Introduction

Stroke is the second leading cause of death, and a leading cause of disability. Current estimates suggest that 15 million people suffer from stroke worldwide, with these resulting in 5 million deaths. The two main subtypes of stroke are ischemic (~80% of cases) and hemorrhagic stroke (~20% of cases) [1]. Crucially, it is estimated that up to 80% of strokes are preventable [2]. With that in mind, it is essential that those at risk are identified early so that effective primary prevention strategies can be implemented.

The retina is the only tissue that permits direct, non-invasive visualization of the microvasculature and central nervous system. Given the retina’s homology with the brain in terms of anatomy, physiology, and embryology, there is great potential for the retina to play a role in the prediction of cerebrovascular disease [3]. During routine clinical practice in both hospital and community (optometry) settings, optometrists and ophthalmologists can assess the retina in a direct, non-invasive, inexpensive manner through ocular imaging, including scanning laser ophthalmoscopy, fundus photography, optical coherence tomography (OCT), and OCT angiography (OCT-A) (Fig. 1). With OCT, scanning laser ophthalmoscopy, and fundus photography technology constantly develo** (and with increasingly widefield imaging available), opportunities to detect clinically useful retinal biomarkers reflecting stroke risk are now abundant. Further, artificial intelligence (AI)-enabled analysis of retinal imaging has been shown to predict stroke with good accuracy using machine learning approaches [4, 5]. AI has also proven capable of inferring a range of cardiovascular disease relevant factors from retinal images including blood pressure, smoking status, sex, and age—inferences impossible for human observers to make [6]. Combining modern imaging technologies with cutting-edge AI systems, the future for stroke prediction from retinal images appears bright.

Fig. 1
figure 1

A selection of retinal images which illustrate normal anatomy in optical coherence tomography (OCT), OCT angiography (OCT-A), and scanning laser ophthalmoscopy: a a structural OCT image of the macula; b a structural OCT with OCT-A flow overlay (highlighted by yellow dots, which indicate blood flow) and with automated segmentation at the internal limiting membrane (ILM), inner plexiform layer (IPL), outer plexiform layer (OPL) and Bruch’s membrane (BM); c an en face macula OCT-A image (superficial vascular plexus slab); d an en face macula OCT-A image (deep vascular plexus slab); e an en face macula OCT-A image (choriocapillaris slab); f a wide-field retinal image acquired using scanning laser ophthalmoscopy

Full size image

Whilst there is an extensive evidence synthesis literature base surrounding the role of stroke risk biomarkers including silent brain infarcts [7], proteinuria [8], pulse pressure [9], and carotid atherosclerosis [10], there is currently no systematic review detailing the relationship between retinal biomarkers and future stroke risk. With a view to summarizing the current evidence and opportunities in this domain, in this review, we explore the role of retinal biomarkers as indicators of future stroke risk and discuss the future of research in this area. To this end, this review 1) describes studies featuring AI-enabled methodologies to predict stroke; 2) summarizes the evidence surrounding the role of specific retinal biomarkers in stroke prediction; and 3) provides an assessment of the limitations and opportunities in this domain.

Methods

Protocol registration

This review was prospectively registered to PROSPERO [11]. The registration number for this review is CRD42023389223.

Literature search strategy

A systematic search was carried out on 23/12/2022 using the PubMed, Embase, and Web of Science databases. For the search, two search terms were used. Term 1 was ‘stroke’, ‘CVA’, or ‘cerebrovascular accident’ and Term 2 was ‘fundus photograph’, ‘fundus image’, ‘fundus autofluorescence’, ‘retinal image’, ‘retinal photograph’, ‘retinal autofluorescence’, ‘OCT’, ‘optical coherence tomography’, ‘OCTA’, ‘OCT A’, ‘adaptive optics’, ‘fluorescein angiography’, ‘optical imaging’, or ‘optical image’. Our search included human species and English language filters. On 21/09/2023, we repeated our search to ensure that this review was up to date. The search strategy is summarized in Fig. 2.

Fig. 2
figure 2

A PRISMA flow diagram outlining the exclusion/inclusion process

Full size image

Inclusion and exclusion criteria

Articles eligible to be included in this review were required to meet the following criteria:

  1. 1.

    The article discussed the role of retinal biomarkers as indicators of stroke risk.

  2. 2.

    The study was original and prospective or used data from biobanks involving prospective recruitment.

  3. 3.

    The study was conducted using human subjects.

  4. 4.

    The article was written in English.

Articles meeting the following criteria were excluded from our review:

  1. 1.

    Animal studies.

  2. 2.

    Non-original articles.

  3. 3.

    Case reports.

  4. 4.

    Conference abstracts.

  5. 5.

    Duplicates.

  6. 6.

    Preprint studies.

  7. 7.

    Articles which could not be obtained in full.

  8. 8.

    Studies which observed the relationship between established stroke and retinal biomarkers.

  9. 9.

    Studies observing the relationship between specific retinal diseases and stroke risk, unless providing individual biomarkers / retinopathy grades which can be ascertained using retinal imaging.

The reference lists of included articles were scanned for further articles which may fall within the scope of this review. All abstracts were screened by two authors independently (AS and ZG) who were blinded to each other's decisions. Conflicts were resolved by consensus between three authors (THJ, ZG, AS). All papers retained during abstract screening were screened in full by one author, with consensus discussions for any papers in which an inclusion/exclusion decision was not clear. Where a study was not available in an open access format, access was sought through three institutional libraries. Two recently published studies were identified outside of the search strategy and had come to the attention of the authors due to the high impact nature of the publications.

Data collection

Data were extracted from each study included: population size; demographics; stroke subtype; ocular imaging used; software for image analysis used; retinal features; and compliance with STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) and Advised Protocol for OCT Study Terminology and Elements (APOSTEL) reporting guidelines[12]. All included studies were reviewed in full by ≧2 authors prior to documentation of their findings.

Synthesis of results

For most measures, meta-analysis was precluded by overlap** populations, heterogeneous retinal biomarker assessment, and heterogeneous populations. Where meta-analysis was performed, the calculations were conducted in R version 4.2.2. using the meta library version 6.5–0 [13]. The generic inverse variance method was used to produce pooled hazard ratios (HR), 95% confidence intervals (CI), and p values. Values reported in text are those derived from the random effects model although fixed effects estimates are also reported in figures. The restricted maximum-likelihood estimator for tau2 was used to identify heterogeneity. Forest plots are provided (Figs. 3, 4, 5). Our study is presented in accordance with the PRISMA guidelines.

Assessment of bias

STROBE recommendations were used to assess the observational studies included in this review [12]. APOSTEL recommendations were used to assess the reporting of OCT results of studies included in this review [14].

Compliance with ethical guidelines

This article is based upon previously published studies. There are no ethical concerns with regards to this study.

Results

Study Characteristics

The literature search heralded 3,097 results, of which 22 were eligible for inclusion. Two additional eligible studies were identified outside of the search strategy. Therefore, a total of 24 articles were included in this review (Figure 2). The included papers pooled hemorrhagic and ischemic stroke into a single outcome unless otherwise specified.

Artificial intelligence (AI) enabled techniques to predict stroke

AI has enabled efficient analysis of enormous datasets, allowing for well-powered analyses; and the nature of models developed are becoming increasingly sophisticated in terms of the types of retinal variables being considered. In the domain of stroke prediction, there are currently three published AI-enabled prediction models, which illuminate the enormous potential for AI technologies to revolutionize this area of research.

The first of these is a methodologically robust study authored by Rudnicka et al., who utilized AI-enabled retinal vasculometry (developed using a mixture of supervised machine learning and deep learning) for the purpose of predicting stroke incidence and mortality [5]. In this study, the model was developed in a UK Biobank (UKBB) cohort (n = 88,052), and externally validated in the European Prospective Investigation into Cancer (EPIC) Norfolk cohort (n = 7,411). The external validation in this study is key and assessed if the models decisions are trustworthy and consistent across the diversity of datasets. QUARTZ was implemented as an automated solution for the processing of a substantial number of retinal images, enabling the extraction of quantitative vascular measurements [15]. The authors compared three models for the prediction of stroke and stroke mortality: 1) the Framingham risk score (FRS), 2) a combination of age, smoking status, medical history, and retinal vasculometry (RV), 3) a combination of FRS and RV. The addition of RV to the FRS model did not improve model performance, but the model combining age, smoking status, medical history, and RV showed similar performance to FRS with C-statistic of 0.73 in men and 0.75 in women (compared with 0.74 for both men and women with FRS). There was a comparable drop in performance during external validation across models. For circulatory mortality, the model inclusive of age, smoking, medical history, and RV performed well in men and women both in UKBB and EPIC Norfolk (C-statistics ranging from 0.75 to 0.77). This study demonstrates potential for AI-enabled retinal image analysis considering numerous retinal traits concurrently to be used for the prediction of stroke but does not outperform traditional risk scores.

Zhu et al. developed a deep learning-based model which estimated a subject’s age based upon their retinal photograph, and then utilized the ‘retinal age gap’ to assess risk of stroke [16]. In order to develop an algorithm to predict age, the authors initially trained a Xception deep learning model (a deep convolutional neural network architecture) [17] on UKBB subjects without any medical history (n = 11,052). Strong correlation was achieved between predicted age and chronological age in the healthy subject dataset (0.81, p <0.001). The study then assessed the role of the retinal age gap (defined as retina determined age minus true chronological age) in the prediction of future stroke in those who were stroke-free at baseline (n = 35,304). In a confounder adjusted model, each one-year increase in the retinal age gap was associated with a 4% increase in the risk of stroke (HR:1.04, 95% CI 1.00–1.08, p = 0.03). Furthermore, compared to participants with retinal age gap in the first quintile, participants with retinal age gap in the fifth quintile had significantly higher risks of stroke events (HR: 2.37, 95% CI 1.37–4.10, p = 0.002), whilst those in other quintiles did not differ significantly. The predictive performance for 10 year stroke risk assessed by the FRS was compared with the predictive value of a retinal age-based model adjusted for cardiovascular risk-related covariates. The AUCs of the retinal age-based model (0.68, 95% CI 0.64–0.71) were marginally higher than that of the FRS (0.66, 95% CI 0.63–0.69); however, this difference was not significant (p = 0.51). A notable weakness of this study was that the findings were not externally validated, and so it was unclear to what extent the model would generalize.

The most recently published study in this area was undertaken by Zhou et al. [18]. The study had a broad focus, using fundus photographs and OCTs for a range of detection and prediction tasks for both ocular and systemic diseases. The authors developed their foundation model, ‘RETFound’ (a masked autoencoder), using 904,170 color fundus photographs and 736,442 OCT images derived primarily from the Moorfields Diabetic imAge dataSet (MEH-MIDAS, n = 37,401, providing 90.2% of fundus photographs and 85.2% of OCTs). They utilized the RETFound encoder (from a masked autoencoder) and a multilayer perceptron for the three-year prediction of systemic disease, such as stroke, myocardial infarction, and heart failure. To test the performance of RETFound for the prediction of systemic disease, the authors utilized the data of subjects from the AlzEye study (n = 353,157, inclusive of n = 1,263 stroke subjects), with external validation in UKBB (n = 82,885, inclusive of 154 stroke subjects) [19, 20]. Although RETFound had impressive performance in some areas (such as detection of ocular disease), the accuracy of the model for the three-year prediction of stroke was limited during external validation, with an AUROC of 0.75 in AlzEye and 0.59 in UKBB when using fundus photographs and 0.75 in AlzEye and 0.56 in UKBB when using OCT. The strengths of the work are the utilization of a range of cohorts, inclusion of ethnically diverse subjects, the large sample size, use of external validation, and the breadth of tasks which RETFound can be utilized for. Key weaknesses include the relative geographic homogeneity of the cohorts (given they are entirely derived from the UK), the development of the autoencoder in principally diabetic cohort, and the limited performance for stroke prediction.

The above studies have illustrated the potential for AI-enabled retinal image analysis to be used for prediction of stroke. At present, these models do not offer a clear advantage over conventional stroke risk scores with regard to predictive performance, but prove that retinal image analysis has the scope to predict cerebrovascular events.

Individual retinal features as indicators of future stroke risk

For the most part, the included studies observed the relationship between individual retinal traits and stroke risk, rather than considering several traits together within a predictive model. Whilst the future of retinal imaging enabled stroke prediction will be founded in more complex models considering numerous retinal characteristics, the following findings are interesting given that they prove the potential of the retina as an indicator of stroke risk, and demonstrate which retinal traits confer the most information about cerebrovascular health.

Vessel calibers and arterio-venous ratio

The calibers of retinal vessels are influenced by numerous systemic factors relevant to stroke, and reflect the health of the vasculature. As such, retinal arteriolar and venular calibers, and a measure of their relative diameters known as the arterio-venous ratio (AVR), have been explored as predictors of stroke. The venules are typically measured in the form of the ‘central retinal vein equivalent’ (CRVE), whilst the arterioles are measured in the form of the ‘central retinal artery equivalent’ (CRAE) [21].

Three studies were sufficiently similar to allow meta-analysis of the relationship between CRVE and CRAE and stroke risk (n = 12,919)[22,45,46,47].

Risk of Bias

STROBE guidelines were used to assess the reporting of observational studies according to accepted standards, with the maximum possible score being 22 [12]. Reporting standards were generally high in papers included in this review, with STROBE scores ranging from 17 to 22 (Table 1). Reporting in studies utilizing OCT was assessed against the APOSTEL recommendations (Table 2) [14]. Whilst study compliance with APOSTEL recommendations was incomplete, the biobanks utilized have clearly protocolled methodologies.

Table 1 STROBE compliance checklist 
Full size table
Table 2 APOSTEL criteria compliance checklist 
Full size table

Conclusions

The current state of evidence

Although the literature surrounding the role of retinal image derived biomarkers for the assessment of stroke risk has been develo** for over two decades, this field in many ways remains in its infancy, with OCT and OCT-A biomarkers having received perishingly little attention. The published evidence illustrates that key individual retinal traits related to increasing stroke risk include wider CRVE, narrower CRAE, lower fractal dimension, increased arterial tortuosity, retinal emboli, and the presence of features of retinopathy. To a lesser extent, there is evidence that the presence of individual features of retinopathy, such as microaneurysms, AVN, and focal arteriolar narrowing, may predict stroke, although further studies in these areas are required given the conflicting evidence surrounding CRAE, and limited number of publications exploring individual retinopathy features.

The future of stroke prediction from retinal imaging will not be founded in the exploration of individual retinal features and their relationship to future stroke risk, but in construction of sophisticated models considering numerous retinal features. Early efforts toward this have been made in three studies which utilized sophisticated, AI-enabled analysis to demonstrate the opportunity for retinal images to be used to predict stroke [5, 16, 18]. Whilst current models do not perform sufficiently well to be rolled out into clinical practice, further work in this area could lead to cost-effective, quick, and accessible primary prevention programs based on retinal imaging.

Opportunities and limitations

Although there is an enormous opportunity in this domain, there are numerous challenges which must be addressed before retinal imaging can be utilized to predict stroke in clinical practice.

First, the current models generally focus on selection of individual hypothesis-driven retinal parameters, rather than considering the retinal image in its entirety in a hypothesis free manner. Whilst this allows clear biological interpretability, it comes at the sacrifice of clinically useful data. It is understood that deep learning algorithms are capable of identifying clinically relevant patterns in images that cannot be appreciated by human investigators, and as such future research in this area should seek to analyze the full image using deep learning techniques, rather than isolating selected traits for consideration.

Second, OCT imaging is currently underexplored. Given the wealth of information conveyed by an OCT in terms of neuronal and vascular health, this is an area which is currently receiving insufficient attention. Particularly interesting imaging modalities include swept source OCT because of the improved visualization of the choroid and OCT-A which enables clear visualization of the microcirculation of the retina and choroid. Notably, no study included in this review explored the role of the choroidal vasculature in the prediction of stroke despite the recognized relationship between choroidal morphology and cardiovascular disease [48].

Careful external validation is essential for future papers modeling stroke prediction from retinal images. None of the AI-enabled predictive models evaluated in this review were tested in cohorts exclusively constituent patients with ocular conditions, such as cataract, age-related macular degeneration (AMD), or retinopathy. In the domain of MI prediction, it has been shown that the presence of AMD degrades predictive model performance, with decreasing performance with increasing grade of AMD [49]. This is a substantial problem, given that those at the highest risk of stroke are the elderly, who are also the most likely to suffer from the above ophthalmic diseases. Further, many studies identified in this review rely on cohorts which are not diverse with regard to age, ethnicity, or comorbidities, likely causing algorithmic bias in the developed models. This problem has been discussed at length in previous research, and it is essential that future research considers effective bias mitigation strategies and representative study design [50]. With this in mind, future work must evaluate how well models perform across diverse patient subgroups.

Finally, the included studies typically combined ischemic and hemorrhagic stroke into a single outcome. Whilst this increases power, it is likely that the degree to which retinal biomarkers predict different forms of stroke will vary, and as such it would be interesting for future research to aim to predict distinct stroke phenotypes. Most importantly, future studies should segregate stroke events into hemorrhagic and ischemic groups and observe the differences with regard to the accuracy of retinal imaging-informed predictions, and the specific retinal phenotypes of relevance to the risk of stroke events.

In conclusion, our review highlights early work that has revealed potential for retinal imaging to be used to predict stroke. In this review, we have highlighted the current evidence suggesting retinal features are linked to future stroke risk, limitations of this work, and opportunities for future researchers​​ to build on the existing evidence to produce models which are suitable for clinical practice. Although the currently available models for stroke prediction are unready for implementation in clinical practice, there is clear scope for clinical utility once the aforementioned limitations have been addressed. With development, a future in which your annual ophthalmic/optometric appointment could serve to analyze your cerebrovascular health in addition to your ocular health is within sight.