Retina as a window to systemic hypertension: From hypertensive retinopathy to artificial intelligence-driven microvascular biomarkers

Abstract

The retinal microvasculature is the only vascular bed in the human body that can be visualised directly, non-invasively and repeatedly in vivo. Because retinal arterioles share embryological origin, calibre and autoregulatory physiology with the cerebral and coronary microcirculation, hypertension-induced retinal changes offer a unique window onto systemic small-vessel disease. For this narrative minireview, PubMed/MEDLINE, EMBASE, Scopus, Web of Science and IEEE Xplore were searched from database inception to April 30, 2026 for peer-reviewed English-language human studies of hypertensive retinopathy, quantitative retinal vascular biomarkers, multimodal retinal imaging and artificial intelligence (AI) applied to fundus images; population-based cohorts, individual-participant meta-analyses and AI studies reporting external validation were prioritised, and evidence was synthesised thematically and tabulated for cross-study comparison. Classical hypertensive retinopathy, graded for over eight decades by the Keith-Wagener-Barker and Scheie systems, has been supplemented by quantitative measurement of retinal vessel calibre and by geometric parameters such as fractal dimension and tortuosity. Large population cohorts have established that these parameters predict incident hypertension, stroke, coronary heart disease and mortality independently of conventional risk factors. Optical coherence tomography (OCT), OCT angiography and adaptive optics have extended assessment to the capillary and cellular level, while deep learning has enabled automated retinopathy grading and direct inference of blood pressure and cardiovascular risk from fundus photographs. Critically, the high accuracies reported on small private datasets are not reproduced on the first public benchmark, where the best algorithms achieved only fair agreement, and no randomised evidence yet shows that acting on a retina-derived biomarker improves outcomes. Clinical translation therefore depends on measurement standardisation, prospective multicentre and multi-ethnic external validation, explainable models, defined referral pathways from an abnormal retinal screen to blood pressure confirmation and cardiovascular risk management, and clarity on the regulatory status of software that infers systemic disease from ocular images.

Key Words: Hypertensive retinopathy; Retinal vessels; Microvascular biomarkers; Optical coherence tomography angiography; Deep learning; Artificial intelligence; Cardiovascular risk

Core Tip: The retina offers a directly visualisable model of the systemic microcirculation, and the integration of quantitative vascular metrics with deep-learning analysis of fundus images now allows blood pressure and cardiovascular risk to be inferred non-invasively at population scale. Realising this potential requires standardisation of measurement, validation across ethnicities and prospective demonstration that retina-derived biomarkers improve clinical decisions beyond established risk scores.

INTRODUCTION

Systemic arterial hypertension is the single largest modifiable contributor to global death and disability, and its damage to small vessels precedes and predicts events in the brain, heart and kidney. The eye is unique in allowing the small vessels themselves to be inspected. Through the transparent ocular media, the retinal arterioles, venules and capillaries can be photographed and quantified non-invasively, providing what amounts to an in vivo biopsy of the systemic microcirculation. Because retinal and cerebral vessels share embryological and anatomical characteristics, including a blood-tissue barrier and the absence of autonomic innervation distal to the central vessels, they exhibit comparable responses to elevated blood pressure.

The clinical study of these responses dates to the nineteenth century but was formalised in 1939, when Keith, Wagener and Barker classified essential hypertension into four groups by ophthalmoscopic appearance and demonstrated a steep gradient in prognosis: In their series of grade IV (malignant) patients, defined by papilloedema, 79% were dead within one year of diagnosis and the five-year survival rate was under 1%[1]. In 1953, Scheie[2] refined the description by grading hypertensive and arteriolosclerotic changes separately, distinguishing acute blood-pressure-related signs from chronic arteriolar wall remodelling. These systems remain in clinical use, but their limitations - poor interobserver reproducibility, weak correlation of early signs with blood pressure, confounding by age-related arteriolosclerosis, and recognition of choroidopathy and optic neuropathy as separate entities[3] - prompted the development of the simplified, prognostically anchored Wong-Mitchell grading in 2004[4] and, in parallel, of quantitative computer-assisted vessel measurement[5]. The trajectory from subjective ophthalmoscopy to objective quantification, and now to artificial intelligence (AI)-driven inference, defines the modern field of oculomics applied to hypertension[6].

Knowledge gap and rationale for this review

Several authoritative syntheses have addressed components of this field. Cheung et al[3] reviewed the retinal microvasculature as a model of hypertensive small-vessel disease; Wong and Mitchell[4] established the clinical grading framework; Wagner et al[6] introduced the concept of oculomics; and Cheung et al[7] provided a comprehensive primer on hypertensive eye disease. These works, however, either predate or address only in passing three developments that now dominate the field. First, the release of the first publicly available, expert-annotated benchmark for hypertension and hypertensive retinopathy classification has permitted, for the first time, a like-for-like comparison of competing algorithms, and the results are markedly less favourable than those reported on private datasets[8]. Second, task-specific convolutional networks are being displaced by self-supervised retinal foundation models trained on more than a million unlabelled images, which change the data economics of the field[9]. Third, accumulated regulatory and implementation experience from autonomous AI screening in diabetic retinopathy has clarified the evidentiary and workflow requirements that retina-derived hypertension biomarkers will have to meet[10,11].

The present minireview is therefore differentiated from earlier syntheses in three respects. First, rather than listing reported accuracies, it deliberately juxtaposes performance obtained on small curated datasets with performance on a common public benchmark, and treats the discrepancy as the central methodological problem of the field. Second, it maps each candidate biomarker to the imaging modality that generates it, with an explicit comparison of cost, accessibility and clinical readiness, so that readers can judge which biomarkers are deployable now and which remain research tools. Third, it foregrounds implementation - who acquires the image, who is accountable for the algorithmic output, and what referral pathway an abnormal result should trigger - rather than treating clinical translation as a closing caveat.

The objectives of this review are to: (1) Summarise the pathophysiology and contemporary grading of hypertensive retinopathy; (2) Appraise the epidemiological evidence linking quantitative retinal vascular parameters to incident hypertension and cardiovascular outcomes; (3) Compare the multimodal imaging platforms from which these biomarkers derive; (4) Critically evaluate AI approaches to retinal hypertension biomarker discovery, with particular attention to reproducibility and external validation; and (5) Define the standardisation, regulatory and workflow requirements for clinical translation, together with a research agenda.

LITERATURE SEARCH STRATEGY

This article is a narrative minireview and was not registered as a systematic review; nonetheless, a structured and reproducible search was performed in order to limit selection bias. PubMed/MEDLINE, EMBASE, Scopus, Web of Science and IEEE Xplore were searched from database inception to 30 April 2026. Search strings combined controlled vocabulary and free-text terms across three concept blocks using Boolean operators. Block 1 (disease): Hypertension, blood pressure, hypertensive retinopathy, hypertensive eye disease. Block 2 (retinal assessment): Retina, retinal vessel calibre, fundus photography, arteriolar-to-venular ratio (AVR), central retinal arteriolar equivalent (CRAE), fractal dimension, tortuosity, optical coherence tomography (OCT), OCT angiography (OCTA), adaptive optics. Block 3 (computational methods): AI, deep learning, machine learning, convolutional neural network (CNN), vision transformer, foundation model, oculomics. Blocks 1 and 2 were combined for the clinical and epidemiological sections, and all three blocks for the AI sections.

Eligible sources were peer-reviewed, English-language studies in human participants that reported quantitative retinal vascular, structural or algorithmic data in relation to blood pressure, hypertensive retinopathy or cardiovascular outcomes. Population-based cohorts, individual-participant meta-analyses, systematic reviews and AI studies reporting external or independent validation were prioritised; landmark historical papers were retained irrespective of date because they define the grading systems still in clinical use. Excluded were conference abstracts without a full-text report, non-peer-reviewed preprints, animal and ex vivo studies, and single case reports, with the exception of one adaptive-optics case retained for illustrative purposes because comparable in vivo wall-to-lumen data in malignant hypertension are not otherwise available. Titles and abstracts were screened independently by two authors, full texts of potentially eligible records were assessed against the same criteria, and disagreements were resolved by discussion with a third author. Reference lists of included articles and of prior reviews were hand-searched to identify additional sources.

Because of substantial heterogeneity in imaging protocols, grading reference standards, biomarker definitions and outcome ascertainment, quantitative pooling was neither appropriate nor attempted. Evidence was synthesised narratively under four thematic domains - pathophysiology and grading, quantitative vascular parameters, imaging modalities, and AI - with cross-study comparison presented in tabulated form (Tables 1, 2, and 3). Reported effect estimates are given as published, with the corresponding population and adjustment set, so that readers can judge comparability directly.

Table 1 Landmark epidemiological studies of quantitative retinal vascular parameters and cardiovascular outcomes.

Ref.	Cohort	Design and sample	Retinal parameter	Outcome	Principal finding
Sharrett et al[19], 1999	ARIC	Population-based cross-sectional, middle-aged adults	CRAE	Blood pressure	Mean arterial pressure was the dominant systemic determinant of narrower CRAE
Wong et al[54], 2004	ARIC	Prospective cohort	AVR	Incident hypertension	Smaller AVR predicted incident hypertension over 3 years
Ding et al[20], 2014	IPD meta-analysis, 6 cohorts	Individual-participant meta-analysis	CRAE, CRVE	Incident hypertension	Narrower arterioles and wider venules independently predicted incident hypertension
Kawasaki et al[21], 2009	MESA	Multi-ethnic prospective cohort	CRAE	Incident hypertension	OR = 1.34 (95%CI: 1.18-1.52) per 1-SD decrease in CRAE
Smith et al[22], 2004	Blue Mountains Eye Study	Population-based prospective cohort	Arteriolar calibre	5-year incident severe hypertension	Arteriolar narrowing predicted incident severe hypertension
Wong et al[23], 2002	ARIC	Prospective cohort	AVR	Incident coronary heart disease	Association present in women, absent in men
Seidelmann et al[24], 2016	ARIC	Long-term prospective cohort	CRAE, CRVE	Stroke, heart failure, mortality	Vessel calibres predicted long-term ischaemic stroke, heart failure and all-cause mortality
McGeechan et al[25], 2009	IPD meta-analysis	20798 participants, 945 stroke events	CRVE, CRAE	Incident stroke	CRVE: HR = 1.15 (95%CI: 1.05-1.25) per 20 μm; CRAE: HR = 1.00 (95%CI: 0.92-1.08)
McGeechan et al[26], 2009	IPD meta-analysis	Pooled cohorts	CRAE, CRVE	Coronary heart disease	HR = approximately 1.16-1.17 per 20 µm in women; no association in men
Liew et al[29], 2021	-	Population-based	Fractal dimension	Stroke mortality	Stroke mortality 7.7% in lowest vs 1.3% in highest fractal-dimension quartile
Triantafyllou et al[55], 2013	-	Cross-sectional, ABPM sub-phenotypes	AVR	Hypertension phenotype	AVR 0.820 normotension, 0.739 white-coat, 0.736 sustained, 0.716 masked hypertension

ARIC: Atherosclerosis Risk in Communities; MESA: Multi-ethnic study of atherosclerosis; CRAE: Central retinal arteriolar equivalent; CRVE: Central retinal venular equivalent; AVR: Arteriolar-to-venular ratio; IPD: Individual participant data; ABPM: Ambulatory blood pressure monitoring.

Table 2 Comparison of retinal imaging modalities for hypertension-related microvascular assessment.

Modality	Principal biomarkers	Resolution/depth	Relative cost	Accessibility	Acquisition burden	Clinical readiness
Non-mydriatic fundus photography	HR grade; CRAE, CRVE, AVR; fractal dimension, tortuosity, branching angle; AI-derived BP and CVD risk	Surface, approximately 10 µm lateral	Low (portable and smartphone systems available)	High; already embedded in national DR screening programmes	Seconds; no dilation; non-specialist operator	Deployable now; the only modality realistically scalable to population screening
Optical coherence tomography	Peripapillary RNFL and macular GC-IPL thickness	Cross-sectional, approximately 5 µm axial	Moderate to high	Widespread in ophthalmology; rare in primary care	1-2 minute; trained operator	Established in ophthalmic practice; not a hypertension screening tool
OCTA	Superficial and deep capillary vessel density; FAZ area; choriocapillaris flow voids	Capillary plexus level, depth-resolved	High	Specialist ophthalmology centres	1-3 minute; motion artefact common; trained operator	Research and specialist use; device-dependent metrics preclude screening
Adaptive optics SLO	Arteriolar wall thickness; wall-to-lumen ratio	Cellular, minute 2 µm	Very high	Few research centres worldwide	Prolonged; expert operator; small field of view	Research tool only

HR: Hypertensive retinopathy; CRAE: Central retinal arteriolar equivalent; CRVE: Central retinal venular equivalent; AVR: Arteriolar-to-venular ratio; BP: Blood pressure; CVD: Cardiovascular disease; DR: Diabetic retinopathy; RNFL: Retinal nerve fibre layer; GC-IPL: Ganglion cell-inner plexiform layer; FAZ: Foveal avascular zone; SLO: Scanning laser ophthalmoscopy; AI: Artificial intelligence; OCTA: Optical coherence tomography angiography.

Table 3 Comparison of key artificial intelligence studies using retinal images in hypertension and cardiovascular risk.

Ref.	Dataset (size, provenance)	Task	Architecture	Reported performance	External validation	Potential clinical application
Abbas et al[41], 2021	1400 graded images, private	5-stage HR classification	DenseNet with semantic/instance segmentation	Sensitivity 90.5%, specificity 91.5%, accuracy 92.6%, AUC 0.915	No	HR grading support
Qureshi et al[42], 2022	Private, small	HR detection	Depth-wise separable CNN + SVM	Sensitivity 94%, specificity 96%, accuracy 95%, AUC 0.96	No	HR detection
Bhimavarapu et al[43], 2024	ODIR, INSPIREVR, VICAVR (augmented, 1200 images)	5-class HR severity	Improved CNN + improved SVM	Accuracy up to 98.99%	No	HR severity grading
Suman et al[44], 2025	HRSG, expert-annotated, private	4-class HR severity	ResNet-50 + modified ViT (locality self-attention), decoupled classifier	Accuracy 0.9688, sensitivity 0.9435, specificity 0.9766	No	HR severity grading
Qian et al[8], 2025 (HRDC challenge)	1000 images per task, public benchmark	Hypertension and HR classification	Multiple submitted algorithms	Best Kappa 0.3819 (hypertension), 0.4154 (HR)	Yes (held-out test set)	Benchmarking; reveals reproducibility gap
Poplin et al[45], 2018	284335 patients; independent validation sets	Prediction of systolic BP, sex, MACE	Inception-v3 CNN	SBP MAE 11.23 mmHg; sex AUC 0.97; MACE AUC 0.70 (95%CI: 0.648-0.740) vs SCORE 0.72	Yes	Population risk stratification
Cheung et al[47], 2021 (SIVA-DLS)	> 70000 images, multi-ethnic, multi-country	Automated CRAE/CRVE measurement and CVD risk	Deep-learning vessel measurement system	ICC 0.82-0.95 vs human graders; predicted incident CVD	Yes	Scalable replacement for semi-automated vessel software
Diaz-Pinto et al[48], 2022	United Kingdom Biobank	Myocardial infarction prediction	Multimodal DL (retina + minimal clinical data)	DL-measured arteriolar narrowing: HR 1.64 (95%CI: 1.16-2.32)	Yes	MI risk prediction
Rim et al[49], 2020	Multi-country retinal photographs	Prediction of systemic biomarkers	DL	Multiple biomarkers predicted with variable accuracy	Yes	Opportunistic systemic screening
Rim et al[50], 2021 (RetiCAC)	216152 photographs; Korea, Singapore, United Kingdom	Retinal surrogate for coronary artery calcium	DL	Predicted CVD events; NRI 22%-26% over Pooled Cohort Equation in borderline/intermediate risk	Yes	CVD risk reclassification
Son et al[51], 2020	Korean cohort	High coronary artery calcium score	DL	Discriminated high CAC from fundus images	Partial	CVD risk stratification
Zhang et al[52], 2020	Central China cross-sectional cohort	Hypertension, hyperglycaemia, dyslipidaemia	DL	Moderate discrimination for chronic disease status	No	Community chronic-disease screening
Zhou et al[9], 2023 (RETFound)	1.6 million unlabelled retinal images	Self-supervised foundation model; ocular and systemic tasks	Masked autoencoder (ViT)	Outperformed comparison models for ocular disease and incident systemic disease with fewer labels	Yes	Label-efficient backbone for HR and oculomics tasks

HR: Hypertensive retinopathy; CNN: Convolutional neural network; SVM: Support vector machine; ViT: Vision transformer; AUC: Area under the receiver operating characteristic curve; BP: Blood pressure; SBP: Systolic blood pressure; MAE: Mean absolute error; MACE: Major adverse cardiac events; CVD: Cardiovascular disease; CAC: Coronary artery calcium; ICC: Intraclass correlation coefficient; NRI: Net reclassification improvement; CRAE: Central retinal arteriolar equivalent; CRVE: Central retinal venular equivalent; MI: Myocardial infarction; DL: Deep learning; HRDC: Hypertensive Retinopathy Diagnosis Challenge.

HYPERTENSIVE RETINOPATHY: PATHOPHYSIOLOGY AND CLINICAL FEATURES

The retinal response to elevated blood pressure proceeds through overlapping pathophysiological phases. The initial vasoconstrictive phase produces generalised and focal arteriolar narrowing as a myogenic autoregulatory response. A sustained sclerotic phase follows, characterised by intimal thickening, medial hyperplasia and hyaline degeneration of the arteriolar wall, manifesting clinically as an enhanced arteriolar light reflex (“copper” and later “silver” wiring) and arteriovenous (AV) nicking, in which a thickened arteriole compresses the venule at crossing points. With breakdown of the inner blood-retinal barrier, an exudative phase ensues, producing flame-shaped (nerve-fibre-layer) haemorrhages, cotton-wool spots representing ischaemic infarction of the retinal nerve fibre layer, hard exudates and microaneurysms. In the most severe form, optic disc swelling (papilloedema) reflects malignant hypertension; a macular star of hard exudates deposited in Henle’s layer and serous retinal detachment may accompany the associated hypertensive choroidopathy[3,12,13].

A crucial distinction separates chronic from accelerated/malignant disease. Chronic hypertension produces arteriolosclerotic changes (AV nicking, arteriolar wall opacification) that accrue gradually and overlap with involutional ageing. Accelerated or malignant hypertension, defined by a rapid rise in blood pressure (often > 180/120 mmHg) with grade 3-4 retinal changes and especially bilateral disc oedema, constitutes a hypertensive emergency. Untreated malignant hypertension historically carried a one-year mortality approaching 90%[14], and although most retinal changes regress within roughly six months of blood pressure control, arteriosclerotic changes do not reverse[12]. The clinical continuum of hypertensive retinal injury, from early arteriolar vasoconstriction to malignant hypertensive retinopathy, is summarised in Figure 1.

Figure 1  Progressive retinal manifestations of systemic hypertension.

Grading systems

The Keith-Wagener-Barker (KWB) system grades retinopathy as: Grade 1 (mild generalised arteriolar narrowing); grade 2 (definite focal narrowing and AV nicking); grade 3 (grade 2 plus flame haemorrhages, cotton-wool spots and hard exudates); and grade 4 (grade 3 plus optic disc swelling)[1]. The Scheie system grades hypertensive changes (stages 0-4) and arteriolosclerosis (stages 0-4) on two separate scales, with arteriolosclerosis graded by the evolution of the arteriolar light reflex and AV crossing changes[2]. The simplified Wong-Mitchell classification, derived from population-based data, collapses findings into three prognostically meaningful categories: Mild (one or more of generalised/focal arteriolar narrowing, AV nicking, arteriolar wall opacity), moderate (haemorrhages, microaneurysms, cotton-wool spots or hard exudates) and malignant (moderate signs plus optic disc swelling)[4]. Wong and McIntosh[15] reported that moderate and malignant grades carry strong associations with stroke, cognitive decline, cardiovascular disease and death, whereas the association of the earliest vasoconstrictive and arteriosclerotic signs is weaker and less consistent. The simplified system shows higher intra- and interobserver agreement than KWB, although both systems were associated with aortic stiffness and carotid hypertrophy only in those under 55 years, with no clear discrimination between KWB grades 1 and 2[16].

QUANTITATIVE RETINAL VASCULAR PARAMETERS

The poor reproducibility of subjective grading motivated computer-assisted, photograph-based quantification of vessel calibre. Using the Hubbard protocol developed in the Atherosclerosis Risk in Communities (ARIC) study[5], the six largest arterioles and venules crossing a zone 0.5-1.0 disc diameters from the optic disc margin are measured and summarised, via the Knudtson-Parr-Hubbard formulae[17], as the CRAE and central retinal venular equivalent (CRVE); their ratio yields the AVR. Narrower CRAE reflects the arteriolar vasoconstriction and remodelling of elevated blood pressure, whereas wider CRVE is associated with smoking, inflammation, obesity and dyslipidaemia. AVR, while historically popular, is a non-specific composite that can mask divergent arteriolar and venular changes, and is now generally superseded by separate reporting of CRAE and CRVE[18]. Key photograph-derived vascular biomarkers of hypertension, including retinal vessel calibre and network geometry, are illustrated in Figure 2.

Figure 2  Quantitative retinal vascular biomarkers of systemic hypertension.

Epidemiological evidence

The epidemiological evidence base is substantial and is summarised in Table 1. In the ARIC study, higher mean arterial blood pressure was the dominant systemic determinant of narrower CRAE[19]. An individual-participant meta-analysis confirmed that narrower arterioles and wider venules independently predict incident hypertension[20], and in the multi-ethnic study of atherosclerosis a smaller CRAE was associated with 34% higher odds of incident hypertension per one standard-deviation decrease (OR = 1.34, 95%CI: 1.18-1.52) after adjustment for age, sex and ethnicity[21]. The Blue Mountains Eye Study demonstrated that retinal arteriolar narrowing predicted five-year incident severe hypertension[22], and in ARIC a smaller AVR was associated with incident coronary heart disease in women but not in men[23]. In the ARIC cohort, retinal vessel calibres also predicted long-term ischaemic stroke, heart failure and all-cause mortality[24].

Comparison across these studies exposes a consistent and clinically important pattern: Predictive value differs systematically by vessel type, endpoint and sex. In the McGeechan individual-participant meta-analysis of incident stroke (20798 participants, 945 events), wider venular calibre predicted stroke (pooled HR = 1.15, 95%CI: 1.05-1.25 per 20-µm increase) while arteriolar calibre did not (HR = 1.00, 95%CI: 0.92-1.08)[25], whereas the companion meta-analysis of coronary heart disease found that both venular widening and arteriolar narrowing predicted events in women (HR = approximately 1.16-1.17 per 20 µm) but not in men[26]. Taken together with the ARIC sex-specific coronary findings[23], this indicates that arteriolar and venular calibre are not interchangeable markers of a single process: Arteriolar narrowing indexes blood-pressure-related myogenic and remodelling change, whereas venular widening appears to index inflammation, endothelial dysfunction and hypoxia. The clinical corollary is that composite indices such as AVR should be abandoned in favour of separate reporting, and that effect sizes derived in one sex or one endpoint cannot be extrapolated to another - a point insufficiently acknowledged in much of the AI literature that uses these measures as ground truth.

Beyond calibre, geometric parameters capture the efficiency and complexity of the vascular network. Fractal dimension quantifies branching complexity and density; a lower (sparser) fractal dimension has been associated with higher blood pressure and with stroke incidence and mortality[27,28], with stroke-related mortality higher in the lowest fractal-dimension quartile (7.7%) than in the highest (1.3%)[29]. Systematic review has confirmed the potential of fractal dimension as a neurovascular biomarker while emphasising that between-study heterogeneity in segmentation and image quality currently precludes a pooled estimate[30]. Increased arteriolar tortuosity and altered branching angles, which deviate from the optimal geometry that minimises energy expenditure, have likewise been linked to blood pressure and cardiovascular risk[31]. These parameters are highly sensitive to image quality and segmentation method, a limitation directly relevant to the standardisation problem discussed below.

RETINAL PHOTOGRAPHY AND IMAGING MODALITIES

The four platforms in current use differ substantially in the biomarkers they generate, in cost and accessibility, and in clinical readiness; these differences are compared directly in Table 2 and illustrated in Figure 3.

Figure 3 Multimodal retinal imaging biomarkers of hypertension-related microvascular injury. OCT: Optical coherence tomography.

Non-mydriatic fundus photography

Non-mydriatic fundus photography is the workhorse of population-level retinal assessment. Requiring no pupil dilation and increasingly delivered by low-cost, portable and even smartphone-based cameras, it underpins existing diabetic retinopathy screening programmes and provides the substrate for nearly all quantitative vessel analysis and AI work. Its scalability is the principal reason the retina is attractive as a screening organ for hypertension[10].

OCT

OCT provides micron-resolution cross-sectional imaging of the neurosensory retina. Hypertension is associated with thinning of the peripapillary retinal nerve fibre layer and the macular ganglion cell-inner plexiform layer (GC-IPL), reflecting the neuronal consequences of chronic microvascular compromise; these changes can be present in hypertensive individuals even before overt retinopathy. One study reported mean GC-IPL thicknesses of 83.5 µm in controls, 82.1 µm in hypertensives without retinopathy and 75.9 µm in those with resolved hypertensive retinopathy (P < 0.001), illustrating a gradient of neuroretinal loss[32]. Some studies describe a U-shaped relationship between blood pressure and OCT metrics, indicating that both very low and very high pressures may be detrimental to retinal perfusion.

OCTA

OCTA images the retinal capillaries without dye, resolving the superficial and deep capillary plexuses and the choriocapillaris, and quantifying vessel density, the foveal avascular zone (FAZ) and flow voids. A meta-analysis of 11 studies found that hypertensive eyes had significantly lower superficial vessel density [standardised mean difference (SMD) = -0.50, 95%CI: -0.70 to -0.30], lower deep vessel density (SMD = -0.38, 95%CI: -0.64 to -0.13) and a larger superficial FAZ (SMD = 0.32, 95%CI: 0.04-0.61) than controls[33], consistent with the microvascular rarefaction that characterises hypertensive end-organ damage. OCTA changes are demonstrable in hypertensive patients without clinically visible retinopathy[34,35], supporting a role in detecting subclinical disease, and choriocapillaris flow voids increase with poorly controlled blood pressure[36,37]. It should be noted that the effect sizes are moderate and the confidence intervals wide, that segmentation and projection-artefact removal differ between devices, and that normative databases are not interchangeable - constraints that currently confine OCTA to research and specialist practice rather than screening.

Adaptive optics scanning laser ophthalmoscopy

Adaptive optics scanning laser ophthalmoscopy (AOSLO) corrects ocular aberrations to achieve cellular-resolution imaging of the arteriolar wall, permitting direct in vivo measurement of the wall-to-lumen ratio (WLR), a validated marker of microvascular remodelling that correlates with the media-to-lumen ratio of subcutaneous small arteries obtained by biopsy[38,39]. In a patient with malignant hypertensive retinopathy, AOSLO documented an arterial wall thickness of 18.7 µm and a WLR of 0.44, both markedly elevated relative to normal[40]. AOSLO and related systems remain research tools constrained by small fields of view, cost, acquisition time and operator expertise, and no population-scale AOSLO dataset exists.

AI AND DEEP LEARNING IN RETINAL HYPERTENSION BIOMARKER DISCOVERY

Key AI studies are compared in Table 3 with respect to dataset size and provenance, task, architecture, reported performance and the presence or absence of external validation.

Automated hypertensive retinopathy grading and the benchmark problem

Deep learning has been applied to classify hypertensive retinopathy and to segment its constituent lesions from fundus photographs. CNN approaches, frequently built on vessel segmentation followed by feature classification, have reported high accuracies. The five-stage DenseNet system of Abbas et al[41], using semantic and instance segmentation across 1400 graded images, reported sensitivity 90.5%, specificity 91.5%, accuracy 92.6% and area under the curve (AUC) 0.915; a depth-wise separable CNN with a support-vector-machine classifier reported sensitivity 94%, specificity 96%, accuracy 95% and AUC 0.96[42]; an improved CNN with a modified support-vector-machine classifier reported accuracy up to 98.99% for five-stage severity classification on augmented public datasets[43]; and a hybrid ResNet-50/Vision Transformer architecture with locality self-attention and decoupled representation-classifier training reported accuracy 0.9688, sensitivity 0.9435 and specificity 0.9766 for four-class severity grading[44].

These figures require careful interpretation, and the most informative way to interpret them is by direct comparison with performance on a common test set. The Hypertensive Retinopathy Diagnosis Challenge (HRDC), organised in conjunction with Computer Graphics International 2023, released the first publicly available expert-annotated benchmark for this task, comprising 1000 fundus images each for hypertension classification and hypertensive retinopathy classification. On that benchmark, the best submitted algorithm achieved a Kappa of 0.3819 (F1: 0.6337, specificity: 0.8472) for hypertension classification and a Kappa of 0.4154 (F1: 0.6122, specificity: 0.8444) for hypertensive retinopathy classification[8]. The contrast between accuracies exceeding 90% on curated private datasets and only fair chance-corrected agreement on a shared benchmark is, in our view, the single most important quantitative observation in this literature.

Three explanations are plausible and not mutually exclusive. First, private datasets are typically small, single-source and class-balanced by design or by augmentation, so reported accuracy is inflated relative to screening populations in which hypertensive retinopathy is uncommon; accuracy is in any case a poor metric under class imbalance, and Kappa or F1 should be reported alongside it. Second, most such studies report internal cross-validation rather than external validation on independent devices and populations, so the estimates reflect within-dataset optimisation. Third, and most fundamentally, the reference standard is itself unreliable: Models are trained to reproduce a human grading system whose poor interobserver reproducibility was the original motivation for quantitative measurement[3,5]. An algorithm cannot exceed the reliability of its labels. Until benchmark-based, externally validated reporting becomes the norm, headline accuracies from hypertensive retinopathy classification studies should not be treated as evidence of clinical readiness.

Direct prediction of blood pressure and cardiovascular risk

The landmark study by Poplin et al[45] demonstrated that deep-learning models trained on fundus images from 284335 patients, and validated on independent datasets, could infer systemic variables not previously thought quantifiable from the retina, predicting systolic blood pressure with a mean absolute error of 11.23 mmHg and sex with an AUC of 0.97. The same models predicted major adverse cardiac events with an AUC of 0.70 (95%CI: 0.648-0.740), comparable to but not significantly better than the European SCORE risk calculator (AUC: 0.72); adding the algorithm to SCORE did not improve discrimination, and wide confidence intervals reflected the small number of events in the validation set. Saliency mapping indicated that the models attended to anatomically plausible features including vessels and the optic disc[45]. A mean absolute error of 11.23 mmHg is of the same order as the difference between hypertension stages, which means that the model is informative at the level of populations but not at the level of an individual treatment decision - a distinction that is frequently blurred in secondary reporting of this work.

Automated vessel quantification tools

Semi-automated platforms - Integrative Vessel Analysis (IVAN), Singapore I Vessel Assessment (SIVA) and the Vascular Assessment and Measurement Platform for Images of the Retina (VAMPIRE) - operationalised CRAE/CRVE/AVR and geometric parameters for research cohorts, but they show only poor-to-moderate cross-platform agreement, with mean CRAE differences of 16-37 µm between systems, which complicates pooling of results and comparison of published thresholds[46]. A fully automated deep-learning successor, SIVA-DLS, was trained and validated on more than 70000 images from diverse multi-ethnic, multi-country datasets; it measured retinal vessel calibre with agreement against human graders of intraclass correlation 0.82-0.95, performed comparably to or better than expert graders in relating calibre to blood pressure and other risk factors, and predicted incident cardiovascular disease[47]. Independent analysis confirmed that deep-learning-measured arteriolar narrowing was associated with incident myocardial infarction (multivariable HR = 1.64, 95%CI: 1.16-2.32)[48]. Notably, the studies with the strongest external validation - SIVA-DLS and the UK Biobank myocardial infarction analysis - are those that predict a well-defined continuous or event outcome rather than a subjective grading label, reinforcing the conclusion that the reference-standard problem, rather than model architecture, is the principal constraint.

Prediction of events, mortality and surrogate markers

Rim et al[49] showed that deep learning could predict multiple systemic biomarkers from retinal photographs, and subsequently developed RetiCAC, a retinal surrogate for coronary artery calcium derived from 216152 photographs across South Korea, Singapore and the United Kingdom; stratified RetiCAC predicted cardiovascular events and added incremental value to the Pooled Cohort Equation in borderline- and intermediate-risk individuals, with a net reclassification improvement of 22%-26%[50]. Other groups have predicted high coronary artery calcium scores and chronic disease status directly from routine screening images[51,52], and a multimodal approach combining fundus photographs with minimal clinical data improved event prediction further[48]. The incremental value demonstrated to date is concentrated in the intermediate-risk stratum, which is also the stratum in which clinical decisions are genuinely uncertain - a favourable finding for potential clinical utility, but one that has not yet been tested prospectively.

Foundation models and label efficiency

The most significant recent architectural development is the emergence of retinal foundation models. RETFound was trained by self-supervised learning on 1.6 million unlabelled retinal images and then adapted to downstream tasks with limited labels, outperforming comparison models in the diagnosis and prognosis of ocular disease and in incident prediction of systemic disorders including heart failure and myocardial infarction[9]. This paradigm is directly relevant to hypertensive retinopathy, where progress has been constrained less by algorithmic sophistication than by the scarcity of large, expert-annotated datasets[8,44]; label-efficient adaptation of a pretrained retinal encoder offers a plausible route out of that constraint. Broader syntheses of oculomics have reached similar conclusions regarding the shift from task-specific to general-purpose retinal representations[11].

Explainability and distributed training

Because clinical adoption requires trust, gradient-weighted class activation mapping and related saliency methods are increasingly used to confirm that models attend to vessels, optic disc and macula rather than to artefacts[53]. Saliency maps are, however, an explanation of where a model looked rather than of why it decided, and they should be regarded as a necessary sanity check rather than sufficient evidence of clinical validity. Federated and privacy-preserving learning frameworks are being explored to train across institutions without sharing identifiable images and to improve generalisability across populations[6].

CLINICAL IMPLICATIONS AND SCREENING POTENTIAL

Several lines of evidence indicate that retinal findings can precede the clinical diagnosis of hypertension. Population studies demonstrate that altered retinal microvascular diameters predict progression from normotension to hypertension[54], positioning the retina as an early sentinel. Retinal vascular abnormalities are also detectable across hypertension sub-phenotypes defined by ambulatory monitoring: In Triantafyllou et al’s study[55], AVR was significantly lower not only in sustained hypertension (0.736) but also in masked (0.716) and white-coat hypertension (0.739) relative to true normotension (0.820), suggesting that retinal photography could help characterise these higher-risk states.

These observations support two complementary clinical roles. First, retinal biomarkers may refine cardiovascular risk stratification beyond office blood pressure and conventional scores; the RetiCAC and SIVA-DLS data provide proof of concept, although the incremental gains demonstrated so far are modest and the Poplin model did not significantly outperform SCORE[45,47,50]. Second, the scalability of non-mydriatic photography combined with AI interpretation enables teleophthalmology and opportunistic screening: A retinal image acquired for diabetic retinopathy screening could simultaneously flag undiagnosed or inadequately controlled hypertension and elevated cardiovascular risk[10].

BARRIERS TO CLINICAL TRANSLATION

Standardisation of retinal measurement

No consensus protocol governs image acquisition, measurement zone, vessel selection or summarisation. The demonstrated disagreement between IVAN, SIVA and VAMPIRE, with mean CRAE differences of 16-37 µm[46], is of the same order as the between-group differences those tools are used to detect, which makes published cut-points non-transferable between platforms. Camera field of view, magnification, refractive error, axial length, pupil size, media opacity and even cardiac cycle phase all influence measured calibre. Progress requires a standardised acquisition and analysis protocol, publication of device- and ethnicity-specific normative ranges, phantom or reference-image calibration across platforms, and routine reporting of repeatability and reproducibility coefficients. Equivalent standardisation is required for OCTA, where vessel density values are not comparable across devices or segmentation algorithms.

External validation and algorithmic generalisability

The great majority of hypertensive retinopathy AI studies report internal validation only (Table 3). Generalisability must be demonstrated prospectively across ethnicities, camera manufacturers, image quality strata, degrees of media opacity and comorbid retinal disease, since diabetic retinopathy, age-related macular degeneration and glaucoma coexist with hypertension and may confound both human and algorithmic assessment. Subgroup performance should be reported rather than pooled, because an algorithm with acceptable aggregate performance may nonetheless be systematically inaccurate in the subgroups at highest absolute risk. Data provenance, class distribution, exclusion of ungradable images and calibration - not only discrimination - should be reported as standard.

Regulatory considerations

Software that infers systemic disease from ocular images is a medical device in most jurisdictions and, when used to inform diagnosis or triage, falls into a higher risk class. Autonomous AI screening in diabetic retinopathy provides the closest regulatory precedent and establishes the expected evidentiary bar: A prospective, pre-registered trial against a rigorous reference standard, conducted in the intended-use population and setting, and by the intended operators[10]. For hypertension the reference-standard problem is harder, because the comparator is not a reading-centre grade but a physiological quantity that varies diurnally and by measurement method; a prospective validation would need standardised office measurement and, ideally, ambulatory monitoring as the reference. In the European Union, AI systems that are safety components of medical devices are classified as high risk, adding requirements for risk management, data and data-governance quality, technical documentation, transparency, human oversight and post-market monitoring on top of the Medical Device Regulation. Continuously learning models raise the further question of change control, for which predetermined change-control plans have emerged as the favoured regulatory instrument. Investigators should anticipate these requirements at the design stage rather than retrofitting them, and should be explicit about intended use, since a claim of “screening for undiagnosed hypertension” carries a materially different regulatory and liability burden from “risk stratification adjunct”.

Integration into screening pathways and referral workflow

The question of who acquires the image, who is accountable for the output and what happens next is at least as consequential as model performance, and is largely unaddressed in the current literature. Three deployment models are plausible. The first is opportunistic capture within existing diabetic retinopathy screening programmes, which is the least costly because the image, the camera, the trained photographer and the governance framework already exist; the marginal cost is an additional algorithmic read. The second is community-based capture in optometric practice or community pharmacy using low-cost non-mydriatic or smartphone-based cameras, which extends reach to people not otherwise engaged with health services but requires new quality-assurance and indemnity arrangements. The third is capture in primary care by a trained health-care assistant during routine chronic-disease review, which places the result directly in front of the clinician who will act on it but adds equipment and training costs to an already pressured setting.

In all three models, an abnormal algorithmic result must not be reported as a diagnosis of hypertension. It should instead trigger a defined pathway: Confirmation by standardised office blood pressure measurement and, where indicated, home or 24-hour ambulatory monitoring; assessment of global cardiovascular risk and of other end-organ damage in primary care; referral to ophthalmology where sight-threatening features are present; and same-day assessment where severe blood-pressure elevation is accompanied by bilateral disc oedema, retinal haemorrhages or visual loss, which constitutes a hypertensive emergency[12,14]. Referral to cardiology, nephrology or an accelerated-hypertension service is appropriate where accelerated hypertension, secondary hypertension or additional end-organ damage is suspected. Operationally, such a pathway requires a named responsible clinician for each algorithmic output, a defined fail-safe route for ungradable images, a documented protocol for discordance between the algorithm and clinical findings, integration of results into the electronic health record rather than a standalone system, and continuous audit of positive predictive value, uptake of confirmatory testing and downstream diagnostic yield. Health-economic evaluation - including the cost of false positives, the workload of confirmatory testing and the effect on ophthalmology referral volumes - should accompany any pilot, since a screening test that generates unabsorbable downstream demand will not be adopted regardless of its accuracy.

Critically, no randomised evidence yet shows that acting on a retina-derived biomarker improves patient outcomes[7].

FUTURE DIRECTIONS

The most promising near-term direction is multimodal integration, combining fundus photography, OCT, OCTA and structured systemic data within a single model, an approach that has already improved cardiovascular risk discrimination relative to single-modality models[6,7]. Such integration may allow the retina to contribute to hypertension sub-phenotyping (white-coat, masked and resistant hypertension), where current diagnosis depends on cumbersome ambulatory monitoring[55]. Foundation models pretrained on large unlabelled image collections and adapted with parameter-efficient fine-tuning may overcome the data scarcity that has constrained hypertensive retinopathy AI to date[9,11].

Five priorities should structure the research agenda. First, an international consensus protocol for retinal vascular measurement, with cross-platform calibration and published normative data stratified by ethnicity, age and axial length. Second, expansion of open, expert-annotated, multi-ethnic benchmarks along the lines of the HRDC dataset, with mandatory benchmark reporting alongside internal results[8]. Third, prospective multicentre external validation of the most promising algorithms against standardised office and ambulatory blood pressure measurement, with pre-registered analysis plans, subgroup reporting and calibration as well as discrimination. Fourth, development and evaluation of explainable models whose outputs clinicians can interrogate, moving beyond post hoc saliency towards inherently interpretable or concept-based architectures[53]. Fifth, and decisively, randomised or well-designed prospective implementation studies testing whether a retina-derived biomarker changes management - detection of previously undiagnosed hypertension, treatment intensification, achieved blood pressure - and ultimately whether it reduces cardiovascular events, accompanied by formal health-economic evaluation.

CONCLUSION

The retina has evolved from a structure examined ophthalmoscopically for prognostic clues in malignant hypertension into a quantitatively measurable, AI-interpretable model of the systemic microcirculation. Hypertensive retinopathy grading retains value for recognising hypertensive emergencies and stratifying systemic risk, while quantitative calibre and geometric parameters, capillary-level OCTA metrics and deep-learning algorithms extend assessment to the subclinical and predictive domains. The evidence nonetheless supports a measured conclusion. That the retina contains blood-pressure and cardiovascular information is established; that this information is currently actionable in an individual patient is not. Reported diagnostic accuracies above 90% on private datasets are not reproduced on shared benchmarks, incremental gains over established risk scores are modest and concentrated in the intermediate-risk stratum, measurement platforms disagree by clinically meaningful margins, and no randomised evidence links a retina-derived biomarker to improved outcomes. The field’s immediate priorities are therefore standardisation of retinal biomarker assessment, prospective multicentre and multi-ethnic validation, development of explainable models, definition of the screening-to-referral pathway with clear clinical accountability, and evaluation of the effect of these tools on clinical decision-making and patient outcomes. If those conditions are met, the eye may become a practical, scalable instrument for hypertension detection and cardiovascular risk reduction; until they are, retinal biomarkers should be regarded as a rapidly maturing research tool rather than a deployable clinical test.