Detection of Systemic Diseases From Ocular Images Using Artificial Intelligence: A Systematic Review : The Asia-Pacific Journal of Ophthalmology

Secondary Logo

Journal Logo

Review Article

Detection of Systemic Diseases From Ocular Images Using Artificial Intelligence: A Systematic Review

Peng, Qingsheng MD∗,†; Tseng, Rachel Marjorie Wei Wen BA; Tham, Yih-Chung PhD∗,‡; Cheng, Ching-Yu MD, PhD∗,‡,§; Rim, Tyler Hyungtaek MD, PhD∗,‡

Author Information
Asia-Pacific Journal of Ophthalmology 11(2):p 126-139, March-April 2022. | DOI: 10.1097/APO.0000000000000515


Among systemic diseases, cardiovascular and cerebrovascular diseases are the leading causes of mortality1 and these extensive effects have been demonstrated by the strong association between ocular neurovascular structure and disease progression and prognosis.2–7 Ocular changes have also been found in other diseases, such as kidney and degenerative neural diseases.8,9 With the increasing number of associations found between the eye and other body systems, it could be worthwhile exploring the possibility of a universal screening tool using ocular images.

Since the last century, studies have employed color fundus photography (CFP) as a standard screening technique for retinal diseases.10,11 Large quantities of CFP images are sufficient to build predictive models, and the recent development of handheld and smartphone-based CFP cameras has increased the availability and usability of CFP.12 However, traditional utilization of information from ocular images, whether through simple statistics or subjective judgment made by the human eye, is limited in scope compared to using artificial intelligence (Ai). Ai presents opportunities to extract essential features from retina images by machine learning (ML) and deep learning (DL) algorithms and this valuable information, demonstrated through heatmaps,13 can be used selectively based on the retinal experts’ region of interest. Although heatmaps have poor capability explaining global pixel connections on the image, the machine vision can help reveal the unknown correlation between the highlighted areas and the outcome, especially in systemic diseases where little is known about their shared pathology with the eye. Furthermore, DL features can be transformed to risk scores, representing the total risk of the systemic disease calculated from ocular image.14 This, in addition to pathological changes, provides crucial information to clinicians when diagnosing a disease.

Although guidelines have not recommended Ai as a diagnostic or prognosis management tool, Ai algorithms are already widely used in clinical practice, and the US Food and Drug Administration (FDA) has approved 130 Ai devices as of 2020.15 While vascular changes on the retina can be homogenous in multiple systemic diseases, some retinal vascular alterations have already been identified as important evidence for specific systemic diseases.16,17 For example, the retinal vasospasm is a distinguishable sign for diagnosing pre-eclampsia in pregnant women.16 Therefore, Ai studies could potentially reveal unknown connections between systemic diseases and ocular images such as specific retinal alterations, and this can then be utilized as a screening tool for multiple systemic diseases.

In addition to CFP, dynamic retinal vessel analysis (DVA) monitors retinal vessel dilation and blood flow volume, and the development of optical coherence tomography (OCT) and OCT angiography (OCTA) enables in vivo view of vitreoretinal neurovascular microstructures.18,19 This contains more specific retinal neural and vascular structure information than CFP. Pilot studies revealed correlations between systemic diseases and OCT/OCTA and DVA parameters. Based on these findings, proof-of-concept studies developed ML approaches to predict systemic diseases. The disadvantage of applying these ocular imaging modalities other than CFP on systemic disease screening is that the average population rarely receives these examinations. However, the newly developed ocular images can contribute different information than CFP and require a smaller sample size as the information density is higher than CFP.

Therefore, applying AI to easily accessible ocular images could increase the feasibility of population-based screening for systemic diseases. This review aims to summarize current proof- of-concept and application studies of ocular image-based AI on detecting systemic diseases.


This systematic review was written following the Preferred Reporting Items for Systematic Reviews and Meta-analysis Protocols (PRISMA) checklist. The complete PRISMA checklist is shown in Supplementary Digital Content, Table 1,

Search Strategy and Inclusion Criteria

A systematic search was conducted on PubMed and Web of Science on September 1, 2021. The additional broad search was done on Google Scholar to find missed papers. Articles that applied artificial intelligence models processing ocular images data to diagnose diseases other than ophthalmic diseases were included. Secondary studies were excluded. The search terms for ocular images were “fundus photo,” “optical coherence tomography,” “optical coherence tomography angiography,” “anterior segment optical coherence tomography,” “ultra-widefield,” “external eye photo,” and “dynamic retina vessel analysis.” Search terms for systemic diseases include “cardiovascular,” “cerebrovascular,” “hypertension,” “diabetes,” “degenerative neural diseases,” “cognitive disorder,” “dementia,” “hepatopathy,” “metabolic disorder,” “kidney,” and “hepatopathy.” Methodology terms were chosen as “machine learning,” “random forest,” “support vector machine,” “deep learning,” “automatic framework,” “ensemble learning,” “embedded network,” and “neural network.” Boolean operator “AND” was applied between the 3 components, and “OR” was used between parallel search terms within each component.

All reviewed papers were available in full text and English. Cross-sectional and retrospective studies were included. Surveys, case reports, case series, reviews, guidelines, statistical models, and papers aiming at screening eye diseases were excluded. AI methodology was defined as ML or DL algorithms that were applied during ocular image-based model construction, and the models were further validated using masked data. AI methodology studies focusing on improving risk stratification were also included.

Data Extraction

Screening of the articles and data extraction was conducted by QP. Data extracted included study setting details (study name, first author, and the year of publication), study population (sample size, internal database, and external database), application (systemic disease name, disease category, and outcome formality), ocular image (imaging type, number of images), AI model used (name of the neural network, training platform), study results and their conclusions (Tables 1 and 2).

Table 1 - Summary of Studies Detecting Systemic Disease From Ocular Image
Author, y Input AI Structure Training/Testing Dataset External Validation Systemic Disease Outcome Type, Results
Chang et al, 2020 CFP Xeception 5,296 images for training; 647 for tuning; 654 for testing NA carotid artery atherosclerosis Binary, AUROC = 0.713; ACC = 0.583;
Wing et al, 2021 contact CFP U-net & RBFNN 58 newborns in total, 80% for training and 20% for validation NA Cyanotic CHD Binary, AUROC = 0.897
Poplin et al, 2018 CFP, age, gender, BMI, SBP, current smoke Inception-3 48,101 images from UK Biobank and 236,234 from EyePACS for training; 12,026 from UK Biobank and 999 from EyePACS for validation NA MACE Binary, AUROC = 0.73
Rim et al, 2021 CFP Efficientnet 15,911 images for training; 2536 for testing 8707 patients from health-screening centre affiliated with the Philip Medical Centre, South Koreaand 527 patients from CMERCHI cohort Coronary artery calcium Three-way: CACs = O; O<CACs<lOO; CACs >100, external AUROC = 0.742
Son et al, 2020 CFP Inception-3 20130 images for training and 5-fold validation NA Coronary artery calcium Binary: CACs = 0; CACs>O, AUROC = 0.75
Zhong et al, 2021 OCTA, ECG, HDL, LDL and sex Lasso regression 508 images for training and 287 for validation NA CAD Binary, AUROC = 0.897
Lim et al, 2019 CFP U-net & VGGNet 4,528 images from stroke patients and 6,622 from controls; 80% for training, 5% for validation and 15% for testing NA Ishemic stroke Binary, AUROC ranged from 0.496 to 0.966 on split datasets
Dai et al, 2020 CFP U-net & CNN 1,419 images in total, 75% for training, 25% for testing NA Hypertension Binary, Mean ACC = 0.6094; Mean AUROC = 0.6506
Zhang et al, 2020 CFP Inception-3 625 images in total, 80% for training, 10% for validation and 10% for testing NA Hypertension Binary, AUROC = 0.766
Benson et al, 2020 CFP VGG-16 46 patients and 96 controls, 80% for training and 20% for validation NA DPN Binary,ACC = 0.89,sensitivity = 78% and specificity = 95%
Aslam et al, 2020 OCTA RF 152 images for training and 30 for validation NA Diabetes Binary, AUROC = 0.8
Govindaswamy et al, 2020 OCT and OCTA SVM, RF 314 images for training and 5fold validation NA Diabetes Three-way: normal; DM; DR, ACC(for DM) = 0.74
Islam et al, 2021 CFP DenseNet-121 492 images in total for training and nested cross-validation NA Diabetes Binary, ACC = 0.84
Zhang et al, 2021 CFP, age, gender, height, weight and BP ResNet-50 43,156 images in total, 70% for training, 10 for tuning and 20% for validation 8,059 patients from CC-FII-C and 3,081 patients from COACS cohort Type 2 diabetes CKDEarly CKD Binary, external AUROC ranged from 0.845 to 0.871Binary, external AUROC ranged from 0.897 to 0.898Binary, external AUROC ranged from 0.845 to 0.848
Kang et al, 2020 CFP VGG-19 6,212 images in total, 80% for training, 10% for validation and 10% for testing NA Early CKD Binary, AUROC = 0.81
Sabanayagam et al, 2020 CFP, age, sex, ethnicity, diabetes, and hypertension cCondenseNet 5188 patients for training, 1297 patients for validation 3735 patients in SP2, 1538 patients in BES CKD Binary, external AUROC ranged from 0.827 to 1.00
Lemmens et al, 2020 HSI and OCT LDA 39 patients in total for training and nested cross-validation NA AD Binary, AUROC ranged from 0.67 to 0.79
Tian et al, 2021 CFP SVM, RBF 122 images from AD patients and 52,492 images from controls, for training and 5-fold cross-validation NA AD Binary, ACC = 0.8244
Wisely et al, 2020 UWF FP, OCT, OCTA, UWF FAF ResNet-18 284 eyes for training and validation, 68 eyes for testing NA AD Binary, AUROC (multimodal) ranged from 0.829 to 0.841
Lai et al, 2020 CFP ARIA, ResNet-50 & SVM 23 patients with paired 23 controls, for training and 10-fold cross validation NA Autism Binary, AUROC = 0.907, sensitivity = 82.6%and specificity = 91.3%
Rim et al, 2020 CFP VGG-16 86,994 images for training, 21,698 for testing 9324 images from a health screening centre affiliated with the Severance Gangnam Hospital.,4234 images from BES, 63275 images from SEED, and 50732 images from UK Biobank Anemia Regression, external r 2 ranged from 0.36 to 0.61
Chen et al, 2019 OCT PCA, LDA 24 patients and 16 controls for training and LOO cross-validation NA Anemia Binary, ACC = 0.836, sensitivity = 82.6% and specificity = 91.3%
Wei et al, 2021 OCT AlexNet 17 patients and 13 controls for training and 5-fold cross validation NA Anemia Binary, AUROC = 0.9983, sensitivity = 98.38% and specificity = 95.94%
Mitani et al, 2019 CFP Inception-4 57,163 in total, 70% for training, 10 for tuning and 20% for validation NA Anemia Binary, AUROC = 0.870
Xiao et al, 2021 Slitlamp image, CFP U-net & ResNet-101 1252 images in total, 75% for training, 25% for tuning 537 patients from the Department of Infectious Diseases of the Third Affiliated Hospital of Sun Yat-sen University and the Huanshidong Medical Centre of Aikang Health Care Hepatobiliary disease Binary, external AUROC for slitlamp =0.74; AUROC for CFP = 0.68
Liver cancerLiver cirrhosischronic viral hepatitisnonalcoholic fatty liver disease CholelithiasisHepatic cyst Binary, external AUROC for slitlamp =0.93; AUROC for CFP = 0.84 Binary, external AUROC for slitlamp =0.90; AUROC for CFP = 0.83Binary, external AUROC for slitlamp =0.69; AUROC for CFP = 0.62Binary, external AUROC for slitlamp =0.55; AUROC for CFP = 0.62Binary, external AUROC for slitlamp =0.58; AUROC for CFP = 0.68Binary, external AUROC for slitlamp =0.66; AUROC for CFP = 0.69
ACC indicates accuracy; AD, Alzheimer disease; AUROC, area under the receiver operating curve; BES, Beijing Eye Study; BMI, body-mass index; BP, blood pressure; CAC, coronary artery calcium; CAD, coronary artery disease; CC-FII-C, China Consortium of Fundus Image Investigation Cross-sectional cohort; CFP, color fundus photo; CHD, congenital heart disease; CKD, chronic kidney disease; CMERCHI, Cardiovascular and Metabolic- Disease Etiology Research Center - High Risk Cohort; COACS, Colchicine for Acute Coronary Syndromes study; DM, diabetes mellitus; DPN, diabetic peripheral neuropathy; DR, diabetic retinopathy; ECG, electrocardiogram; HDL, high-density lipoprotein; HSI, hyperspectral retinal imaging; LDA, linear discriminant analysis; LDL, low-density lipoprotein; MACE, major adverse cardiovascular events; OCT, optical coherence- tomography; OCTA, optical coherence tomography angiography; PCA, primary component analysis; RBF, radial basis function; RF, random forest; SBP, systolic blood pressure; SEED, Singapore Epidemiology of Eye Diseases; SP2, Singapore Prospective Study Program; SVM, support vector machine.

Table 2 - Summary for Studies Predicting Systemic Risk Factors
Author, y Input Model Structure Training/Testing Dataset External Validation Systemic Risk Factors Outcome Type, Results
Poplin et al, 2018 CFP Inception-3 48,101 images from UK Biobank and 236,234 from EyePACS for training; 12,026 from UK Biobank and 999 from EyePACS for validation NA SBPDBPAgeGenderBMIHbAlcSmoking Regression, r 2 = 0.36Regression, r 2 = 0.32Regression, UK Biobank: r 2 = 0.74; EyePACS: r 2 = 0.82Binary, UK Biobank: AUROC =0.97; EyePACS: AUROC = 0.97Regression, r 2 = 0.13Regression, r 2 = 0.09Binary, r 2 = 0.32
Zhang et al, 2020 CFP Inception-3 625 images in total, 80% for training, 10% for validation and 10% for testing NA HyperglycemiaDyslipidemiaAgeGenderSmokingDrinking Salty tasteBMIWHR HCTMCHCTBILDBIL Binary, ACC = 0.748, AUROC = 0.850Binary, ACC = 0.667, AUROC = 0.703Binary, ACC = 0.748, AUROC = 0.850Binary, ACC = 0.624, AUROC = 0.704Binary, ACC = 0.732, AUROC = 0.794Binary, ACC = 0.863, AUROC = 0.948Binary, ACC = 0.757, AUROC = 0.809Binary, ACC = 0.712, AUROC = 0.731Binary, ACC = 0.646, AUROC = 0.704Binary, ACC = 0.698, AUROC = 0.759Binary, ACC = 0.605, AUROC = 0.686Binary, ACC = 0.700, AUROC = 0.764Binary, ACC = 0.650, AUROC = 0.703
Betzler et al, 2021 CFP VGG-16 9.956 images in total, 79.9% for training and 20.1% for validation NA Gender Binary, overall AUROC = 0.94
Kim et al, 2020 CFP ResNet-152 216,866 normal images for training, 24,36 images for validation and 24,366 images for testing 40,659 hypertensive patients’ images, 14,189 DM patients’ images and 113,510 smokers’ images from the health promotion centre in SNUBH AgeGender Regression, normal participants, MAE = 3.06; hypertensive patients, MAE = 3.46; DM patients, MAE = 3.55; smokers, MAE = 2.76Binary, normal participants, AUROC = 0.97; hypertensive patients, AUROC = 0.96; DM patients, AUROC = 0.96; smokers, AUROC = 0.98
Ma et al, 2021 Ocular anterior segment image LR, XGB and NN 76,065 images for training 100 volunteers recruited in Tianjing Age Regression, r 2 = 0.76
Cheung et al, 2021 CFP SIVA-DLS 2,267 images from SCES, 2,528 from SINDI, and 514 from SiMES, 80% for training, 20% for validation 45,644 from UK Biobank, 627 from the Kangbuk Samsung health study, and 243 from the Austin health study AgeGenderMABPBMIHbAlcSmokingTC Regression, external r 2 ranged from 0.195 to 0.212Regression, external r 2 ranged from 0.063 to 0.089Regression, external r 2 ranged from 0.199 to 0.201Regression, external r 2 ranged from 0.244 to 0.251Regression, external r 2 ranged from 0.202 to 0.215Regression, external r 2 = 0.045Regression, external r 2 ranged from 0.067 to 0.069
Rim et al, 2020 CFP VGG-16 86,994 images for training, 21,698 for testing 9324 images from a health screening centre affiliated with the Severance Gangnam Hospital.,4234 images from BES, 63275 images from SEED, and 50732 images from UK Biobank AgeGenderBody muscular massHeightWeightPercentage body fatBMISBPDBPCreatinineHCTHemoglobin Regression, external r 2 ranged from 0.36 to 0.61Binary, external AUROC ranged from 0.80 to 0.91Regression, external r 2 = 0.33Regression, external r 2 ranged from 0.08 to 0.28Regression, external r 2 ranged from 0.04 to 0.19Regression, external r 2 = 0.08Regression, external r 2 ranged from 0.04 to 0.14Regression, external r 2 ranged from 0.17 to 0.21Regression, external r 2 ranged from 0.16 to 0.27Regression, external r 2 ranged from 0.01 to 0.26Regression, external r 2 ranged from 0.09 to 0.26Regression, external r 2 ranged from 0.06 to 0.33
Vaghefi et al, 2019 CFP CNN 165,104 images from DM patients, 60% for training, 20% for validation and 20% for testing NA Smoking Binary, ACC = 0.8888, AUROC = 0.86
Lau et al, 2018 CFP FNN and PRNN 180 participants for training and 10-fold cross-validation NA WMH Binary, sensitivity at 0.929 and specificity at 0.984
Parameswari et al. 2020 CFP Bayesian and MLP DRIVE and STARE for training NA Atherosclerosis Binary, sensitivity at 0.97 and specificity at 0.97
Zee et al, 2021 CFP ARIA, ResNet-50 and SVM 180 participants for training and 10-fold cross-validation 60 collected prospectively participants WMH Binary, external sensitivity at 0.897 and specificity at 0.968
ACC indicates accuracy; AUROC, area under the operating curve; BES, Beijing Eye Study; BMI, body-mass index; CFP, color fundus photography; D-BIL, direct biliary acid; DBP, diastolic blood pressure; HbA1c, glycosylated hemoglobin; HCT, hematocrit; LR, logistic regression; MABP, mean arterial blood pressure; MCHC, mean corpuscular hemoglobin concentration; NN, neural network; SBP, systolic blood pressure; SCES, Singapore Chinese Eye Study; SEED, Singapore Epidemiology of Eye Diseases; SiMES, Singapore Malay Eye Study; SINDI, Singapore Indian Eye Study; SNUBH, Seoul National University Bundang Hospital; T-BIL, total biliary acid; TC, total cholesterol; WHR, waist-hip ratio; WMH, white matter hyperintensities.

As AI studies vary in study design, model structure, and performance evaluation, meta-analysis was not applicable. Preliminary study results and findings were discussed narratively based on systemic disease, methodological approach, and study findings. The study outcomes were assessed by QP. The area under the receiver operating curve (AUROC), accuracy (ACC), sensitivity, and specificity were chosen to assess classification models’ performance and mean absolute error (MAE) and r2 for regression models’ performance. All results related to the major findings of each paper were extracted. QP and RT conducted the risk of bias assessment following the Joanna Briggs Institute (JBI) critical appraisal tools.20 Included studies were categorized as low, unclear, and high risk of bias (Supplementary Digital Content, Table 2,


A total of 500 papers containing 335 articles from PubMed, 147 articles from Web of Science (WOS), and 18 papers from Google Scholar were retrieved in the initial literature search. Only articles available in full text and published in English were included, leaving 278 papers for review for their titles and abstracts after removing duplicates. 214 papers were further excluded based on exclusion criteria (9 reviews and 3 editorial materials from WOS; 34 reviews and meta-analyses from PubMed). 59 papers were evaluated for eligibility based on their full text, leaving 33 included papers in the review.

For the risk of bias analysis (Supplementary Digital Content, Table 2,, all 33 studies were classified as diagnostic accuracy tests for ocular image – based AI and were assessed for their patient selection, ocular image types, reference standard, medical examination flow, and the interval between ocular and gold standard examinations. All studies were designed retrospectively, and researchers were not blinded to patient and control selection. However, AI predictions are objective and therefore unblinding patients will not increase the risk of bias. All studies used confirmed diagnoses in each systemic disease as the reference standards. The input data only contain the ocular images and limited health parameters, no identity information leak to the algorithms was found in all studies. The studies using registered clinical databases and more evenly distributed datasets have relatively lower bias, as the biases from ML and DL algorithms mainly come from the training dataset. One study did not demonstrate the enrollment process, putting it at high risk of selection bias. The time intervals between the point when ocular images were taken and the point when diagnoses were made in 3 studies were not indicated21–23; the interval was critical in the study predicting coronary artery disease (CAD), which may lead to bias as CAD changes are time-dependent.21 In the other 29 studies, chronic diseases diagnoses were made before ocular examination or at the same time using blood test results, which posed a low risk of bias. 1 study did not present any information on the exclusion criteria and how the diagnoses were made.23


Detection of Cardio-Cerebral Vascular Diseases and Risk Factors

Studies Using Retinal Images

Cardiovascular and cerebrovascular diseases (CVDs) share similar pathophysiology and risk factors, and their associations to the eye were well described in a series of cross-sectional and prospective epidemiology studies.2,3,5 The decrease in the arterio- les-venules ratio was a predictor of hypertension, stroke morbidity, and long-term survival,3 while retinal arterioles narrowing was a significant risk factor for CAD in women and carotid arterial atherosclerosis.2 There were also other signs such as lower fractal dimension (FD), arteriole-venule nicking, microaneurysms, focal arteriole narrowing, hemorrhages, and exudates that increased the risk of CVDs.2,5 The complexity involved in analyzing these myriads of features can be simplified with automatic feature extraction using ML algorithms, allowing these retinal characteristics to be combined into a single modality for evaluation.

Hypertension was the first CVD shown to affect ocular vasculature and was also deemed a risk factor for other lifethreatening CVDs such as stroke. Poplin et al24 first used retinal fundus photo-based DL to predict blood pressure (BP). However, the performance of similar AI systems detecting systemic hypertension in the East Asian population varied,25,26 possibly due to hypertension phenotypes and comorbid diseases. The OCTA studies also suggested that better BP control in the past 6 months correlated with retinal blood flow positively27,28 and that retinal vasculature related more to long-term BP control than BP at a specific timepoint.28 For people with poor compliance to home BP monitoring in lower-income regions or with white coat hypertension (WCH), ocular image - based AI could come in handy when off-office BP measurements are absent.

Studies have shown that arterial atherosclerosis predisposed ischemic heart disease and stroke and was correlated to the progression of hypertension.29 These findings led to the development of AI tools for identifying arterial atherosclerosis on the retina,23 in which a high sensitivity of 89.1% was achieved.30 However, the model was built on a dataset with a small sample size with limited diagnosis information, and its diagnostic power was not tested on external datasets. Using DL to detect coronary artery calcium (CAC) with retina photos alone or combining retina images with other risk factors also proved feasible.14 The algorithms achieved AUROC ranging from 0.742 to 0.750 when coronary computed tomography (CT) scans were used to distinguish healthy controls from patients with CAC.14,31 Considering the benefit of detecting CAC during regular ophthalmic follow-up examinations, the DL models can be a supplementary tool to help refer patients with CAC to cardiologists.

As the retina extends from cranial nerves, the avascular zone and edema neurons in the nerve center impact retinal neural structures through transneuronal retrograde degeneration, as shown in OCT changes of stroke patients.6,32 Apart from lowered retinal vascular FD,3 dilated venule,5 and other retinal signs, neural damage on retinal nerve fiber and ganglion cells worsened with time after ischemic stroke happened.32 An AUROC of 0.966 was observed on the prediction of stroke using retinal images, and the feature isolation analysis suggested that both vascular and neural information was critical to the model. However, the training datasets used for the positive and negative outcomes were different, resulting in a chance of identity leakage during both training and validation.33 In predicting the occurrence of major adverse cardiovascular events (MACE) in 5 years, Poplin et al24 trained a DL model with inputs consisting of cyan fluorescent protein (CFP) image, age, systolic blood pressure, body-mass index (BMI), gender, and smoking, and this model demonstrated better performance than the Systematic Coronary Risk Evaluation (SCORE) calculator.

Lastly, studies have shown that patients with congenital heart disease (CHD), the leading cause of child death, have underdeveloped retinal macroscopic vasculature.34 The vessel density in the macula and radial pericapillary capillaries (RPC) dropped significantly compared to the normal population.35 Retina images taken from contact fundus photography systems during retinopathy of prematurity screening in the neonatal intensive care unit (NICU) were demonstrated to hold diagnostic power over cyanosis in CHD patients and help assess rectification surgery risk by predicting cardiopulmonary bypass time (CBT).36

Studies Using Other Ocular Images

OCTA studies revealed deeper pathology of retinal and choroidal microvasculature changes in CVD patients. In CAD patients, the density of superficial and deep capillary plexus (SCP and DCP) at the macula and choroidal capillaries dropped.37 This reduction was associated with Gensini Scores and aligned with the American Heart Association (AHA) risk assessment system.37 Zhong et al21 further explored the diagnostic value of OCTA image over CAD morbidity, and OCTA parameters singled out by LASSO regression combined with electrocardiogram and clinical characteristics achieved an AUROC of 0.897 on validation dataset using nomogram. However, lower retinal and choroidal vessel density were reported in hypertension,27,38 stroke,39 degenerative neural diseases, and diabetes.39 Thus, retinal microvascular change alone has limited specificity.

On the other hand, OCT studies found that retinal neural damage occurs with microvascular rarefication.28 The cerebrovascular disease can be identified for its distinct neurovascular alteration pattern on the retina.40 Hence, further exploration is required to uncover the specificity of OCTA on detecting CVDs fully.

Detection of Central Nervous System Diseases

Besides cerebrovascular diseases, degenerative neural diseases such as Alzheimer disease (AD), Parkinson disease (PD), and dementia correlate to retinal neurovascular changes as well. These changes include the loss of retinal ganglion cells (RGC) and thinning of retinal layers due to amyloid b-protein accumulation in AD patients, as demonstrated in many OCT studies.8,41 Furthermore, studies have shown that photoreceptors in the outer retina also decrease across degenerative neural disorders, including AD or non-AD dementia, PD, and mild cognitive impairment (CI).8,42 As such, the comparatively fast eye examination with high repeatability makes ocular images useful to detect degenerative neural diseases such as AD and CI during regular health examinations.43,44

With the small number of AD cases in the UK Biobank, the DL model built from retina images could distinguish AD patients from healthy controls using a feature selection method.45 Another method, hyperspectral retinal imaging (HSI), which uses a multi- spectral light source, could also identify the specific wavelength of light reflected from amyloids b-proteins.46 The DL model predicting AD from HSI images still rendered considerable accuracy with a limited sample size of 39 participants (17 AD patients vs 22 normal controls).47 Despite the small sample size, significant clinical connections and customized AI models could minimize this disadvantage. Wisely et al developed a multimodal ocular image convolutional neural network that included ultra- widefield color and fundus autofluorescence (UWF FAF), OCT ganglion cell-inner plexiform layer (GC-IPL) map, and OCTA modalities and performed well on the validation dataset considering the small sample size. So far, the multimodal input combined with patients’ records achieved the best performance with an AUROC at 0.841 in detecting AD patients.48

The number of children developing autism spectrum disorder (ASD) has increased in recent years.49 The most urgent task regarding ASD is the lack ofan objective screening method which leads to earlier intervention that betters one's quality of life.50 Findings of the significant difference in retinal nerve fiber layer (RNFL) thickness between high function ASD and other ASD patients provide a fresh insight into the distinguished neural structural development patterns.51,52 Hence, researchers built a OCT image - based DL model to extract image features capable of ASD prediction when combined with automatic features generated with automatic retinal image analysis (ARIA). The model performance on 10-fold cross-validation reached 95.7% sensitivity and 91.3% specificity.53 Comparing to the long observation periods for ASD diagnosis, AI-enhanced ocular image sheds new light on screening ASD in patients’ early years.

Detection of Endocrinological Disorder

Diabetic retinopathy (DR) is one of the leading causes of blindness worldwide. Studies on DR demonstrated that microscopic lesions already took place before clinical-stage DR appeared. In diabetic eyes without retinopathy, dilated retinal arterioles were reported,54 neural layer thinning was observed on OCT images, and RNFL, ganglion cell layer (GCL) thickness, SCP, and DCP decreased as the course of diabetes extended.55–57 The thinning of GCL occurred in type 1 diabetes patients even when blood glucose was well controlled, which suggests that ocular image holds accumulative information of diabetes regardless of hyperglycemia.58 Using an alternative noninvasive approach through ocular images to help screen early diabetes could increase population compliance as no blood draw is needed.

In a cross-sectional study focusing on detecting chronic diseases from CFP, a multitask DL model was built to distinguish patients with hypertension, hyperglycemia, and dyslipidemia while also predicting CVD risk factors. The model rendered sufficient performance on internal validation test sets, achieving AUROCs of 0.880 and 0.703 when predicting hyperglycemia and dyslipidemia, respectively.25 Another study used ML to predict diabetic eyes based on OCT and OCTA data such as vessel density, fractal dimension, and foveal avascular zone (FAZ), demonstrating that OCT and OCTA images could be helpful in the early diagnosis of diabetes.59 In recent years, metabolic syndrome was proposed as the complete pathophysiological process of dyslipidemia, obesity, hypertension, atherosclerosis, diabetes, stroke, and CAD.60,61 Longitudinal cohort studies confirmed that an improved metabolic condition by losing weight had significantly lowered insulin resistance in diabetic patients.62,63 Furthermore, an AI study found CFP capable of predicting hyperglycemia, hypertension, and dyslipidemia with the same CNN features.25 Therefore, regarding diabetes as part of the metabolic syndrome may also help identify diabetic patients at an early stage.

Besides DR, diabetic peripheral neuropathy (DPN) is another major complication that leads to disability, for example, diabetic foot ulcer and amputation, in diabetic patients.64 The DPN happens almost in parallel with DR during the progression of diabetes and is highly associated with DR stages.65 Efforts were made to screen DPN using retina vessel parameter calculated by ARIA (Singapore “I” Vessel Assessment, SIVA), and the logistic regression of vessel parameters resulted in AUROCs around 0.8 while vibration perception threshold (VPT), an examination for DPN diagnosis, was 0.941 in the same sample population.66 However, another study used a modified VGG-16 network to screen DPN from CFP images. The accuracy was roughly 0.1 higher than the logistic regression model.67 The results demonstrated the superiority of DL algorithms to ML models. Further clinical trials should explore whether ocular image-based AI can be an adjunct diagnostic tool for DPN.

Detection of Kidney Diseases and Renal Function

Other than DPN, DR stages strongly correlate with stages of diabetic kidney disease (DKD) as well.68 Diabetic patients with DKD are at high risk of developing DR and vice versa.69,70 The damaged renal function also affects macular structures as the intraretinal vascular leakage accumulates and forms macular edema (ME).70 The correlation between nondiabetic chronic kidney disease (CKD), clinically characterized as the decrease of estimated glomerular filtration rate (eGFR) and increase of urine albumin/creatinine ratio (UACR), and retinal macroscopic vessel parameters was significant yet inconsistent due to confounding factors.71 The renal function deterioration affects retinal structure through multiple pathways such as vascular endothelial dysfunction, chronic inflammation, and thrombosis state,72 and more comprehensive work is required to uncover these ocular changes. On the other hand, OCTA studies found retinal and choroidal microvascular rarefaction and macular thickening correlated to elevated UACR, a risk factor to endothelial dysfunction (ED).9,73 This finding is supported by another study assessing retinal vessel ED by flicker light – induced retinal vasodilation and DVA, which indicates simultaneous ED in the eye and kidney.74

Sabanayagam et al developed a DL algorithm to detect CKD (defined as eGFR less than 60 mL/min per 1.73 m2) from CFP images in the Singapore Epidemiology of Eye Diseases study database.75 The results showed that the retinal image – based model was comparable to the model built upon risk factors including age, gender, ethnicity, diabetes, and hypertension in both internal and external validation datasets. The performance increased in the SEED validation dataset when adding the retinal image to the risk factor model.

As CKD and diabetes are highly associated with the retina, a DL study tried to predict both diabetes and CKD. The study suggests that retinal images hold predictive values over early- stage CKD and highlight the shared pathophysiology among retina changes, diabetes, and CKD.12 Taken together, ocular image-based AI could monitor CKD patients from early stage to end-stage.

Detection of Hematological Diseases

It is known that anemia causes hypoxia in the eye, leading to increased venous tortuosity, chronic inflammation, and extravas- cular lesions such as hemorrhages, exudates, and cotton-wool spots. More often, hematological studies used eyelid conjunctiva images to predict anemia. However, conjunctiva images are rare and rely on arbitrary judgment. Mitani et al implemented DL algorithms on CFP to screen for anemia.76 Afterward, Rim et al14 applied VGG-16 on retinal images to predict systemic biomarkers, including hemoglobin. Zhang et al25 also used CFP to predict hematocrit and mean corpuscular hemoglobin concentration (MCHC), which further provided tools to identify anemia subtypes.

Recent OCT/OCTA studies discovered that children with iron deficiency anemia (IDA) had significantly lower retinal vessel density and lighter OCT reflectance of the vessel.77 Hence, an OCT study proposed an ML algorithm extracting texture features from OCT images to classify anemia.78 Furthermore, by building a DL network from OCT images, the AneNet reached a state-of-the-art performance; the accuracy in identifying anemia was 98.65%, and AUROC was 99.83% on test dataset.79 Although further testing to confirm generalizability is needed, the results still suggest the potential of scaling AI-based ocular imaging for anemia screening.

Detection of Hepatobiliary Disease

The hepatobiliary disease is a significant public health problem as the percentage of individuals with hepatitis B virus, hepatitis C virus, obesity, and type 2 diabetes mellitus (T2DM) remain high.80 Therefore, screening for the hepatobiliary disease remains essential. Ocular signs were observed in many hepatobiliary diseases like Kayser-Fleischer Ring in Wilson disease and yellow sclera during hyperbilirubinemia, suggesting that retinal images could provide a window for observing the effect of hepatobiliary disease on macro- and microvasculature and neural tissue. Zhang et al predicted total bilirubin (T-BIL) and direct bilirubin (D-BIL) from CFP images using a multitask DL model.25 Xiao et al22 built 6 DL models based on ResNet-101 to screen 6 major hepatobiliary diseases from slit-lamp external eye images and CFP images, achieving the best performance when detecting liver cancer and liver cirrhosis using slitlamp images. The DL model outperformed manual evaluation done by 6 ophthalmologists too. In addition, the researchers compared the modular information on the slitlamp images by applying a grey mask on specific areas such as the iris and conjunctiva and compared the diagnostic performance. The most prominent part of the image input was the conjunctiva, on which pathological changes such as jaundice and telangiectasis were more likely to be observed. Therefore, potential patients with liver cancer and liver cirrhosis will likely benefit from applying ocular image-based AI, and AI studies also contribute to the knowledge development of hepatobiliary diseases.

Prediction of Systemic Disease Prognosis

Ocular image alterations being regarded as an integrated risk factor suits clinical practice. The DL features can be translated to probability scales and function as a new risk factor representing the entire ocular image input at the last fully connected layer of the neural network. For predicting carotid artery atherosclerosis, researchers developed DL funduscopic atherosclerosis score (DL-FAS). The hazard ratio of higher tertile to lower tertile of the DL-FAS reached 8.33.30 Furthermore, Rim et al14 proposed a DL-based retinal CAC score (RetiCAC) system to stratify the longitudinal risk of developing CAC. The tertiles on the cumulative cardiovascular event separated from each other suggested possible retinal imaging biomarkers of developing CVDs.

In the study predicting both CKD and T2DM, a lower survival probability was found between the high-risk and other groups stratified from the DL predictor features. The authors also applied time-dependent ROC analysis among the risk groups. The development of future CKD and T2DM was both significantly higher in the high-risk group.12 Altogether, the scores generated from extracted features of the ocular image, similar to the imaging marker seen in cancer studies, can provide early insight into systemic diseases prognosis.

Prediction of Systemic Risk Factors

Predicting risk factors of CVDs including age,24,25,81,82 gender,25,81–83 smoking,25,84 BMI,82 waist-hip ratio,25 and dis- lipidemia25 using retinal images rendered sufficient accuracy, while detecting age from anterior segment ocular photo and corneal images85 were feasible too. Furthermore, with AI- enhanced quantification tool, traditional retinal parameters such as central retinal artery equivalent (CRAE) and central retinal vein equivalent (CRVE) were remeasured and found to have higher power predicting age, gender, mean arterial BP, BMI, and smoking behavior than the same parameters extracted manually,86 demonstrating the value of retinal parameters in AI studies.

Cerebral small vessel disease (SVD) is another cerebrovascular disease with high morbidity and is considered a cause of stroke.87 As a predisease status of SVD, stroke, cognitive disorder, dementia, and severe white matter hyperintensities (WMH) were predicted accurately using a DL-based approach on retinal images (sensitivity: 89.7%; specificity: 96.8%).88 Thus, using ocular input as a predictor of other risk factors can be an alternative to expensive diagnostic tests in detecting diseases and provide an additional reference for missing medical records in clinical practice.


The financial burden imposed on society due to medical examinations is gradually increasing.89 At the same time, the shortage of medical resources for extensive population-based screening remains a pressing matter as the diagnostic tools for major diseases like CVDs and degenerative neural diseases are complicated and unsuitable. All these issues culminate in a poor compliance rate for medical screening. As a noninvasive examination with high repeatability, AI-based ocular imaging could act as an alternative solution by minimizing the massive ocular image feature extraction workload and alleviate the burden of medical resources while maintaining high accuracy. Furthermore, the retinal image scoring systems were developed to integrate ocular image input as one risk factor for disease prognosis, and this process can evolve to locating imaging markers in the future while offering insights into systemic diseases development.

While AI-based ocular imaging presents exciting possibilities, there are still urgent problems to address. Firstly, the lack of explainability that many AI models present with must be mitigated especially for systemic diseases where their correlation with retinal characteristics have not been proven previously in clinical trials. Through the combination of prior clinical knowledge and ocular image data using ML methodology such as joint latent space.90 Sufficient transparency may increase the trust and reliability of AI models in clinical settings while reducing reporting bias and errors in correlations.

Secondly, the generalizability of systemic diseases’ prediction models must be ensured. Internal validation methods are not sufficient for both proof-of-concept and AI application studies. External validation datasets that contain real-world data is essential as it would be a better representation of societal demographics by bearing into consideration the differences in the distribution of race and ethnicity between the training and validation datasets. For example, retinal pigmentation differs among Asians and Europeans, as the loss of retina pigment cells and rod photoreceptors increases with aging.91 Hence the age prediction models trained on people with less retina pigmentation would have lower performance when tested in a population of people with more retina pigmentation.24,82 For future studies, training datasets with higher biodiversity and more resemblance to validation datasets may result in better generalizability. As data privacy and protection gradually become more stringent, external validation data will become increasingly harder to acquire. This has been demonstrated by the discrepancy in sample sizes when comparing OCT/OCTA AI studies to CFP studies. Although OCT/OCTA images have a smaller pixel matrix than CFP, the imaging interference is more extensive, requiring extra samples to reach superior generalization ability. Blockchain technology may be a solution for including more data without compromising on data privacy and protection in the future.92

Thirdly, some systemic diseases have similar effects on ocular structures such as the retina and choroid. Without identifying unique feature patterns for each disease, there is a risk for false-positive diagnosis in the event that several diagnostic tasks are given to the AI model in the same timeframe. Although some studies have attempted to integrate more systemic disease screening tasks into a universal AI system, further work is still needed to differentiate between different diagnostic tasks.

Lastly, to date, there have only been 2 AI devices approved by regulatory bodies for CVD and CKD risk stratifications,14,75 and there has also been a lack of specific clinical guidelines or evidence for the integration of AI devices into clinical settings. Getting regulatory approval and including the use of AI devices in clinical guidelines are important steps especially for physicians who are not as familiar with the use of AI systems to manage specific systemic disease conditions. These are barriers that must be overcome if the full integration of AI systems into clinical practice is to be warranted.


This systematic review has presented the current applications of AI on detecting systemic diseases from multimodal ocular images. However, there are some limitations that have to be acknowledged. AI studies evolve daily, and given the robust search strategy, this review may have missed out on cutting-edge research that did not undergo peer review or were published in abstract form. For example, a preprint study using multiple DL networks to screen type 2 diabetes was excluded because it did not undergo peer review and therefore possesses a high risk of bias.93 The definition of AI was restricted to the automatic frameworks that predict systemic diseases, and hence, studies using manual quantifications were excluded. By excluding review papers, going through reference lists to help find missed studies was no longer an option. A single assessor carried out the paper screening without cross-checking by other authors, and only including papers published in English may also leave out some new research (Supplementary Digital Content, Figure 1,


This review summarized the current results of ocular imagebased AI to screen systemic diseases. The wide variety of studies demonstrated the detection of diseases such as CVDs, CHD, and hematological diseases, along with the risk stratification of CVDs and CKD. The development and performance of ocular image - based AI models for large-scale screening and early detection potential when diagnosing specific diseases such as DPN and liver cancer may be promising, but more work should be done to evaluate the cost effectiveness in the real-world application.


1. Hay SI, Abajobir AA, Abate KH, et al. Global, regional, and national disability-adjusted life-years (DALYs) for 333 diseases and injuries and healthy life expectancy (HALE) for 195 countries and territories, 1990–2016: a systematic analysis for the Global Burden of Disease Study 2016. Lancet 2017; 390:1260–1344. doi:10.1016/S0140-6736(17)32130-X.
2. Wong TY, McGeechan K, Liew G, et al. Meta-analysis: retinal vessel caliber and risk for coronary heart disease. Ann Intern Med 2009; 151:404–413.
3. Cheung CY, Thomas GN, Tay W, et al. Retinal vascular fractal dimension and its relationship with cardiovascular and ocular risk factors. Am J Ophthalmol 2012; 154:663–674. e661. doi:10.1016/j.ajo.2012.04.016.
4. Wang J, Jiang J, Zhang Y, et al. Retinal and choroidal vascular changes in coronary heart disease: an optical coherence tomography angiography study. Biomed Opt Express 2019; 10:1532–1544. doi:10.1364/BOE.10.001532.
5. Cheung CY, Tay W, Ikram MK, et al. Retinal microvascular changes and risk of stroke: the Singapore Malay Eye Study. Stroke 2013; 44:2402–2408. doi:10.1161/STROKEAHA.113.001738.
6. Anjos R, Vieira L, Costa L, et al. Macular ganglion cell layer and peripapillary retinal nerve fibre layer thickness in patients with unilateral posterior cerebral artery ischaemic lesion: an optical coherence tomography study. Neuroophthalmology 2016; 40:8–15. doi:10.3109/01658107.2015.1122814.
7. Guo S, Yin S, Tse G, et al. Association between caliber of retinal vessels and cardiovascular disease: a systematic review and meta-analysis. Curr Atheroscler Rep 2020; 22:1–13.
8. Kim BJ, Irwin DJ, Song D, et al. Optical coherence tomography identifies outer retina thinning in frontotemporal degeneration. Neurology 2017; 89:1604–1611. doi:10.1212/WNL.0000000000004500.
9. Vadalà M, Castellucci M, Guarrasi G, et al. Retinal and choroidal vasculature changes associated with chronic kidney disease. Graefes Arch Clin Exp Ophthalmol 2019; 257:1687–1698. doi:10.1007/s00417-019-04358-3.
10. Yau JW, Xie J, Kawasaki R, et al. Retinal arteriolar narrowing and subsequent development of CKD Stage 3: the Multi-Ethnic Study of Atherosclerosis (MESA). Am J Kidney Dis 2011; 58:39–46. doi:10.1053/j.ajkd.2011.02.382.
11. Kawasaki R, Cheung N, Mosley T, et al. Retinal microvascular signs and 10-year risk of cerebral atrophy. Stroke 2010; 41:1826–1828. doi:10.1161/strokeaha.110.585042.
12. Zhang K, Liu X, Xu J, 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 2021; 5:533–545. doi:10.1038/s41551-021-00745-6.
13. Mitsuhara M, Fukui H, Sakashita Y, et al. Embedding human knowledge in deep neural network via attention map. arXiv preprint. arXiv:1905 2019; 03540.
14. Rim TH, Lee CJ, Tham YC, et al. Deep-learning-based cardiovascular risk stratification using coronary artery calcium scores predicted from retinal photographs. Lancet Digit Health 2021; 3:e306–e316. doi:10.1016/S2589-7500(21)00043-1.
15. Wu E, Wu K, Daneshjou R, et al. How medical AI devices are evaluated: limitations and recommendations from an analysis of FDA approvals. Nat Med 2021; 27:582–584. doi:10.1038/s41591-021-01312-x.
16. Queensland Health, Queensland Government. Hypertension and pregnancy. Queensland Clinical Guidelines. 2021. Available from
17. Brown MA, Magee LA, Kenny LC, et al. The hypertensive disorders of pregnancy: ISSHP classification, diagnosis, and management recommendations for international practice. Hypertension 2018; 72:24–43. doi:10.1161/HYPERTENSIONAHA.117.10803.
18. Huang D, Swanson EA, Lin CP, et al. Optical coherence tomography. Science 1991; 254:1178–1181. doi:10.1126/science.1957169.
19. Gao SS, Jia Y, Zhang M, et al. Optical coherence tomography angiography. Invest Ophthalmol Vis Sci 2016; 57:OCT27–36. doi:10.1167/iovs.15-19043.
20. Campbell JM, K M, Ding S, Carmody DP, Hakonsen SJ, Jadotte YT, White S, Munn Z. Diagnostic test accuracy: methods for systematic review and meta-analysis. Int J Evid Based Healthc 2015; 13:154–162.
21. Zhong P, Qin J, Li Z, et al. Development and validation of retinal vasculature nomogram in suspected angina due to coronary artery disease. J Atheroscler Thromb 2021; Publish ahead of print. doi:10.5551/jat.62059.
22. Xiao W, Huang X, Wang JH, et al. Screening and identifying hepatobiliary diseases through deep learning using ocular images: a prospective, multicentre study. Lancet Digit Health 2021; 3:e88–e97. doi:10.1016/S2589-7500(20)30288-0.
23. Parameswari C, Siva Ranjani S. Prediction of atherosclerosis pathology in retinal fundal images with machine learning approaches. J Ambient Intell Human Comput 2020; 12:6701–6711. doi:10.1007/s12652-020-02294-3.
24. Poplin R, Varadarajan AV, Blumer K, et al. Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. Nat Biomed Eng 2018; 2:158–164. doi:10.1038/s41551-018-0195-0.
25. Zhang L, Yuan M, An Z, et al. Prediction of hypertension, hyperglycemia and dyslipidemia from retinal fundus photographs via deep learning: A cross-sectional study of chronic diseases in central China. PLoS One 2020; 15:e0233166doi:10.1371/journal.pone.0233166.
26. Dai G, He W, Xu L, et al. Exploring the effect of hypertension on retinal microvasculature using deep learning on East Asian population. PLoS One 2020; 15:e0230111doi:10.1371/journal.pone.0230111.
27. Chua J, Chin CWL, Hong J, et al. Impact of hypertension on retinal capillary microvasculature using optical coherence tomographic angiography. J Hypertens 2019; 37:572–580. doi:10.1097/HJH.0000000000001916.
28. Peng Q, Hu Y, Peng M, et al. Retinal neurovascular impairment in patients with essential hypertension: an optical coherence tomography angiography study. Invest Ophthalmol Vis Sci 2020; 61:42doi:10.1167/iovs.61.8.42.
29. Lonardo A, Nascimbeni F, Mantovani A, Targher G. Hypertension, diabetes, atherosclerosis and NASH: Cause or consequence? J Hepatol 2018; 335–352.
30. Chang J, Ko A, Park SM, et al. Association of cardiovascular mortality and deep learning-funduscopic atherosclerosis score derived from retinal fundus images. Am J Ophthalmol 2020; 217:121–130. doi:10.1016/j.ajo.2020.03.027.
31. Son J, Shin JY, Chun EJ, et al. Predicting high coronary artery calcium score from retinal fundus images with deep learning algorithms. Transl Vis Sci Technol 2020; 9:28doi:10.1167/tvst.9.2.28.
32. Park HY, Park YG, Cho AH, Park CK. Transneuronal retrograde degeneration of the retinal ganglion cells in patients with cerebral infarction. Ophthalmology 2013; 120:1292–1299. doi:10.1016/j.ophtha.2012.11.021.
33. Lim G, Lin ZW, Xu D, et al. Feature isolation for hypothesis testing in retinal imaging: an ischemic stroke prediction case study. Proc AAAI Confer Artif Intell 2019; 33:9510–9515. doi:10.1609/aaai.v33i01.33019510.
34. Cordina R, Leaney J, Golzan M, et al. Ophthalmological consequences of cyanotic congenital heart disease: vascular parameters and nerve fibre layer. Clin Exp Ophthalmol 2015; 43.:115–123.
35. Li C, Zhong P, Yuan H, et al. Retinal microvasculature impairment in patients with congenital heart disease investigated by optical coherence tomography angiography. Clin Exp Ophthalmol 2020; 48:1219–1228. doi:10.1111/ceo.13846.
36. Ng WWY, Liang H, Peng Q, et al. An automatic framework for perioperative risks classification from retinal images of complex congenital heart disease patients. Int J Machine LearnCyber 2022; 13:471–483. doi:10.1007/s13042-021-01419-0.
37. Zhong P, Li Z, Lin Y, et al. Retinal microvasculature impairments in patients with coronary artery disease: An optical coherence tomography angiography study. Acta Ophthalmol 2022; 100:225–233. doi:10.1111/aos.14806.
38. Chua J, Chin CWL, Tan B, et al. Impact of systemic vascular risk factors on the choriocapillaris using optical coherence tomography angiography in patients with systemic hypertension. Sci Rep 2019; 9:5819doi:10.1038/s41598-019-41917-4.
39. Liu B, Hu Y, Ma G, et al. Reduced retinal microvascular perfusion in patients with stroke detected by optical coherence tomography angiography. Front Aging Neurosci 2021; 13:628336doi:10.3389/fnagi.2021.628336.
40. Zhang X, Xiao X, Liu C, et al. Optical coherence tomography angiography reveals distinct retinal structural and microvascular abnormalities in cerebrovascular disease. Front Neurosci 2020; 14:588515–1588515. doi:10.3389/fnins.2020.588515.
41. Lad EM, Mukherjee D, Stinnet SS, et al. Evaluation of inner retinal layers as biomarkers in mild cognitive impairment to moderate Alzheimer's disease. PLoS One 2018; 13:e0192646doi:10.1371/journal.pone.0192646.
42. Uchida A, Pillai JA, Bermel R, et al. Outer retinal assessment using spectral-domain optical coherence tomography in patients with Alzheimer's and Parkinson's disease. Invest Ophthalmol Vis Sci 2018; 59:2768–2777. doi:10.1167/iovs.17-23240.
43. Lee S, Jiang K, McIlmoyle B, et al. Amyloid beta immunoreactivity in the retinal ganglion cell layer of the alzheimer's eye. Front Neurosci 2020; 14doi:10.3389/fnins.2020.00758.
44. van de Kreeke JA, Nguyen HT, Konijnenberg E, et al. Optical coherence tomography angiography in preclinical Alzheimer's disease. Br J Ophthalmol 2020; 104:157doi:10.1136/bjophthalmol-2019-314127.
45. Tian J, Smith G, Guo H, et al. Modular machine learning for Alzheimer's disease classification from retinal vasculature. Sci Rep 2021; 11:238doi:10.1038/s41598-020-80312-2.
46. More SS, Beach JM, Vince R. Early detection of amyloidopathy in Alzheimer's mice by hyperspectral endoscopy. Invest Ophthalmol Vis Sci 2016; 57:3231–3238. doi:10.1167/iovs.15-17406.
47. Lemmens S, Van Craenendonck T, Van Eijgen J, et al. Combination of snapshot hyperspectral retinal imaging and optical coherence tomography to identify Alzheimer's disease patients. Alzheimers Res Ther 2020; 12:144doi:10.1186/s13195-020-00715-1.
48. Wisely CE, Wang D, Henao R, et al. Convolutional neural network to identify symptomatic Alzheimer's disease using multimodal retinal imaging. Br J Ophthalmol 2022; 106:388–395.
49. Seretopoulos K, Lamnisos D, Giannakou K. The epidemiology of autism spectrum disorder. Arch Hellenic Med 2020; 37:169–180.
50. Roman-Urrestarazu A, Yanez C, Lopez-Gari C, et al. Autism screening and conditional cash transfers in Chile: Using the Quantitative Checklist (Q- CHAT) for early autism detection in a low resource setting. Autism 2021; 25:932–945. doi:10.1177/1362361320972277.
51. Garcia-Medina JJ, Garcia-Pinero M, Del-Rio-Vellosillo M, et al. Comparison of foveal, macular, and peripapillary intraretinal thicknesses between autism spectrum disorder and neurotypical subjects. Invest Ophthalmol Vis Sci 2017; 58:5811–5818. doi:10.1167/iovs.17-22238.
52. Emberti Gialloreti L, Pardini M, Benassi F, et al. Reduction in retinal nerve fiber layer thickness in young adults with autism spectrum disorders. J Autism Dev Disord 2014; 44:873–882. doi:10.1007/s10803-013-1939-z.
53. Lai M, Lee J, Chiu S, et al. A machine learning approach for retinal images analysis as an objective screening method for children with autism spectrum disorder. EClinicalMedicine 2020; 28:100588doi:10.1016/j.eclinm.2020.100588.
54. Sabanayagam C, Tai ES, Lee J, et al. Retinal vessel caliber and peripheral neuropathy in diabetic participants. Microcirculation 2010; 17:297–302.
55. Shahidi AM, Sampson GP, Pritchard N, et al. Retinal nerve fibre layer thinning associated with diabetic peripheral neuropathy. Diabet Med 2012; 29:e106–e111.
56. Vujosevic S, Midena E. Retinal layers changes in human preclinical and early clinical diabetic retinopathy support early retinal neuronal and muller cells alterations. J Diab Res 2013; 2013:905058doi:10.1155/2013/905058.
57. Cao D, Yang D, Huang Z, et al. Optical coherence tomography angiography discerns preclinical diabetic retinopathy in eyes of patients with type 2 diabetes without clinical diabetic retinopathy. Acta Diabetol 2018; 55:469–477. doi:10.1007/s00592-018-1115-1.
58. Gerendas BS, Hatz K, Kaider A, et al. Ganglion cell layer thickening in well- controlled patients with type 1 diabetes: an early sign for diabetic retinopathy? Acta Ophthalmol 2020; 98:E292–E300. doi:10.1111/aos.14273.
59. Govindaswamy N, Ratra D, Dalan D, et al. Vascular changes precede tomographic changes in diabetic eyes without retinopathy and improve artificial intelligence diagnostics. J Biophotonics 2020; 13:e202000107doi:10.1002/jbio.202000107.
60. Mendrick DL, Diehl AM, Topor LS, et al. Metabolic syndrome and associated diseases: from the bench to the clinic. Toxicol Sci 2017; 162:36–42. doi:10.1093/toxsci/kfx233.
61. McCracken E, Monaghan M, Sreenivasan S. Pathophysiology of the metabolic syndrome. Clin Dermatol 2018; 36:14–20.
62. Lean ME, Leslie WS, Barnes AC, et al. Primary care-led weight management for remission of type 2 diabetes (DiRECT): an open-label, cluster-randomised trial. Lancet 2018; 391:541–551.
63. Lean ME, Leslie WS, Barnes AC, et al. Durability of a primary care-led weight-management intervention for remission of type 2 diabetes: 2-year results of the DiRECT open-label, cluster-randomised trial. Lancet Diabetes Endocrinol 2019; 7:344–355.
64. Margolis DJ, Jeffcoate W. Epidemiology of foot ulceration and amputation: can global variation be explained? Med Clin North Am 2013; 97:791–805.
65. Karvestedt L, Martensson E, Grill V, et al. Peripheral sensory neuropathy associates with micro-or macroangiopathy: results from a population-based study of type 2 diabetic patients in Sweden. DiabCare 2009; 32:317–322.
66. Hu G, Wu H, Kuang L, et al. Clinical study of diabetic peripheral neuropathy screening by retinal vascular geometric parameters. Sci Rep 2021; 11:6784doi:10.1038/s41598-021-85831-0.
67. Benson J, Estrada T, Burge M, Soliz P. Diabetic peripheral neuropathy risk assessment using digital fundus photographs and machine learning. Annu Int Conf IEEE Eng Med Biol Soc 2020; 1988–1991. doi:10.1109/EMBC44109.2020.9175982.
68. Grunwald JE, Alexander J, Ying GS, et al. Retinopathy and chronic kidney disease in the chronic renal insufficiency cohort (cric) study. Arch Ophthalmol 2012; 130:1136–1144. doi:10.1001/archophthalmol.2012.1800.
69. Xu X, Sun F, Wang Q, et al. Comprehensive retinal vascular measurements: a novel association with renal function in type 2 diabetic patients in China. Sci Rep 2020; 10:13737doi:10.1038/s41598-020-70408-0.
70. Zhuang X, Cao D, Yang D, et al. Association of diabetic retinopathy and diabetic macular oedema with renal function in southern Chinese patients with type 2 diabetes mellitus: a single-centre observational study. BMJ Open 2019; 9:e031194doi:10.1136/bmjopen-2019-031194.
71. Lim LS, Cheung CYL, Sabanayagam C, et al. Structural changes in the retinal microvasculature and renal function. Invest Ophthalmol Vis Sci 2013; 54:2970–2976. doi:10.1167/iovs.13-11941.
72. Satoh M. Endothelial dysfunction as an underlying pathophysiological condition of chronic kidney disease. Clin Exp Nephrol 2012; 16:518–521.
73. Zhuang X, Cao D, Zeng Y, et al. Associations between retinal microvasculature/microstructure and renal function in type 2 diabetes patients with early chronic kidney disease. Diabetes Res Clin Pract 2020; 168:108373.
74. Gunthner R, Hanssen H, Hauser C, et al. Impaired retinal vessel dilation predicts mortality in end-stage renal disease. Circ Res 2019; Publish ahead of print. doi:10.1161/CIRCRESAHA.118.314318.
75. Sabanayagam C, Xu D, Ting DSW, et al. A deep learning algorithm to detect chronic kidney disease from retinal photographs in community-based populations. Lancet Digit Health 2020; 2:e295–e302. doi:10.1016/S2589-7500(20)30063-7.
76. Mitani A, Huang A, Venugopalan S, et al. Detection of anaemia from retinal fundus images via deeplearning. Nat Biomed Eng 2020; 4:18–27. doi:10.1038/s41551-019-0487-z.
77. Korkmaz MF, Can ME, Kazanci EG. Effects of iron deficiency anemia on peripapillary and macular vessel density determined using optical coherence tomography angiography on children. Graefes Arch Clin Exp Ophthalmol 2020; 258:2059–2068.
78. Chen Z, Mo Y, Ouyang P, et al. Retinal vessel optical coherence tomography images for anemia screening. Med Biol Eng Comput 2019; 57:953–966. doi:10.1007/s11517-018-1927-8.
79. Wei H, Shen H, Li J, et al. AneNet: a lightweight network for the real-time anemia screening from retinal vessel optical coherence tomography images. Optics Laser Technol 2021; 136:106773doi:10.1016/j.optlastec.2020.106773.
80. da Rocha MC, Marinho RT, Rodrigues T. Mortality associated with hepatobiliary disease in Portugal between 2006 and 2012. GE Port J Gastroenterol 2018; 25:123–131. doi:10.1159/000484868.
81. Kim YD, Noh KJ, Byun SJ, et al. Effects of hypertension, diabetes, and smoking on age and sex prediction from retinal fundus images. Sci Rep 2020; 10:4623doi:10.1038/s41598-020-61519-9.
82. Rim TH, Lee G, Kim Y, et al. Prediction of systemic biomarkers from retinal photographs: development and validation of deep-learning algorithms. Lancet Digit Health 2020; 2:e526–e536. doi:10.1016/S2589-7500(20)30216-8.
83. Betzler BK, Yang HHS, Thakur S, et al. Gender prediction for a multiethnic population via deep learning across different retinal fundus photograph fields: retrospective cross-sectional study. JMIR Med Inform 2021; 9:e25165doi:10.2196/25165.
84. Vaghefi E, Yang S, Hill S, et al. Detection of smoking status from retinal images; a Convolutional Neural Network study. Sci Rep 2019; 9:7180doi:10.1038/s41598-019-43670-0.
85. Ma J, Xu X, Li M, et al. Predictive models of aging of the human eye based on ocular anterior segment morphology. J Biomed Inform 2021; 120:103855doi:10.1016/j.jbi.2021.103855.
86. Cheung CY, Xu D, Cheng CY, et al. A deep-learning system for the assessment of cardiovascular disease risk via the measurement of retinalvessel calibre. Nat Biomed Eng 2021; 5:498–508. doi:10.1038/s41551-020-00626-4.
87. Lau AY, Mok V, Lee J, et al. Retinal image analytics detects white matter hyperintensities in healthy adults. Ann Clin Transl Neurol 2019; 6:98–105. doi:10.1002/acn3.688.
88. Zee B, Wong Y, Lee J, et al. Machine-learning method for localization of cerebral white matter hyperintensities in healthy adults based on retinal images. Brain Commun 2021; 3:fcab124doi:10.1093/braincomms/fcab124.
89. Cunningham PJ. The Growing Financial Burden Of Health Care: National And State Trends, 2001–2006. Health Affairs 2010;29:1037–1044. doi:10.1377/hlthaff.2009.0493.
90. Diaz-Pinto A, Ravikumar N, Attar R, et al. Predicting myocardial infarction through retinal scans and minimal personal information. Nat Mach Intell 2022; 4:55–61. doi:10.1038/s42256-021-00427-7.
91. Campello L, Singh N, Advani J, et al. Aging of the retina: molecular and metabolic turbulences and potential interventions. Annu Rev Vis Sci 2021; 7:633–664. doi:10.1146/annurev-vision-100419-114940.
92. Tan TE, Anees A, Chen C, et al. Retinal photograph-based deep learning algorithms for myopia and a blockchain platform to facilitate artificial intelligence medical research: a retrospective multicohort study. Lancet Digit Health 2021; 3:e317–e329.
93. Yun JS, Kim J, Jung SH, et al. A deep learning model for screening type 2 diabetes from retinal photographs. Nutr Metabol Cardiovasc Dis 2022; In press. doi:10.1016/j.numecd.2022.01.010.

artificial intelligence; fundus photography; ocular image; public health; systemic disease

Supplemental Digital Content

Copyright © 2022 Asia-Pacific Academy of Ophthalmology. Published by Wolters Kluwer Health, Inc. on behalf of the Asia-Pacific Academy of Ophthalmology.