieved good performance and, thanks to their SS OCT-A evaluation, may improve the management of patients referred to ophthalmologists. For the generalizability of our results, it is a priority to validate the models in future studies. Methods Neurocardiovascular risk profile We used the CHA 2 DS 2 -VAS C score (Table 3 ) as the score for neurocardiovascular risk assessment. It is an embolic risk stratification tool originally used to assess the risk of stroke in patients with non-valvular atrial fibrillation 14 . It has been recently presented as an effective model for patients without atrial fibrillation 23 – 29 . In contrast to other CV risk scores, it does not need any biological sampling. Table 3 CHA 2 DS 2 -VASc point-based scoring system Risk Factor Score C ongestive heart failure / Left ventricular dysfunction 1 H ypertension 1 A ge ≥ 75 years 2 D iabetes mellitus 1 S troke / TIA / TE 2 V ascular disease (prior myocardial infarction, peripheral artery disease, or aortic plaque) 1 A ge 65–74 years 1 S ex c ategory (i.e., female gender) 1 Dataset We previously published the Retinal oct-Angiography and cardiovascular STAtus (RASTA) dataset 18 , which was acquired from February 2018 to June 2023 in the Department of Ophthalmology at the University Hospital of Dijon, France. The RASTA dataset was anonymized and processed in accordance with the rules established by the Ethics Committee of the University Hospital of Dijon. The RASTA dataset is hosted and publicly available at https://rasta.u-bourgogne.fr/ . We assembled a cross-sectional set of retinal SS OCT-A images consisting of en face images and angiocubes combined with clinical and demographic characteristics from healthy and at-risk patients with complete clinical CV phenotypes. Information on data accessibility and specifications is provided in Table 4 . Each participant was included in one of two groups according to their neurocardiovascular risk category following the risk scheme used for RASTA (Table 5 ): Low neurocardiovascular risk – CHA 2 DS 2 -VASc = [0; 1] Intermediate–high neurocardiovascular risk – CHA 2 DS 2 -VASc = [2; 9] For each participant, corresponding images of the SS OCT-A 6 × 6-mm acquisitions were obtained with the PLEX Elite 9000® device (Carl Zeiss Meditec Inc., Dublin, OH, USA). En face images (Fig. 1 ) were based on the plexuses. Moreover, quantitative variables from the 2D SS OCT-A images were available and represented the different measurable characteristics of each patient's superficial and deep retinal vascular plexuses. These quantitative variables were calculated using the ARI Network segmentation and analysis platform developed by the manufacturer Carl Zeiss Meditec Inc. Table 4 Specifications table Subject Area Biomedical Imaging, Ophthalmology More specific subject area Retinal OCT-A volume analysis for cardiovascular risk prediction Type of data Image, CSV How data were acquired Swept-source OCT-A Instrument name: PLEX Elite 9000® (Carl Zeiss Meditec Inc., Dublin, OH, USA) Data format DICOM for volumes, Bitmap for en face images Experimental factors Pupillary dilatation with tropicamide 0.5% if signal strength < 8/10 Experimental features Macular angiography 6 × 6-mm Main data source location University Hospital of Dijon, Dijon 21000, France Data accessibility https://rasta.u-bourgogne.fr/ Table 5 Risk scheme used for neurocardiovascular risk stratification Risk scheme Low risk [0 ;1] Intermediate-high risk [2 ;9] RASTA (2023) One or no combination risk factor At least 1 definitive risk factor and 1 or no combination risk factor, or ≥ 2 combination risk factors Definitive risk factors: previous stroke/TIA/TE, age > 75 Combination risk factors: heart failure/left ventricular ejection fraction ≤ 40%, hypertension, diabetes, vascular disease, female sex, age 65–74 Machine-learning design To develop the models, we worked with raw data of 11 CV risk factors including age, gender, diabetes, smoking, and 15 quantitative variables obtained from en face SS OCT-A images (Fig. 2 and Table 6 ). These quantitative variables represent the different measurable characteristics of each patient's superficial and deep retinal vascular plexuses. We used all the quantitative variables of each eye as an independent feature, and thus we finally had 30 retinal quantitative data for each row when information for both eyes was available. These quantitative variables were calculated using the ARI Network segmentation and analysis platform developed by the manufacturer Carl Zeiss Meditec Inc. as described above. The neurocardiovascular risk category was the target. We built and evaluated four ML algorithms and compared their results: decision tree, random forest (RF), support vector machines (SVM), and logistic regression. Then, given the multivariate and multidimensional nature of the data, we conducted a principal component analysis using Pearson's correlation coefficient to eliminate the effect of scale in the data, moving from a space of 41 to 19 dimensions for a cumulative variability over 95% (Fig. 3 ). Table 6 Retinal SS OCT-A quantitative parameters used in ML models Vascular characteristics Features Superficial plexus Deep plexus Fovea avascular zone Raw length (mm) Circularity (index) Raw size (mm 2 ) Vessel density (mm − 1 ) Density average Density average Density in a circle of 3-mm diameter Density in a circle of 3-mm diameter Density in a circle of 6-mm diameter Density in a circle of 6-mm diameter Perfusion density (index) Perfusion average Perfusion average Perfusion in a circle of 3-mm diameter Perfusion in a circle of 3-mm diameter Perfusion in a circle of 6-mm diameter Perfusion in a circle of 6-mm diameter Machine-learning model assessment Due to the limited and imbalanced number of images in the RASTA dataset, a normal train–test split was not possible. Therefore, we employed the k-fold cross-validation method. We opted for a two-stage cross-validation procedure with k = 5 and then with k = 10 (Fig. 4 ). Deep-learning network design We employed cutting-edge DL methodologies to automate the assessment of neurocardiovascular risk using microvascular imaging. Our approach involved constructing a deep convolutional network with the EfficientNetV2 architecture 30 as the backbone, a convolutional neural network (CNN) architecture specifically designed for classification tasks. EfficientNetV2 employs a compound scaling strategy that simultaneously increased the model’s depth, width, and resolution. In our implementation, we used the EfficientNetV2-B3 backbone pre-trained on the ImageNet dataset. Input images were resized to a resolution of 300×300×3 pixels. The network architecture incorporated three dense layers, with a 20% dropout between each layer. The top two dense layers employed rectified linear unit (ReLU) activation, while the final layer utilized SoftMax activation, producing two outputs corresponding to the probabilities of different neurocardiovascular risk category. The network architecture is illustrated in Fig. 5 . Following the initial training phase, we performed fine-tuning by unfreezing the last 20 layers and retraining the model. Fine-tuning involved making subtle adjustments to the pre-trained model’s weights or parameters using data from the target task. Due to the limited number of images in one of the two categories in the RASTA dataset, we employed the stratified k-fold cross-validation with k = 5 for training a less biased model, where the ratio between the two target classes was the same in each fold as it was in the full dataset. Details on the network implementation are given in Table 7 . Table 7 Deep-learning network implementation details Model used EfficientNetB3-V2 Image resolution 300 Batch size 16 Epochs 100 K-folds 5 Learning rate 1.00E-03 Fine-tune epochs 50 Fine-tune learning rate 1.00E-03 Loss function Weighted Categorical Cross Entropy Early stopping True Early stopping patience 25 Labeling methods One-Hot Input scaling 0-255 Optimizer Adamv2 Activation function ReLU Cross-validation Stratified k-folds Train test splits 20 Pretrained weights ImageNet Dense layers 4 Pooling Average Dropout 20 Kernel regularizers L1_L2 Batch norm Yes LR scheduler Yes Input augmentation Flip, Brightness (25%), Saturation (25%) Ensemble voting Soft Deep-learning network implementation The entire network was implemented in TensorFlow 2.8 using the EfficientNetV2-B3 backbone from the Keras library. The reported metrics for each model represent the average performance across folds and are detailed in the Results section. Training involved 50 epochs with a batch size of 4, a learning rate set to 1e − 4 , and the application of weighted sparse categorical cross entropy (CCE) loss. Following the initial training phase, the last 20 layers were unfrozen, and the network underwent an additional 50 epochs of fine-tuning. Since en face SS OCT-A images are grayscale, they were converted to RGB format by triplicating the channel, adhering to EfficientNet’s requirement for three-channel images. Deep-learning model assessment For the evaluation, we used balanced accuracy as a metric for our imbalanced dataset, which is calculated as the arithmetic mean of sensitivity (true positive rate) and specificity (true negative rate). To assess the performance of the trained EfficientNetV2-B3 network, we performed a comparative analysis employing a random forest (RF) classifier based on the features extracted from the network. Predictions for each image were generated using a truncated version of the trained network, specifically capturing the features inferred at the last dense layer with a kernel size of 32. These features were used to train an RF classifier comprising 100 trees, and subsequently the test set was classified using this trained RF model. To ensure a fair evaluation, a second RF classifier was trained on the same dataset that was used for the DL network. Consequently, we obtained three distinct sets of inference results in the test set: one from the RF classifier, another from the RF classifier using only extracted features (RF-FO), and the third from the trained DL network. Declarations Acknowledgments We would like to thank Professor Ramin Tadayoni and Professor Aude Couturier for sharing data regarding the EVIRED study in Dijon. Author contributions All authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by G.C. Model development was performed by G.C. and A.A. The first draft of the manuscript was written by G.C. and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Data availability statement The RASTA dataset is hosted and publicly available at https://rasta.u-bourgogne.fr/. Competing interests Clément Germanèse has no relevant relationships to disclose. Fabrice Meriaudeau has no relevant relationships to disclose. Atif Anwer has no relevant relationships to disclose. Charles Guenancia has no relevant relationships to disclose. Catherine Creuzot-Garcher has no relevant relationships to disclose. Pierre-Henry Gabrielle has no relevant relationships to disclose. Louis Arnould has no relevant relationships to disclose. References WHO. World Health Organization reveals leading causes of death and disability worldwide: 2000-2019 , (2020). Goff, D. C., Jr. et al. 2013 ACC/AHA guideline on the assessment of cardiovascular risk: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines. Circulation 129 , S49-73, doi:10.1161/01.cir.0000437741.48606.98 (2014). Anderson, K. M., Wilson, P. W., Odell, P. M. & Kannel, W. B. An updated coronary risk profile. A statement for health professionals. Circulation 83 , 356-362, doi:10.1161/01.cir.83.1.356 (1991). group, S. w. & collaboration, E. C. r. SCORE2 risk prediction algorithms: new models to estimate 10-year risk of cardiovascular disease in Europe. European heart journal 42 , 2439-2454, doi:10.1093/eurheartj/ehab309 (2021). Lin, B. et al. Cumulative risk of stroke recurrence over the last 10 years: a systematic review and meta-analysis. Neurological sciences : official journal of the Italian Neurological Society and of the Italian Society of Clinical Neurophysiology 42 , 61-71, doi:10.1007/s10072-020-04797-5 (2021). Jokhadar, M., Jacobsen, S. J., Reeder, G. S., Weston, S. A. & Roger, V. L. Sudden death and recurrent ischemic events after myocardial infarction in the community. American journal of epidemiology 159 , 1040-1046, doi:10.1093/aje/kwh147 (2004). Arnould, L. et al. The EYE-MI Pilot Study: A Prospective Acute Coronary Syndrome Cohort Evaluated With Retinal Optical Coherence Tomography Angiography. Investigative Ophthalmology & Visual Science 59 , 4299-4306, doi:10.1167/iovs.18-24090 (2018). Spaide, R. F., Klancnik, J. M., Jr. & Cooney, M. J. Retinal vascular layers imaged by fluorescein angiography and optical coherence tomography angiography. JAMA ophthalmology 133 , 45-50, doi:10.1001/jamaophthalmol.2014.3616 (2015). Mariampillai, A. et al. Speckle variance detection of microvasculature using swept-source optical coherence tomography. Optics letters 33 , 1530-1532, doi:10.1364/ol.33.001530 (2008). Huang, Y. et al. Swept-source OCT angiography of the retinal vasculature using intensity differentiation-based optical microangiography algorithms. Ophthalmic surgery, lasers & imaging retina 45 , 382-389, doi:10.3928/23258160-20140909-08 (2014). Stanga, P. E. et al. Swept-Source Optical Coherence Tomography Angio™ (Topcon Corp, Japan): Technology Review. Developments in ophthalmology 56 , 13-17, doi:10.1159/000442771 (2016). Jia, Y. et al. Split-spectrum amplitude-decorrelation angiography with optical coherence tomography. Optics express 20 , 4710-4725, doi:10.1364/oe.20.004710 (2012). Laíns, I. et al. Retinal applications of swept source optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA). Prog Retin Eye Res 84 , 100951, doi:10.1016/j.preteyeres.2021.100951 (2021). Lip, G. Y., Nieuwlaat, R., Pisters, R., Lane, D. A. & Crijns, H. J. Refining clinical risk stratification for predicting stroke and thromboembolism in atrial fibrillation using a novel risk factor-based approach: the euro heart survey on atrial fibrillation. Chest 137 , 263-272, doi:10.1378/chest.09-1584 (2010). Wagner, S. K. et al. Insights into Systemic Disease through Retinal Imaging-Based Oculomics. Transl Vis Sci Technol 9 , 6, doi:10.1167/tvst.9.2.6 (2020). Arnould, L. et al. Using Artificial Intelligence to Analyse the Retinal Vascular Network: The Future of Cardiovascular Risk Assessment Based on Oculomics? A Narrative Review. Ophthalmol Ther 12 , 657-674, doi:10.1007/s40123-022-00641-5 (2023). Khan, S. M. et al. A global review of publicly available datasets for ophthalmological imaging: barriers to access, usability, and generalisability. Lancet Digit Health 3 , e51-e66, doi:10.1016/s2589-7500(20)30240-5 (2021). Germanèse, C. et al. A Retinal Oct-Angiography and Cardiovascular STAtus (RASTA) Dataset of Swept-Source Microvascular Imaging for Cardiovascular Risk Assessment. Data 8 , 147 (2023). Poplin, R. et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng 2 , 158-164, doi:10.1038/s41551-018-0195-0 (2018). Rudnicka, A. R. et al. Artificial intelligence-enabled retinal vasculometry for prediction of circulatory mortality, myocardial infarction and stroke. The British journal of ophthalmology 106 , 1722-1729, doi:10.1136/bjo-2022-321842 (2022). Sun, Z., Yang, D., Tang, Z., Ng, D. S. & Cheung, C. Y. Optical coherence tomography angiography in diabetic retinopathy: an updated review. Eye (London, England) 35 , 149-161, doi:10.1038/s41433-020-01233-y (2021). Monteiro-Henriques, I., Rocha-Sousa, A. & Barbosa-Breda, J. Optical coherence tomography angiography changes in cardiovascular systemic diseases and risk factors: A Review. Acta Ophthalmol 100 , e1-e15, doi:10.1111/aos.14851 (2022). Aker, A., Volis, I., Saliba, W., Naoum, I. & Zafrir, B. CHA2DS2-VASc Score as a Predictor of Adverse Outcomes after Ischemic Stroke in Patients without Atrial Fibrillation. The Israel Medical Association journal : IMAJ 26 , 143-148 (2024). Melgaard, L. et al. Assessment of the CHA2DS2-VASc Score in Predicting Ischemic Stroke, Thromboembolism, and Death in Patients With Heart Failure With and Without Atrial Fibrillation. Jama 314 , 1030-1038, doi:10.1001/jama.2015.10725 (2015). Su, C. H. et al. CHA2DS2-VASc score as an independent outcome predictor in patients hospitalized with acute ischemic stroke. PLoS One 17 , e0270823, doi:10.1371/journal.pone.0270823 (2022). Guerra, F. et al. CHA(2)DS(2)-VASc risk factors as predictors of stroke after acute coronary syndrome: A systematic review and meta-analysis. European heart journal. Acute cardiovascular care 7 , 264-274, doi:10.1177/2048872616673536 (2018). Tu, H. T. et al. Pre-stroke CHADS2 and CHA2DS2-VASc scores are useful in stratifying three-month outcomes in patients with and without atrial fibrillation. Cerebrovascular diseases (Basel, Switzerland) 36 , 273-280, doi:10.1159/000353670 (2013). Xing, Y. et al. CHA(2)DS(2)-VASc score as a predictor of long-term cardiac outcomes in elderly patients with or without atrial fibrillation. Clinical interventions in aging 13 , 497-504, doi:10.2147/cia.s147916 (2018). Akboğa, M. K., Yılmaz, S. & Yalçın, R. Prognostic value of CHA2DS2-VASc score in predicting high SYNTAX score and in-hospital mortality for non-ST elevation myocardial infarction in patients without atrial fibrillation. Anatolian journal of cardiology 25 , 789-795, doi:10.5152/AnatolJCardiol.2021.03982 (2021). Tan, M. & Le, Q. (2019). Decencière, E. et al. FEEDBACK ON A PUBLICLY DISTRIBUTED IMAGE DATABASE: THE MESSIDOR DATABASE. Image Analysis & Stereology , 231-234, doi:10.5566/ias.1155 (2014). Cheung, C. Y. et al. A deep-learning system for the assessment of cardiovascular disease risk via the measurement of retinal-vessel calibre. Nat Biomed Eng 5 , 498-508, doi:10.1038/s41551-020-00626-4 (2021). Zhang, K. et al. Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images. Nat Biomed Eng 5 , 533-545, doi:10.1038/s41551-021-00745-6 (2021). Hassan, O. et al. Deep learning prediction of age and sex from optical coherence tomography. IEEE 18th International Symposium on Biomedical Imaging (ISBI). Nice, France: IEEE , 238-242 (2021). Munk, M. R. et al. Assessment of patient specific information in the wild on fundus photography and optical coherence tomography. Scientific reports 11 , 8621, doi:10.1038/s41598-021-86577-5 (2021). Arnould, L. et al. Prediction of Cardiovascular Parameters With Supervised Machine Learning From Singapore "I" Vessel Assessment and OCT-Angiography: A Pilot Study. Transl Vis Sci Technol 10 , 20, doi:10.1167/tvst.10.13.20 (2021). Zhong, P. et al. Development and Validation of Retinal Vasculature Nomogram in Suspected Angina Due to Coronary Artery Disease. Journal of atherosclerosis and thrombosis 29 , 579-596, doi:10.5551/jat.62059 (2022). VARPA Working fields: Public databases , ( Ma, Y. et al. ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model. IEEE transactions on medical imaging 40 , 928-939, doi:10.1109/tmi.2020.3042802 (2021). Li, M. et al. OCTA-500: A retinal dataset for optical coherence tomography angiography study. Medical image analysis 93 , 103092, doi:10.1016/j.media.2024.103092 (2024). Additional Declarations No competing interests reported. Supplementary Files SSOCTAsupplementaryappendix.docx Cite Share Download PDF Status: Published Journal Publication published 07 Nov, 2024 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 24 Jun, 2024 Reviews received at journal 06 Jun, 2024 Reviews received at journal 31 May, 2024 Reviewers agreed at journal 21 May, 2024 Reviewers agreed at journal 20 May, 2024 Reviewers invited by journal 19 May, 2024 Editor assigned by journal 15 May, 2024 Editor invited by journal 28 Apr, 2024 Submission checks completed at journal 28 Apr, 2024 First submitted to journal 25 Apr, 2024 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4326028","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":297659659,"identity":"3913a07e-f353-4d00-831a-8f562da1f6e7","order_by":0,"name":"Clement Germanese","email":"","orcid":"","institution":"Department of Ophthalmology, Dijon University Hospital, Dijon, France","correspondingAuthor":false,"prefix":"","firstName":"Clement","middleName":"","lastName":"Germanese","suffix":""},{"id":297659660,"identity":"645d94d0-26a0-4bb9-8614-d6b542ce6d95","order_by":1,"name":"Atif Anwer","email":"","orcid":"","institution":"Institut de Chimie Moléculaire Université de Bourgogne (ICMUB), Imagerie Fonctionnelle et moléculaire et Traitement des Images Médicales (IFTIM), Burgundy University","correspondingAuthor":false,"prefix":"","firstName":"Atif","middleName":"","lastName":"Anwer","suffix":""},{"id":297659661,"identity":"3b4c4bb3-9adf-4cbd-ae9b-283e6e52b033","order_by":2,"name":"Petra Eid","email":"","orcid":"","institution":"Department of Ophthalmology, Dijon University Hospital, Dijon, France","correspondingAuthor":false,"prefix":"","firstName":"Petra","middleName":"","lastName":"Eid","suffix":""},{"id":297659662,"identity":"131a8968-1b83-471a-bf12-e319bd66069a","order_by":3,"name":"Laure-Anne Steinberg","email":"","orcid":"","institution":"Department of Ophthalmology, Dijon University Hospital, Dijon, France","correspondingAuthor":false,"prefix":"","firstName":"Laure-Anne","middleName":"","lastName":"Steinberg","suffix":""},{"id":297659663,"identity":"b24a3ff2-74ea-4865-82ea-9c1aa47ab89e","order_by":4,"name":"Charles Guenancia","email":"","orcid":"","institution":"Department of Cardiology, Dijon University Hospital, Dijon, France","correspondingAuthor":false,"prefix":"","firstName":"Charles","middleName":"","lastName":"Guenancia","suffix":""},{"id":297659664,"identity":"0982e1ee-46d3-4fe3-ab0b-058d6a284fbc","order_by":5,"name":"Pierre-Henry Gabrielle","email":"","orcid":"","institution":"Department of Ophthalmology, Dijon University Hospital, Dijon, France","correspondingAuthor":false,"prefix":"","firstName":"Pierre-Henry","middleName":"","lastName":"Gabrielle","suffix":""},{"id":297659665,"identity":"bbf1cb66-0220-411f-8bfb-86cc34295a49","order_by":6,"name":"Catherine Creuzot-Garcher","email":"","orcid":"","institution":"Department of Ophthalmology, Dijon University Hospital, Dijon, France","correspondingAuthor":false,"prefix":"","firstName":"Catherine","middleName":"","lastName":"Creuzot-Garcher","suffix":""},{"id":297659666,"identity":"71b2f8a9-3319-4775-a00a-f3305d70c926","order_by":7,"name":"Fabrice Meriaudeau","email":"","orcid":"","institution":"Institut de Chimie Moléculaire Université de Bourgogne (ICMUB), Imagerie Fonctionnelle et moléculaire et Traitement des Images Médicales (IFTIM), Burgundy University","correspondingAuthor":false,"prefix":"","firstName":"Fabrice","middleName":"","lastName":"Meriaudeau","suffix":""},{"id":297659667,"identity":"2ed943d8-cf3b-4a3c-8b9e-777f173a5bca","order_by":8,"name":"Louis Arnould","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBklEQVRIie3RMWvCQBTA8SeBZHnRNaXgZ3idrEv8KoaspQa6BFqoIDgF94L6GRRB6PbCQVzuA3QQdHJKoaWLY48miLQ5abci94cL7yA/uOMATKZ/WF0tLucaYwxdNRB76uv0q4l9RIBRHhFkPTnE7rAgXzstcZKrdA9+r+UkxK9T/5bYmvN1vAa8rDY2ShIIYfs5kZROluEdsR2xJ3eA9W418W5IAFhEL2pwlxzMNjnxxVBABzUHU0Qd7LEkY0UYC4InCCOIkvR/QzCLBNKKSGZROs7C4Km4i0AdaTiDxfs+vidaDebb/MEPRmwtPrxYNHWkiL7tLfWUJ8HPam9/+99kMpnOu09z2l/zVLGHuAAAAABJRU5ErkJggg==","orcid":"","institution":"Department of Ophthalmology, Dijon University Hospital, Dijon, France","correspondingAuthor":true,"prefix":"","firstName":"Louis","middleName":"","lastName":"Arnould","suffix":""}],"badges":[],"createdAt":"2024-04-25 20:23:55","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4326028/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4326028/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-024-78587-w","type":"published","date":"2024-11-07T15:57:38+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":55762394,"identity":"37c65721-7580-4899-904f-423b78cdc3f8","added_by":"auto","created_at":"2024-05-02 19:11:20","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":643873,"visible":true,"origin":"","legend":"\u003cp\u003eRight eye\u003cstrong\u003e \u003c/strong\u003een face images of (\u003cstrong\u003ea\u003c/strong\u003e) superficial plexus, (\u003cstrong\u003eb\u003c/strong\u003e) deep plexus, and (\u003cstrong\u003ec\u003c/strong\u003e) choriocapillaris plexus\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4326028/v1/3bf4360497838353028a3d37.jpg"},{"id":55761696,"identity":"3e6b5d6d-bfa2-4ffa-ba03-55cf1d4c95ea","added_by":"auto","created_at":"2024-05-02 19:03:20","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":311852,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic illustration of machine-learning model design\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea\u003c/strong\u003e Angio-retina layer image shows the fovea avascular zone (FAZ) outline with 3 associated measures: area of the FAZ in mm\u003csup\u003e2\u003c/sup\u003e, perimeter of the FAZ in mm, and circularity index of the FAZ (ranging from 0 to 1). \u003cstrong\u003eb \u003c/strong\u003eand \u003cstrong\u003ed\u003c/strong\u003e show vessel density, which is defined as the total length of perfused vasculature per unit area in a region of measurement in mm\u003csup\u003e-1\u003c/sup\u003e. \u003cstrong\u003ec\u003c/strong\u003e and \u003cstrong\u003ee\u003c/strong\u003e show perfusion density, which is defined as the total area of perfused vasculature per unit area in a region of measurement (ranging from 0 to 1).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eNeurocardiovascular risk category\u003c/strong\u003e is the target of machine-learning models, it should not be considered as a feature.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4326028/v1/c39761388b12fa9b106f2131.jpg"},{"id":55761690,"identity":"d69979af-93ee-4ad1-b1fa-4bf279070553","added_by":"auto","created_at":"2024-05-02 19:03:18","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":221829,"visible":true,"origin":"","legend":"\u003cp\u003eScree plot\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4326028/v1/69c75f984075b5ba5773d9fb.jpg"},{"id":55761691,"identity":"fb3227f8-a2db-4ce7-9506-7888eee17732","added_by":"auto","created_at":"2024-05-02 19:03:18","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":166297,"visible":true,"origin":"","legend":"\u003cp\u003eK-fold cross-validation\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4326028/v1/d0fa4f2e3c2c2818e722dcbc.jpg"},{"id":55761694,"identity":"2ac01dd0-0ad2-4bd3-80bd-06eb20e4474f","added_by":"auto","created_at":"2024-05-02 19:03:20","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":353251,"visible":true,"origin":"","legend":"\u003cp\u003eDeep-learning network based on EfficientNetv2-B3 backbone\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4326028/v1/498e1ac8122775a6d7b39b7c.jpg"},{"id":68749908,"identity":"a2b4146b-42c8-4066-8ce7-2b5d08257c8b","added_by":"auto","created_at":"2024-11-11 16:07:31","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2441653,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4326028/v1/e06c0c5d-716b-4cbc-b25b-30aae025b731.pdf"},{"id":55761692,"identity":"7cba5021-c9cf-432b-b92f-176a29462f70","added_by":"auto","created_at":"2024-05-02 19:03:19","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":715216,"visible":true,"origin":"","legend":"","description":"","filename":"SSOCTAsupplementaryappendix.docx","url":"https://assets-eu.researchsquare.com/files/rs-4326028/v1/0c3cf709a8f09252d32c8971.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Artificial intelligence-based prediction of neurocardiovascular risk score from retinal swept-source microvascular imaging: the RASTA dataset","fulltext":[{"header":"Introduction","content":"\u003cp\u003eWith approximately 17.9\u0026nbsp;million deaths every year\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e, cardiovascular diseases (CVD) are the leading cause of death worldwide, and the assessment of cardiovascular (CV) risk plays a major role in their prevention. Today, CV risk profiles can be estimated with numerous validated score models such as the Framingham Risk Score (FRS), the American Heart Association risk score (AHA risk score), or the SCORE2\u003csup\u003e2\u0026ndash;4\u003c/sup\u003e. Nevertheless, these clinical scores do not take into account personalized vascular conditions, and recurrent neurovascular and cardiovascular events are still difficult to predict\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. We previously demonstrated through a conventional regression approach in post-myocardial infarction patients in the EYE-MI study that retinal optical coherence tomography\u0026ndash;angiography (OCT-A) could be an effective tool for predicting cardiovascular risk scores\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Swept-source OCT\u0026ndash;angiography (SS OCT-A) technology, with more thorough retinal microvasculature description and better resolution, could be used to refine the EYE-MI pilot study results in patients with broader phenotypes\u003csup\u003e\u003cspan additionalcitationids=\"CR9 CR10 CR11 CR12\" citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Moreover, the CHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VASc score, a cardioembolic risk score\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e, may better reflect the connection between the retinal, neurovascular, and cardiovascular systems. The role of advanced imaging analysis is becoming increasingly important in daily clinical practice and biomedical research due to the recent advancements in oculomics\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e and other artificial intelligence (AI)-based image analysis methods\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. In the framework of open-source datasets in ophthalmology\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e, we previously published the Retinal oct-Angiography and cardiovascular STAtus (RASTA) dataset\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. It combined SS OCT-A retinal imaging and CHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VASc score calculation in patients with various conditions (e.g., diabetes mellitus, giant cell arthritis, dyslipidemia, and ischemic CVD) and accurate phenotype. Recent studies using deep-learning (DL) convolutional neural networks (CNNs) have associated retinal photographs with several CVD risk factors, including diabetes mellitus, blood pressure, body mass index (BMI), smoking, and glycated-hemoglobin level\u003csup\u003e\u003cspan additionalcitationids=\"CR20 CR21\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. Nevertheless, the literature is sparse regarding the interest of SS OCT-A for artificial intelligence-based prediction of the neurocardiovascular risk score.\u003c/p\u003e \u003cp\u003eThe purpose of this study was to train and validate machine-learning- (ML) and DL-based models capable of predicting neurocardiovascular risk scores (CHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VASc score) using the detection method from SS OCT-A acquisitions in the RASTA dataset.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eDataset\u003c/h2\u003e \u003cp\u003eOverall, 491 patients were registered in this study. Of the 491 patients, 225 (45.8%) patients had a low neurocardiovascular risk and 266 (54.2%) had an intermediate\u0026ndash;high neurocardiovascular risk. The mean age of the patients was 52.4\u0026thinsp;\u0026plusmn;\u0026thinsp;18.4 years and 51.5% were women. All the cardiovascular clinical data and quantitative SS OCT-A variables were significantly different between the two neurocardiovascular risk categories. A comparison of the clinical and demographic variables is available in the supplementary material (\u003cb\u003esupplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eMachine-learning model results\u003c/h2\u003e \u003cp\u003eThe SVM model performed better than the others in predicting neurocardiovascular risk categories (AUC 0.98\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03 versus 0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02 and 0.91\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04 and 0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12 for logistic regression, RF, and decision tree, respectively). The results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMachine-learning model results\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eDecision tree\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c5\" namest=\"c4\"\u003e \u003cp\u003eRandom forest\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c7\" namest=\"c6\"\u003e \u003cp\u003eSVM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c9\" namest=\"c8\"\u003e \u003cp\u003eLogistic regression\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003eAccuracy (%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ek folds (k\u0026thinsp;=\u0026thinsp;10)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.78\u0026thinsp;\u0026plusmn;\u0026thinsp;0.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e77.2\u0026thinsp;\u0026plusmn;\u0026thinsp;3.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.91\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e81.2\u0026thinsp;\u0026plusmn;\u0026thinsp;2.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.98\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e85.1\u0026thinsp;\u0026plusmn;\u0026thinsp;5.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e84.9\u0026thinsp;\u0026plusmn;\u0026thinsp;5.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ek folds (k\u0026thinsp;=\u0026thinsp;5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.77\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e75.8\u0026thinsp;\u0026plusmn;\u0026thinsp;3.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e80.8\u0026thinsp;\u0026plusmn;\u0026thinsp;2.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.98\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e84.2\u0026thinsp;\u0026plusmn;\u0026thinsp;5.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.96\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e84.4\u0026thinsp;\u0026plusmn;\u0026thinsp;5.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"9\" nameend=\"c9\" namest=\"c1\"\u003e \u003cp\u003eSVM: support vector machines. AUC: area under the curve. Descriptive results for models are presented as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eDeep-learning model results\u003c/h2\u003e \u003cp\u003eFor classifying the two categories, the network was able to achieve a balanced accuracy of 68% as compared to 61% and 54% for RF and RF-FO, respectively. In evaluating the performance of the networks on the test set, the EfficientNet model and both variants of the RF model (RF and RF-FO) exhibited similar results. All the models demonstrated the same mean absolute error (MAE) of approximately 0.697. The \u003cem\u003eR\u003c/em\u003e\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e scores for the two models were identical at \u0026minus;\u0026thinsp;0.9446. In terms of balanced accuracy, the EfficientNet model outperformed the RF models, achieving a balanced accuracy of 39%, compared to 33% for both RF variants. Main results are shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The learning curves are presented in the supplementary material (\u003cb\u003esupplementary Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e\u003c/b\u003e). Classification results details are shown in the supplementary material (\u003cb\u003esupplementary Figure S2\u003c/b\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePerformance of deep-learning-based model\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eEfficientNetB3-V2\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRandom forest\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRandom forest features only\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBalanced accuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eR\u003c/em\u003e\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e score\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-0.3444\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-0.7111\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e-0.9556\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we found good results for neurocardiovascular risk category prediction based on retinal SS OCT-A, with a strong predictive accuracy of up to 98% for ML models and 68% for the EfficientNetV2-B3 backbone model. To our knowledge, this is the first study to investigate artificial intelligence-based prediction of neurocardiovascular risk score with SS OCT-A. Our results highlight the potential value of SS OCT-A as a biomarker of global neurocardiovascular status.\u003c/p\u003e \u003cp\u003eFundus photography (FP) was the first imaging modality in ophthalmology to prove its value in automatic CV risk assessment. Most hospital departments and practitioners have FP equipment, which has led to the creation of many rich databases such as the UK Biobank and MESSIDOR\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. These databases, containing hundreds or thousands of fundus images, have been used in several studies to assess and predict CVD. Our results on artificial intelligence-based prediction of neurocardiovascular risk score with SS OCT-A were in line with previous studies based on FP. Poplin et al. used FPs to demonstrate the contribution of retinal vasculature to the automatic detection of CVD and CVD risk factors, where AUC values of DL models were greater than 0.70 and demonstrated their effectiveness in predicting some CVD risk factor and the occurrence of major adverse cardiovascular events (MACE) over a 5-year period\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Cheung et al. published their work on the assessment of CVD risk via automatic measurement of retinal-vessel caliber (RVC)\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. They developed and tested a DL model to specifically measure RVC from more than 70,000 FPs. They assessed the agreement of the RVC measurement between the DL model and a human expert. The DL models predicted CVD risk factors significantly better than the human-based models or were at least comparable\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. More recently, Zhang et al. demonstrated the capability of DL models to identify chronic kidney disease and diabetes mellitus using 115,344 FPs alone or in combination with clinical metadata (i.e., age, sex, BMI, and blood pressure), with AUCs ranging from 0.85 to 0.93. Additionally, the models could predict glomerular filtration rates and blood glucose levels, yielding MAEs of 11.1\u0026ndash;13.4 mL/min per 1.73 m\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e and 0.65\u0026ndash;1.1 mmol/L, respectively\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eHowever, the exploration of microvasculature at the micrometer level in various plexuses and vascular networks made possible by SS OCT-A could offer hope for an even more accurate and earlier assessment of CV risk compared to FP. Based on previous ophthalmological research, Hassan et al. conducted an evaluation using three-dimensional CNNs to predict the individual age and sex directly from 3D retinal OCT scans, using a large dataset comprising 66,767 participants from the UK Biobank dataset. Model results showed accurate predictions for age (MAE\u0026thinsp;=\u0026thinsp;3.30 years, \u003cem\u003eR\u003c/em\u003e\u0026sup2;=0.89) and for sex (AUC\u0026thinsp;=\u0026thinsp;0.86)\u003csup\u003e34\u003c/sup\u003e. In the same vein, Munk et al. focused on evaluating the performance of DL models in predicting patient age or sex using FPs and OCT scans. Their dataset comprised 135,667 FPs and 85,536 volumetric OCT scans. For sex prediction, the DL models achieved AUC values of 0.80 for FPs, 0.84 for OCT cross sections, and 0.90 for OCT volumes. In terms of age prediction, the input OCT volume models were better than OCT cross sections and better than FPs (MAE\u0026thinsp;=\u0026thinsp;4.541 years, 5.625 years, and 6.328 years, respectively)\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. These findings showed the varying predictive capabilities between FPs and OCT scans, where OCT seems to yield a better prediction of CV risk factors. Considering the favorable measurability and the wealth of retinal information offered by OCT, it is expected that OCT-A studies could expand the research on the potential of CV risk assessment. Initial work was undertaken to estimate the CV risk score (American Hospital Association [AHA] risk score, Syntax risk, and SCORE risk score) with ML models based on retinal vascular quantitative parameters measured with FPs and OCT-A scans through a multimodal approach\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e. Using OCT-A data, the K-nearest neighbor (KNN) and the na\u0026iuml;ve Bayes (NB) approaches more accurately predicted the three CV risk scores, with prediction rates ranging from 76.09\u0026thinsp;\u0026plusmn;\u0026thinsp;3.08 to 96.13\u0026thinsp;\u0026plusmn;\u0026thinsp;1.08 for KNN and 76.19\u0026thinsp;\u0026plusmn;\u0026thinsp;5.30 to 96.23\u0026thinsp;\u0026plusmn;\u0026thinsp;1.88 for NB. With FP-based vascular parameters, these two ML models also performed better than the others in CV risk assessment, with prediction rates ranging from 70.54\u0026thinsp;\u0026plusmn;\u0026thinsp;8.56 to 95.83\u0026thinsp;\u0026plusmn;\u0026thinsp;1.19 for KNN and 74.36\u0026thinsp;\u0026plusmn;\u0026thinsp;6.17 to 96.28\u0026thinsp;\u0026plusmn;\u0026thinsp;1.21 for NB. When combining both FP and OCT-A quantitative data, NB was the best fitted model, with an accuracy ranging from 75.64\u0026thinsp;\u0026plusmn;\u0026thinsp;5.96 to 96.53\u0026thinsp;\u0026plusmn;\u0026thinsp;1.25\u003csup\u003e36\u003c/sup\u003e. Concurrently, Zhong et al. investigated the prediction of coronary artery disease using a combination of clinical, electrocardiographic (ECG), and OCT-A data. The model trained on the combined clinical, ECG, and OCT-A data was presented as the individual prediction nomogram, exhibiting good discrimination (AUC\u0026thinsp;=\u0026thinsp;0.897 [95% CI, 0.861\u0026ndash;0.933]). Notably, the OCT-A model outperformed the ECG model in predicting individuals with coronary heart disease (AUC\u0026thinsp;=\u0026thinsp;0.730 [95% CI, 0.673\u0026ndash;0.788])\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eResearch in CV risk assessment using SS OCT-A was held back by the lack of data. In fact, only a few datasets such as the OCTAGON and FOCTAIR datasets\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e, the Retinal OCTA SEgmentation dataset (ROSE)\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e, and OCTA-500\u003csup\u003e40\u003c/sup\u003e were publicly available. However, none of these datasets combined the OCT-A scans with the CV data of the patients included. We therefore chose to use the RASTA dataset, the first open-source dataset that combined clinical CV data and SS OCT-A scans. To the best of our knowledge, using this database enabled us to be the first research team to focus on predicting the neurocardiovascular risk category from SS OCT-A. This retinal imaging modality provides precise quantitative measurements of vascular density and blood perfusion, unlike FPs where assessment of vascular flow is not possible. Moreover, SS OCT-A has made it possible to explore the chorioretinal vasculature in much greater depth than PFs, enabling visualization of the anatomical vascular layers, i.e., the superficial and deep plexuses. The other major advantage is that this approach could monitor the evolution of lesions by allowing for repeated examinations, thereby assessing the effectiveness of treatments and the progression of the CVD. Although there were several ways to quantify the density of the retinal vasculature, the 15 quantitative datasets used to train our models could be considered a comprehensive representation of retinal microvasculature complexity because they included an analysis of the foveal avascular zone (FAZ), perfusion, and vascular density in a central area of 3 \u0026times; 3 and 6 \u0026times; 6 mm. In our study we found that the ML models performed better than the CNN model; however, our CNN model only had SS OCT-A scans as input, unlike the ML models that combined raw data of 11 CV risk factors. Therefore, we could have obtained better DL results in our study if we had included other CV risk factors in the algorithm.\u003c/p\u003e \u003cp\u003eThe models exhibited a consistent absolute prediction error (MAE\u0026thinsp;=\u0026thinsp;0.697) and strong negative correlation (\u003cem\u003eR\u003c/em\u003e\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;\u0026minus;\u0026thinsp;0.9446).\u003c/p\u003e \u003cp\u003eA major limitation of our study was the small sample size compared to FP-based algorithms and consisted solely of European individuals. The generalizability of our models beyond the RASTA dataset requires further validation on larger external datasets with different ethnic groups. Second, our dataset was based on a specific SS OCT-A device (PLEX Elite 9000\u0026reg;, Carl Zeiss Meditec Inc., Dublin, OH, USA), which could limit our results with other manufacturers. Furthermore, longitudinal follow-up of neurocardiovascular events could strengthen our results.\u003c/p\u003e \u003cp\u003eThe ML and DL models described in this study accurately predicted the CHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VASc score. The models were validated on a public dataset that registers patients with different CV risk factors. The models achieved good performance and, thanks to their SS OCT-A evaluation, may improve the management of patients referred to ophthalmologists. For the generalizability of our results, it is a priority to validate the models in future studies.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\n\u003ch2\u003eNeurocardiovascular risk profile\u003c/h2\u003e\n\u003cp\u003eWe used the CHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VAS\u003csub\u003eC\u003c/sub\u003e score (Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e) as the score for neurocardiovascular risk assessment. It is an embolic risk stratification tool originally used to assess the risk of stroke in patients with non-valvular atrial fibrillation\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. It has been recently presented as an effective model for patients without atrial fibrillation\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. In contrast to other CV risk scores, it does not need any biological sampling.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab3\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eCHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VASc point-based scoring system\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eRisk Factor\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eScore\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cspan class=\"BoldUnderline\"\u003eC\u003c/span\u003eongestive heart failure / Left ventricular dysfunction\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cspan class=\"BoldUnderline\"\u003eH\u003c/span\u003eypertension\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cspan class=\"BoldUnderline\"\u003eA\u003c/span\u003ege\u0026thinsp;\u0026ge;\u0026thinsp;75 years\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e2\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cspan class=\"BoldUnderline\"\u003eD\u003c/span\u003eiabetes mellitus\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cspan class=\"BoldUnderline\"\u003eS\u003c/span\u003etroke / TIA / TE\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e2\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cspan class=\"BoldUnderline\"\u003eV\u003c/span\u003eascular disease (prior myocardial infarction, peripheral artery disease, or aortic plaque)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cspan class=\"BoldUnderline\"\u003eA\u003c/span\u003ege 65\u0026ndash;74 years\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cspan class=\"BoldUnderline\"\u003eS\u003c/span\u003eex \u003cspan class=\"BoldUnderline\"\u003ec\u003c/span\u003eategory (i.e., female gender)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e1\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\n\u003ch2\u003eDataset\u003c/h2\u003e\n\u003cp\u003eWe previously published the Retinal oct-Angiography and cardiovascular STAtus (RASTA) dataset\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e, which was acquired from February 2018 to June 2023 in the Department of Ophthalmology at the University Hospital of Dijon, France. The RASTA dataset was anonymized and processed in accordance with the rules established by the Ethics Committee of the University Hospital of Dijon. The RASTA dataset is hosted and publicly available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://rasta.u-bourgogne.fr/\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e\n\u003cp\u003eWe assembled a cross-sectional set of retinal SS OCT-A images consisting of en face images and angiocubes combined with clinical and demographic characteristics from healthy and at-risk patients with complete clinical CV phenotypes.\u003c/p\u003e\n\u003cp\u003eInformation on data accessibility and specifications is provided in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e. Each participant was included in one of two groups according to their neurocardiovascular risk category following the risk scheme used for RASTA (Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e):\u003c/p\u003e\n\u003cp\u003eLow neurocardiovascular risk \u0026ndash; CHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VASc = [0; 1]\u003c/p\u003e\n\u003cdiv class=\"BlockQuote\"\u003e\n\u003cp\u003eIntermediate\u0026ndash;high neurocardiovascular risk \u0026ndash; CHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VASc = [2; 9]\u003c/p\u003e\n\u003c/div\u003e\n\u003cp\u003eFor each participant, corresponding images of the SS OCT-A 6 \u0026times; 6-mm acquisitions were obtained with the PLEX Elite 9000\u0026reg; device (Carl Zeiss Meditec Inc., Dublin, OH, USA). En face images (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e) were based on the plexuses. Moreover, quantitative variables from the 2D SS OCT-A images were available and represented the different measurable characteristics of each patient's superficial and deep retinal vascular plexuses. These quantitative variables were calculated using the ARI Network segmentation and analysis platform developed by the manufacturer Carl Zeiss Meditec Inc.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab4\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eSpecifications table\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eSubject Area\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eBiomedical Imaging, Ophthalmology\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMore specific subject area\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eRetinal OCT-A volume analysis for cardiovascular risk prediction\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eType of data\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eImage, CSV\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eHow data were acquired\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eSwept-source OCT-A\u003c/p\u003e\n\u003cp\u003eInstrument name: PLEX Elite 9000\u0026reg; (Carl Zeiss Meditec Inc., Dublin, OH, USA)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eData format\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDICOM for volumes, Bitmap for en face images\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eExperimental factors\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePupillary dilatation with tropicamide 0.5% if signal strength\u0026thinsp;\u0026lt;\u0026thinsp;8/10\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eExperimental features\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMacular angiography 6 \u0026times; 6-mm\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eMain data source location\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eUniversity Hospital of Dijon, Dijon 21000, France\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eData accessibility\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://rasta.u-bourgogne.fr/\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003cdiv class=\"colspec\" align=\"left\"\u003e\u0026nbsp;\u003c/div\u003e\n\u003ctable id=\"Tab5\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eRisk scheme used for neurocardiovascular risk stratification\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eRisk scheme\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eLow risk [0\u0026nbsp;;1]\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eIntermediate-high risk [2\u0026nbsp;;9]\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eRASTA (2023)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eOne or no combination risk factor\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eAt least 1 definitive risk factor and 1 or no combination risk factor, or \u0026ge;\u0026thinsp;2 combination risk factors\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd colspan=\"3\" align=\"left\"\u003e\n\u003cp\u003eDefinitive risk factors: previous stroke/TIA/TE, age\u0026thinsp;\u0026gt;\u0026thinsp;75\u003c/p\u003e\n\u003cp\u003eCombination risk factors: heart failure/left ventricular ejection fraction\u0026thinsp;\u0026le;\u0026thinsp;40%, hypertension, diabetes, vascular disease, female sex, age 65\u0026ndash;74\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\n\u003ch2\u003eMachine-learning design\u003c/h2\u003e\n\u003cp\u003eTo develop the models, we worked with raw data of 11 CV risk factors including age, gender, diabetes, smoking, and 15 quantitative variables obtained from en face SS OCT-A images (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e and Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e6\u003c/span\u003e). These quantitative variables represent the different measurable characteristics of each patient's superficial and deep retinal vascular plexuses. We used all the quantitative variables of each eye as an independent feature, and thus we finally had 30 retinal quantitative data for each row when information for both eyes was available. These quantitative variables were calculated using the ARI Network segmentation and analysis platform developed by the manufacturer Carl Zeiss Meditec Inc. as described above. The neurocardiovascular risk category was the target.\u003c/p\u003e\n\u003cp\u003eWe built and evaluated four ML algorithms and compared their results: decision tree, random forest (RF), support vector machines (SVM), and logistic regression. Then, given the multivariate and multidimensional nature of the data, we conducted a principal component analysis using Pearson's correlation coefficient to eliminate the effect of scale in the data, moving from a space of 41 to 19 dimensions for a cumulative variability over 95% (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab6\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 6\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eRetinal SS OCT-A quantitative parameters used in ML models\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eVascular characteristics\u003c/p\u003e\n\u003c/th\u003e\n\u003cth colspan=\"2\" align=\"left\"\u003e\n\u003cp\u003eFeatures\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eSuperficial plexus\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eDeep plexus\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eFovea avascular zone\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd colspan=\"2\" align=\"left\"\u003e\n\u003cp\u003eRaw length (mm)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd colspan=\"2\" align=\"left\"\u003e\n\u003cp\u003eCircularity (index)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd colspan=\"2\" align=\"left\"\u003e\n\u003cp\u003eRaw size (mm\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eVessel density (mm\u003csup\u003e\u0026minus;\u0026thinsp;1\u003c/sup\u003e)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDensity average\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDensity average\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDensity in a circle of 3-mm diameter\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDensity in a circle of 3-mm diameter\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDensity in a circle of 6-mm diameter\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDensity in a circle of 6-mm diameter\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePerfusion density\u003c/p\u003e\n\u003cp\u003e(index)\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePerfusion average\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePerfusion average\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePerfusion in a circle of 3-mm diameter\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePerfusion in a circle of 3-mm diameter\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePerfusion in a circle of 6-mm diameter\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePerfusion in a circle of 6-mm diameter\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\n\u003ch2\u003eMachine-learning model assessment\u003c/h2\u003e\n\u003cp\u003eDue to the limited and imbalanced number of images in the RASTA dataset, a normal train\u0026ndash;test split was not possible. Therefore, we employed the k-fold cross-validation method. We opted for a two-stage cross-validation procedure with k\u0026thinsp;=\u0026thinsp;5 and then with k\u0026thinsp;=\u0026thinsp;10 (Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\n\u003ch2\u003eDeep-learning network design\u003c/h2\u003e\n\u003cp\u003eWe employed cutting-edge DL methodologies to automate the assessment of neurocardiovascular risk using microvascular imaging. Our approach involved constructing a deep convolutional network with the EfficientNetV2 architecture\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e as the backbone, a convolutional neural network (CNN) architecture specifically designed for classification tasks. EfficientNetV2 employs a compound scaling strategy that simultaneously increased the model\u0026rsquo;s depth, width, and resolution. In our implementation, we used the EfficientNetV2-B3 backbone pre-trained on the ImageNet dataset. Input images were resized to a resolution of 300\u0026times;300\u0026times;3 pixels. The network architecture incorporated three dense layers, with a 20% dropout between each layer. The top two dense layers employed rectified linear unit (ReLU) activation, while the final layer utilized SoftMax activation, producing two outputs corresponding to the probabilities of different neurocardiovascular risk category. The network architecture is illustrated in Fig.\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e. Following the initial training phase, we performed fine-tuning by unfreezing the last 20 layers and retraining the model. Fine-tuning involved making subtle adjustments to the pre-trained model\u0026rsquo;s weights or parameters using data from the target task. Due to the limited number of images in one of the two categories in the RASTA dataset, we employed the stratified k-fold cross-validation with k\u0026thinsp;=\u0026thinsp;5 for training a less biased model, where the ratio between the two target classes was the same in each fold as it was in the full dataset. Details on the network implementation are given in Table\u0026nbsp;\u003cspan class=\"InternalRef\"\u003e7\u003c/span\u003e.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n\u003ctable id=\"Tab7\" border=\"1\"\u003e\u003ccaption\u003e\n\u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e\n\u003cdiv class=\"CaptionContent\"\u003e\n\u003cp\u003eDeep-learning network implementation details\u003c/p\u003e\n\u003c/div\u003e\n\u003c/caption\u003e\n\u003cthead\u003e\n\u003ctr\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eModel used\u003c/p\u003e\n\u003c/th\u003e\n\u003cth align=\"left\"\u003e\n\u003cp\u003eEfficientNetB3-V2\u003c/p\u003e\n\u003c/th\u003e\n\u003c/tr\u003e\n\u003c/thead\u003e\n\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eImage resolution\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e300\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eBatch size\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e16\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eEpochs\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e100\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eK-folds\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e5\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLearning rate\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e1.00E-03\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eFine-tune epochs\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e50\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eFine-tune learning rate\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e1.00E-03\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLoss function\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eWeighted Categorical Cross Entropy\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eEarly stopping\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eTrue\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eEarly stopping patience\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e25\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLabeling methods\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eOne-Hot\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eInput scaling\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e0-255\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eOptimizer\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eAdamv2\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eActivation function\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eReLU\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eCross-validation\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eStratified k-folds\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eTrain test splits\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e20\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePretrained weights\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eImageNet\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDense layers\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e4\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003ePooling\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eAverage\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eDropout\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003e20\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eKernel regularizers\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eL1_L2\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eBatch norm\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eYes\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eLR scheduler\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eYes\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eInput augmentation\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eFlip, Brightness (25%), Saturation (25%)\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003ctr\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eEnsemble voting\u003c/p\u003e\n\u003c/td\u003e\n\u003ctd align=\"left\"\u003e\n\u003cp\u003eSoft\u003c/p\u003e\n\u003c/td\u003e\n\u003c/tr\u003e\n\u003c/tbody\u003e\n\u003c/table\u003e\n\u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\n\u003ch2\u003eDeep-learning network implementation\u003c/h2\u003e\n\u003cp\u003eThe entire network was implemented in TensorFlow 2.8 using the EfficientNetV2-B3 backbone from the Keras library. The reported metrics for each model represent the average performance across folds and are detailed in the \u003cspan class=\"InternalRef\"\u003eResults\u003c/span\u003e section. Training involved 50 epochs with a batch size of 4, a learning rate set to 1e\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e, and the application of weighted sparse categorical cross entropy (CCE) loss. Following the initial training phase, the last 20 layers were unfrozen, and the network underwent an additional 50 epochs of fine-tuning. Since en face SS OCT-A images are grayscale, they were converted to RGB format by triplicating the channel, adhering to EfficientNet\u0026rsquo;s requirement for three-channel images.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\n\u003ch2\u003eDeep-learning model assessment\u003c/h2\u003e\n\u003cp\u003eFor the evaluation, we used balanced accuracy as a metric for our imbalanced dataset, which is calculated as the arithmetic mean of sensitivity (true positive rate) and specificity (true negative rate). To assess the performance of the trained EfficientNetV2-B3 network, we performed a comparative analysis employing a random forest (RF) classifier based on the features extracted from the network. Predictions for each image were generated using a truncated version of the trained network, specifically capturing the features inferred at the last dense layer with a kernel size of 32. These features were used to train an RF classifier comprising 100 trees, and subsequently the test set was classified using this trained RF model. To ensure a fair evaluation, a second RF classifier was trained on the same dataset that was used for the DL network. Consequently, we obtained three distinct sets of inference results in the test set: one from the RF classifier, another from the RF classifier using only extracted features (RF-FO), and the third from the trained DL network.\u003c/p\u003e\n\u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe would like to thank Professor Ramin Tadayoni and Professor Aude Couturier for sharing data regarding the EVIRED study in Dijon.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by G.C. Model development was performed by G.C. and A.A. The first draft of the manuscript was written by G.C. and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe RASTA dataset is hosted and publicly available at https://rasta.u-bourgogne.fr/.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCl\u0026eacute;ment German\u0026egrave;se has no relevant relationships to disclose. Fabrice Meriaudeau has no relevant relationships to disclose. Atif Anwer has no relevant relationships to disclose. Charles Guenancia has no relevant relationships to disclose. Catherine Creuzot-Garcher has no relevant relationships to disclose. Pierre-Henry Gabrielle has no relevant relationships to disclose. Louis Arnould has no relevant relationships to disclose.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eWHO. \u003cem\u003eWorld Health Organization reveals leading causes of death and disability worldwide: 2000-2019\u003c/em\u003e, \u0026lt;https://www.who.int/news/item/09-12-2020-who-reveals-leading-causes-of-death-and-disability-worldwide-2000-2019\u0026gt; (2020).\u003c/li\u003e\n\u003cli\u003eGoff, D. C., Jr.\u003cem\u003e et al.\u003c/em\u003e 2013 ACC/AHA guideline on the assessment of cardiovascular risk: a report of the American College of Cardiology/American Heart Association Task Force on Practice Guidelines. \u003cem\u003eCirculation\u003c/em\u003e \u003cstrong\u003e129\u003c/strong\u003e, S49-73, doi:10.1161/01.cir.0000437741.48606.98 (2014).\u003c/li\u003e\n\u003cli\u003eAnderson, K. M., Wilson, P. W., Odell, P. M. \u0026amp; Kannel, W. B. An updated coronary risk profile. A statement for health professionals. \u003cem\u003eCirculation\u003c/em\u003e \u003cstrong\u003e83\u003c/strong\u003e, 356-362, doi:10.1161/01.cir.83.1.356 (1991).\u003c/li\u003e\n\u003cli\u003egroup, S. w. \u0026amp; collaboration, E. C. r. SCORE2 risk prediction algorithms: new models to estimate 10-year risk of cardiovascular disease in Europe. \u003cem\u003eEuropean heart journal\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, 2439-2454, doi:10.1093/eurheartj/ehab309 (2021).\u003c/li\u003e\n\u003cli\u003eLin, B.\u003cem\u003e et al.\u003c/em\u003e Cumulative risk of stroke recurrence over the last 10 years: a systematic review and meta-analysis. \u003cem\u003eNeurological sciences : official journal of the Italian Neurological Society and of the Italian Society of Clinical Neurophysiology\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, 61-71, doi:10.1007/s10072-020-04797-5 (2021).\u003c/li\u003e\n\u003cli\u003eJokhadar, M., Jacobsen, S. J., Reeder, G. S., Weston, S. A. \u0026amp; Roger, V. L. Sudden death and recurrent ischemic events after myocardial infarction in the community. \u003cem\u003eAmerican journal of epidemiology\u003c/em\u003e \u003cstrong\u003e159\u003c/strong\u003e, 1040-1046, doi:10.1093/aje/kwh147 (2004).\u003c/li\u003e\n\u003cli\u003eArnould, L.\u003cem\u003e et al.\u003c/em\u003e The EYE-MI Pilot Study: A Prospective Acute Coronary Syndrome Cohort Evaluated With Retinal Optical Coherence Tomography Angiography. \u003cem\u003eInvestigative Ophthalmology \u0026amp; Visual Science\u003c/em\u003e \u003cstrong\u003e59\u003c/strong\u003e, 4299-4306, doi:10.1167/iovs.18-24090 (2018).\u003c/li\u003e\n\u003cli\u003eSpaide, R. F., Klancnik, J. M., Jr. \u0026amp; Cooney, M. J. Retinal vascular layers imaged by fluorescein angiography and optical coherence tomography angiography. \u003cem\u003eJAMA ophthalmology\u003c/em\u003e \u003cstrong\u003e133\u003c/strong\u003e, 45-50, doi:10.1001/jamaophthalmol.2014.3616 (2015).\u003c/li\u003e\n\u003cli\u003eMariampillai, A.\u003cem\u003e et al.\u003c/em\u003e Speckle variance detection of microvasculature using swept-source optical coherence tomography. \u003cem\u003eOptics letters\u003c/em\u003e \u003cstrong\u003e33\u003c/strong\u003e, 1530-1532, doi:10.1364/ol.33.001530 (2008).\u003c/li\u003e\n\u003cli\u003eHuang, Y.\u003cem\u003e et al.\u003c/em\u003e Swept-source OCT angiography of the retinal vasculature using intensity differentiation-based optical microangiography algorithms. \u003cem\u003eOphthalmic surgery, lasers \u0026amp; imaging retina\u003c/em\u003e \u003cstrong\u003e45\u003c/strong\u003e, 382-389, doi:10.3928/23258160-20140909-08 (2014).\u003c/li\u003e\n\u003cli\u003eStanga, P. E.\u003cem\u003e et al.\u003c/em\u003e Swept-Source Optical Coherence Tomography Angio\u0026trade; (Topcon Corp, Japan): Technology Review. \u003cem\u003eDevelopments in ophthalmology\u003c/em\u003e \u003cstrong\u003e56\u003c/strong\u003e, 13-17, doi:10.1159/000442771 (2016).\u003c/li\u003e\n\u003cli\u003eJia, Y.\u003cem\u003e et al.\u003c/em\u003e Split-spectrum amplitude-decorrelation angiography with optical coherence tomography. \u003cem\u003eOptics express\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e, 4710-4725, doi:10.1364/oe.20.004710 (2012).\u003c/li\u003e\n\u003cli\u003eLa\u0026iacute;ns, I.\u003cem\u003e et al.\u003c/em\u003e Retinal applications of swept source optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA). \u003cem\u003eProg Retin Eye Res\u003c/em\u003e \u003cstrong\u003e84\u003c/strong\u003e, 100951, doi:10.1016/j.preteyeres.2021.100951 (2021).\u003c/li\u003e\n\u003cli\u003eLip, G. Y., Nieuwlaat, R., Pisters, R., Lane, D. A. \u0026amp; Crijns, H. J. Refining clinical risk stratification for predicting stroke and thromboembolism in atrial fibrillation using a novel risk factor-based approach: the euro heart survey on atrial fibrillation. \u003cem\u003eChest\u003c/em\u003e \u003cstrong\u003e137\u003c/strong\u003e, 263-272, doi:10.1378/chest.09-1584 (2010).\u003c/li\u003e\n\u003cli\u003eWagner, S. K.\u003cem\u003e et al.\u003c/em\u003e Insights into Systemic Disease through Retinal Imaging-Based Oculomics. \u003cem\u003eTransl Vis Sci Technol\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 6, doi:10.1167/tvst.9.2.6 (2020).\u003c/li\u003e\n\u003cli\u003eArnould, L.\u003cem\u003e et al.\u003c/em\u003e Using Artificial Intelligence to Analyse the Retinal Vascular Network: The Future of Cardiovascular Risk Assessment Based on Oculomics? A Narrative Review. \u003cem\u003eOphthalmol Ther\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 657-674, doi:10.1007/s40123-022-00641-5 (2023).\u003c/li\u003e\n\u003cli\u003eKhan, S. M.\u003cem\u003e et al.\u003c/em\u003e A global review of publicly available datasets for ophthalmological imaging: barriers to access, usability, and generalisability. \u003cem\u003eLancet Digit Health\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, e51-e66, doi:10.1016/s2589-7500(20)30240-5 (2021).\u003c/li\u003e\n\u003cli\u003eGerman\u0026egrave;se, C.\u003cem\u003e et al.\u003c/em\u003e A Retinal Oct-Angiography and Cardiovascular STAtus (RASTA) Dataset of Swept-Source Microvascular Imaging for Cardiovascular Risk Assessment. \u003cem\u003eData\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 147 (2023).\u003c/li\u003e\n\u003cli\u003ePoplin, R.\u003cem\u003e et al.\u003c/em\u003e Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. \u003cem\u003eNat Biomed Eng\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, 158-164, doi:10.1038/s41551-018-0195-0 (2018).\u003c/li\u003e\n\u003cli\u003eRudnicka, A. R.\u003cem\u003e et al.\u003c/em\u003e Artificial intelligence-enabled retinal vasculometry for prediction of circulatory mortality, myocardial infarction and stroke. \u003cem\u003eThe British journal of ophthalmology\u003c/em\u003e \u003cstrong\u003e106\u003c/strong\u003e, 1722-1729, doi:10.1136/bjo-2022-321842 (2022).\u003c/li\u003e\n\u003cli\u003eSun, Z., Yang, D., Tang, Z., Ng, D. S. \u0026amp; Cheung, C. Y. Optical coherence tomography angiography in diabetic retinopathy: an updated review. \u003cem\u003eEye (London, England)\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 149-161, doi:10.1038/s41433-020-01233-y (2021).\u003c/li\u003e\n\u003cli\u003eMonteiro-Henriques, I., Rocha-Sousa, A. \u0026amp; Barbosa-Breda, J. Optical coherence tomography angiography changes in cardiovascular systemic diseases and risk factors: A Review. \u003cem\u003eActa Ophthalmol\u003c/em\u003e \u003cstrong\u003e100\u003c/strong\u003e, e1-e15, doi:10.1111/aos.14851 (2022).\u003c/li\u003e\n\u003cli\u003eAker, A., Volis, I., Saliba, W., Naoum, I. \u0026amp; Zafrir, B. CHA2DS2-VASc Score as a Predictor of Adverse Outcomes after Ischemic Stroke in Patients without Atrial Fibrillation. \u003cem\u003eThe Israel Medical Association journal : IMAJ\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 143-148 (2024).\u003c/li\u003e\n\u003cli\u003eMelgaard, L.\u003cem\u003e et al.\u003c/em\u003e Assessment of the CHA2DS2-VASc Score in Predicting Ischemic Stroke, Thromboembolism, and Death in Patients With Heart Failure With and Without Atrial Fibrillation. \u003cem\u003eJama\u003c/em\u003e \u003cstrong\u003e314\u003c/strong\u003e, 1030-1038, doi:10.1001/jama.2015.10725 (2015).\u003c/li\u003e\n\u003cli\u003eSu, C. H.\u003cem\u003e et al.\u003c/em\u003e CHA2DS2-VASc score as an independent outcome predictor in patients hospitalized with acute ischemic stroke. \u003cem\u003ePLoS One\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, e0270823, doi:10.1371/journal.pone.0270823 (2022).\u003c/li\u003e\n\u003cli\u003eGuerra, F.\u003cem\u003e et al.\u003c/em\u003e CHA(2)DS(2)-VASc risk factors as predictors of stroke after acute coronary syndrome: A systematic review and meta-analysis. \u003cem\u003eEuropean heart journal. Acute cardiovascular care\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 264-274, doi:10.1177/2048872616673536 (2018).\u003c/li\u003e\n\u003cli\u003eTu, H. T.\u003cem\u003e et al.\u003c/em\u003e Pre-stroke CHADS2 and CHA2DS2-VASc scores are useful in stratifying three-month outcomes in patients with and without atrial fibrillation. \u003cem\u003eCerebrovascular diseases (Basel, Switzerland)\u003c/em\u003e \u003cstrong\u003e36\u003c/strong\u003e, 273-280, doi:10.1159/000353670 (2013).\u003c/li\u003e\n\u003cli\u003eXing, Y.\u003cem\u003e et al.\u003c/em\u003e CHA(2)DS(2)-VASc score as a predictor of long-term cardiac outcomes in elderly patients with or without atrial fibrillation. \u003cem\u003eClinical interventions in aging\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 497-504, doi:10.2147/cia.s147916 (2018).\u003c/li\u003e\n\u003cli\u003eAkboğa, M. K., Yılmaz, S. \u0026amp; Yal\u0026ccedil;ın, R. Prognostic value of CHA2DS2-VASc score in predicting high SYNTAX score and in-hospital mortality for non-ST elevation myocardial infarction in patients without atrial fibrillation. \u003cem\u003eAnatolian journal of cardiology\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 789-795, doi:10.5152/AnatolJCardiol.2021.03982 (2021).\u003c/li\u003e\n\u003cli\u003eTan, M. \u0026amp; Le, Q. (2019).\u003c/li\u003e\n\u003cli\u003eDecenci\u0026egrave;re, E.\u003cem\u003e et al.\u003c/em\u003e FEEDBACK ON A PUBLICLY DISTRIBUTED IMAGE DATABASE: THE MESSIDOR DATABASE. \u003cem\u003eImage Analysis \u0026amp; Stereology\u003c/em\u003e, 231-234, doi:10.5566/ias.1155 (2014).\u003c/li\u003e\n\u003cli\u003eCheung, C. Y.\u003cem\u003e et al.\u003c/em\u003e A deep-learning system for the assessment of cardiovascular disease risk via the measurement of retinal-vessel calibre. \u003cem\u003eNat Biomed Eng\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 498-508, doi:10.1038/s41551-020-00626-4 (2021).\u003c/li\u003e\n\u003cli\u003eZhang, K.\u003cem\u003e et al.\u003c/em\u003e Deep-learning models for the detection and incidence prediction of chronic kidney disease and type 2 diabetes from retinal fundus images. \u003cem\u003eNat Biomed Eng\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 533-545, doi:10.1038/s41551-021-00745-6 (2021).\u003c/li\u003e\n\u003cli\u003eHassan, O.\u003cem\u003e et al.\u003c/em\u003e Deep learning prediction of age and sex from optical coherence tomography. \u003cem\u003eIEEE 18th International Symposium on Biomedical Imaging (ISBI). Nice, France: IEEE\u003c/em\u003e, 238-242 (2021).\u003c/li\u003e\n\u003cli\u003eMunk, M. R.\u003cem\u003e et al.\u003c/em\u003e Assessment of patient specific information in the wild on fundus photography and optical coherence tomography. \u003cem\u003eScientific reports\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 8621, doi:10.1038/s41598-021-86577-5 (2021).\u003c/li\u003e\n\u003cli\u003eArnould, L.\u003cem\u003e et al.\u003c/em\u003e Prediction of Cardiovascular Parameters With Supervised Machine Learning From Singapore \u0026quot;I\u0026quot; Vessel Assessment and OCT-Angiography: A Pilot Study. \u003cem\u003eTransl Vis Sci Technol\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 20, doi:10.1167/tvst.10.13.20 (2021).\u003c/li\u003e\n\u003cli\u003eZhong, P.\u003cem\u003e et al.\u003c/em\u003e Development and Validation of Retinal Vasculature Nomogram in Suspected Angina Due to Coronary Artery Disease. \u003cem\u003eJournal of atherosclerosis and thrombosis\u003c/em\u003e \u003cstrong\u003e29\u003c/strong\u003e, 579-596, doi:10.5551/jat.62059 (2022).\u003c/li\u003e\n\u003cli\u003e\u003cem\u003eVARPA Working fields: Public databases\u003c/em\u003e, \u0026lt;http://www.varpa.es/research/ophtalmology.html#octagon\u0026gt; (\u003c/li\u003e\n\u003cli\u003eMa, Y.\u003cem\u003e et al.\u003c/em\u003e ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model. \u003cem\u003eIEEE transactions on medical imaging\u003c/em\u003e \u003cstrong\u003e40\u003c/strong\u003e, 928-939, doi:10.1109/tmi.2020.3042802 (2021).\u003c/li\u003e\n\u003cli\u003eLi, M.\u003cem\u003e et al.\u003c/em\u003e OCTA-500: A retinal dataset for optical coherence tomography angiography study. \u003cem\u003eMedical image analysis\u003c/em\u003e \u003cstrong\u003e93\u003c/strong\u003e, 103092, doi:10.1016/j.media.2024.103092 (2024).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"OCT-A, CHA2DS2-VASc, machine learning, deep learning, neurocardiovascular, prediction","lastPublishedDoi":"10.21203/rs.3.rs-4326028/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4326028/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe recent rise of artificial intelligence represents a revolutionary way of improving current medical practices, including cardiovascular (CV) assessment scores. Retinal vascular alterations may reflect systemic processes such as the presence of CV risk factors. The value of swept-source retinal optical coherence tomography\u0026ndash;angiography (SS OCT-A) imaging is significantly enhanced by image analysis tools that provide rapid and accurate quantification of vascular features. We report on the interest of using machine-learning (ML) and deep-learning (DL) models for CV assessment from SS OCT-A microvasculature imaging. We assessed the accuracy of ML and DL algorithms in predicting the CHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VASc neurocardiovascular score based on SS OCT-A retinal images of patients from the open-source RASTA dataset. The ML and DL models were trained on data from 491 patients. The ML models tested here achieved good performance with area under the curve (AUC) values ranging from 0.71 to 0.96. According to a classification into two or three CV risk groups, the EfficientNetV2-B3 tool predicted risk correctly in 39% and 68% of cases, respectively, with a mean absolute error (MAE) of approximately 0.697. Our models enable a confident prediction of the CHA\u003csub\u003e2\u003c/sub\u003eDS\u003csub\u003e2\u003c/sub\u003e-VASc score from SS OCT-A imaging, which could be a useful tool contributing to the assessment of neurocardiovascular profiles in the future.\u003c/p\u003e","manuscriptTitle":"Artificial intelligence-based prediction of neurocardiovascular risk score from retinal swept-source microvascular imaging: the RASTA dataset","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-05-02 19:03:12","doi":"10.21203/rs.3.rs-4326028/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-06-24T06:14:19+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-06-06T18:17:30+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-05-31T21:59:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"13749260664336508209384836917977481420","date":"2024-05-21T12:25:28+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"217242021273070547144150923935366923070","date":"2024-05-20T19:46:28+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-05-19T12:20:05+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-05-15T04:21:35+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2024-04-28T15:54:17+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-04-28T15:52:14+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2024-04-25T20:09:52+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"88029f3c-d52c-41dc-9d18-9c097958b74b","owner":[],"postedDate":"May 2nd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":31398157,"name":"Health sciences/Biomarkers"},{"id":31398158,"name":"Health sciences/Cardiology"},{"id":31398159,"name":"Health sciences/Neurology"},{"id":31398160,"name":"Health sciences/Risk factors"}],"tags":[],"updatedAt":"2024-11-11T16:01:18+00:00","versionOfRecord":{"articleIdentity":"rs-4326028","link":"https://doi.org/10.1038/s41598-024-78587-w","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2024-11-07 15:57:38","publishedOnDateReadable":"November 7th, 2024"},"versionCreatedAt":"2024-05-02 19:03:12","video":"","vorDoi":"10.1038/s41598-024-78587-w","vorDoiUrl":"https://doi.org/10.1038/s41598-024-78587-w","workflowStages":[]},"version":"v1","identity":"rs-4326028","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4326028","identity":"rs-4326028","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}