ArticlePDF AvailableLiterature Review

Abstract and Figures

Among the most noteworthy developments in ophthalmology over the last decade has been the emergence of quantifiable high-resolution imaging modalities, which are typically non-invasive, rapid and widely available. Such imaging is of unquestionable utility in the assessment of ocular disease however evidence is also mounting for its role in identifying ocular biomarkers of systemic disease, which we term oculomics. In this review, we highlight our current understanding of how retinal morphology evolves in two leading causes of global morbidity and mortality, cardiovascular disease and dementia. Population-based analyses have demonstrated the predictive value of retinal microvascular indices, as measured through fundus photography, in screening for heart attack and stroke. Similarly, the association between the structure of the neurosensory retina and prevalent neurodegenerative disease, in particular Alzheimer’s disease, is now well-established. Given the growing size and complexity of emerging multimodal datasets, modern artificial intelligence techniques, such as deep learning, may provide the optimal opportunity to further characterize these associations, enhance our understanding of eye-body relationships and secure novel scalable approaches to the risk stratification of chronic complex disorders of ageing.
Content may be subject to copyright.
Special Issue
Insights into Systemic Disease through Retinal
Imaging-Based Oculomics
Siegfried K. Wagner1, Dun Jack Fu1, Livia Faes1,2, Xiaoxuan Liu3,4, Josef Huemer1,
Hagar Khalid1, Daniel Ferraz1, Edward Korot1, Christopher Kelly5,
Konstantinos Balaskas1, Alastair K. Denniston1,3,4, and Pearse A. Keane1
1NIHR Biomedical Research Center at Moorelds Eye Hospital NHS Foundation Trust and UCL Institute of Ophthalmology, London, UK
2Department of Ophthalmology, Cantonal Hospital Lucerne, Lucerne, Switzerland
3Department of Ophthalmology, University Hospitals Birmingham NHS Foundation Trust, Birmingham, UK
4Academic Unit of Ophthalmology, Institute of Inammation & Ageing, University of Birmingham, Birmingham, UK
5Google Health, London, UK
Correspondence: Pearse A. Keane,
NIHR Biomedical Research Centre for
Ophthalmology, Moorelds Eye
Hospital NHS Foundation Trust and
UCL Institute of Ophthalmology,
London, UK. e-mail:
Received: October 8, 2019
Accepted: October 9, 2019
Published: February 12, 2020
Keywords: deep learning; articial
intelligence; optical coherence
Citation: Wagner SK, Fu DJ, Faes L,
Liu X, Huemer J, Khalid H, Ferraz D,
Korot E, Kelly C, Balaskas K,
Denniston AK, Keane PA. Insights
into systemic disease through retinal
imaging-based oculomics. Trans Vis
Sci Tech. 2020;9(2):6,
Among the most noteworthy developments in ophthalmology over the last decade
has been the emergence of quantiable high-resolution imaging modalities, which are
typically non-invasive, rapid and widely available. Such imaging is of unquestionable
utility in the assessment of ocular disease however evidence is also mounting for its
role in identifying ocular biomarkers of systemic disease, which we term oculomics.In
this review, we highlight our current understanding of how retinal morphology evolves
in two leading causes of global morbidity and mortality, cardiovascular disease and
dementia. Population-based analyses have demonstrated the predictive value of retinal
microvascular indices, as measured through fundus photography, in screening for heart
attack and stroke. Similarly, the association between the structure of the neurosensory
retina and prevalent neurodegenerative disease, in particular Alzheimer’s disease, is
now well-established. Given the growing size and complexity of emerging multimodal
datasets, modern articial intelligence techniques, such as deep learning, may provide
the optimal opportunity to further characterize these associations, enhance our under-
standing of eye-body relationships and secure novel scalable approaches to the risk
stratication of chronic complex disorders of ageing.
The convergence of modern multimodal imaging
techniques and large-scale data sets has fostered an
extraordinary opportunity to exhaustively charac-
terize the macroscopic, microscopic, and molecular
ophthalmic features associated with health and disease
(i.e., the oculome). One of the potential avenues of
this oculomics revolution is the leveraging of the retina
to gain insights beyond the eye. As the only human
tissue allowing direct noninvasive in vivo visualization
of the microvascular circulation and central nervous
system, the neurosensory retina aords a unique setting
for the characterization of systemic disease. Microvas-
cular changes precede clinical manifestation, and so
their detection should have predictive value.1Indeed,
ophthalmoscopic changes in retinal microvasculature
structure have been identied as independent predic-
tors for hypertension, diabetes, coronary disease, renal
Copyright 2020 The Authors | ISSN: 2164-2591 1
This work is licensed under a Creative Commons Attribution 4.0 International License.
Downloaded from on 02/24/2020
Oculomics: An Eye on Systemic Disease TVST | Special Issue | Vol. 9 | No. 2 | Article 6 | 2
disease, and stroke.26Alterations in the thickness of
the retinal nerve ber layer and macular volume, most
easily revealed through optical coherence tomography
(OCT), may highlight those individuals most at risk
of developing cognitive decline and neurodegenera-
tive disease.710 Furthermore, certain disorders may
exhibit distinct retinal manifestations signifying their
presence—the sea fan neovascularization of sickle cell
anemia, the macular crystals of cystinosis, or the astro-
cytic hamartomas of tuberous sclerosis.11
A major facilitator for this has been the advance-
ment in retinal imaging modalities over the past two
decades. Primitive methods of direct ophthalmoscopy
have evolved to encompass a diverse armamentar-
ium of high-resolution imaging techniques, which are
predominantly easy to acquire, risk free, and often
demanding only nominal expertise and time for acqui-
sition. In particular, both modern retinal photogra-
phy and OCT are now of unprecedented resolution
and increasingly carried out on a routine basis in
both hospital eye settings and within the commu-
nity.12 At Moorelds Eye Hospital NHS Founda-
tion Trust, the largest ophthalmic unit in North
America and Europe, we have seen a 14-fold increase
in the capture of OCT from 23,500 scans in 2008 to
>330,000 in 2016.13 Similarly, the availability of such
cross-sectional imaging has exploded in primary care
settings—the largest optical franchise in the United
Kingdom, Specsavers, announced in 2017 that each of
its >700 branches would have an OCT device by the
end of 2019.14 The primary objective of this transfor-
mation has been to enhance the diagnosis of sight-
threatening retinal disease, but an important paral-
lel opportunity is emerging—the ability to use ocular
biomarkers to detect systemic disease, predict its future
onset, and provide noninvasive surrogates of its sever-
ity and treatment response.
Meaningful quantitative relationships between
retinal structure and systemic disease have now been
established using population-based analyses in cardio-
vascular disease (CVD) and dementia. In the former,
changes in retinal microvasculature, from vascular
caliber to tortuosity indices, have been associated
with CVD risk factors and may have predictive value
for relevant events, such as myocardial infarction
and stroke.4,5,1517 Traditionally, this has relied upon
onerous manual segmentation of digitized images.
The development of semiautomated retinal imaging
analysis software has alleviated this burden but still
demands signicant time and researcher input for
large-scale data sets highlighting the requirement for
fully automated means, which may be addressed by
modern methods of articial intelligence (AI). One
form of AI, deep learning, may be the answer. In 2018,
researchers at Google Brain not only constructed a
model capable of predicting CVD risk factors with
reasonable accuracy but, more surprisingly, was also
able to predict age and sex with impressive con-
dence.18 Reassurance of the rationale behind the
model’s output came from interpretability techniques
highlighting retinal vasculature, the optic nerve, and
macular morphology in decision making.
Although our understanding of eye-body relation-
ships has evolved from decades of traditional statistical
modeling in large population-based studies incorporat-
ing ophthalmic assessments, the application of AI to
this eld is still in its early stages. In this article, we
focus on where AI-based studies may build on these
traditional analyses to reveal novel insights leveraging
retinal biomarkers of systemic disease. Particular focus
is paid to common chronic disorders of aging, such as
cardiovascular and neurodegenerative disease.
Cardiovascular Disease
The rst reported association between systemic
vascular disease and the eye likely comes from the
British nephrologist, Richard Bright, who in 1836
described a series of patients with albuminuria and
vision loss.19 The ophthalmic features of what would
become known as Bright disease, an umbrella term for
all forms of glomerulonephritis, would not be ratio-
nalized until after the invention of the ophthalmo-
scope, when Marcus Gunn noted features of severe
hypertensive retinopathy in a cohort of patients with
chronic renal impairment in 1892.20 The concept of
retinal-based quantication of cardiovascular disease
risk would subsequently come in 1939, when Keith et
al.21 would describe their prominent grading system for
hypertensive retinopathy, which “permitted an intel-
ligent appraisal of the individual patient and has
increased considerably the accuracy of prognosis.”
As CVD is the leading cause of global mortal-
ity, accounting for more than 30% of deaths world-
wide, there has been signicant motivation to develop
eective tools to identify those most at risk.22 The
2019 guidelines from the American College of Cardi-
ology and American Heart Association recommend
use of the ASCVD Risk Estimator Plus, which calcu-
lates a 10-year CVD risk score based on certain
risk factors (age, sex, ethnicity), bedside tests (e.g.,
blood pressure), and blood parameters (e.g., total
cholesterol).23 However, even such risk stratica-
tion algorithms can have limited calibration and
discriminative ability when externally validated.24,25
Moreover, generating these scores depends on signif-
Downloaded from on 02/24/2020
Oculomics: An Eye on Systemic Disease TVST | Special Issue | Vol. 9 | No. 2 | Article 6 | 3
icant health care professional input and laboratory
The use of a single noninvasive “eye check” to
assess CVD risk is an attractive alternative, in part
because of the importance that most of the population
places on matters related to their vision and eye health.
When surveyed, the general public ranks sight as our
most important sense.26 This translates into signicant
dierences between the extent to which members of
the public attend eye checks compared to screening for
CVD. For example, the free “Over-40” check estab-
lished by the UK National Health Service for CVD
risk stratication by primary care physicians in 2009
was attended by only 12.8% of the population from
2009 to 2013.27 In contrast, more than half the popula-
tion attended their community optometric practice for
regular eye checks in 2016.28
A role for retinal photography can therefore be
envisaged in three settings. First, it could be used as
an additional investigation, enabling risk renement.
There is evidence that the addition of retinal photog-
raphy can positively aect reclassication indices for
current risk stratication scoring systems.29,30 Its use as
a mandated additional investigation would have signif-
icant resource implications, but an alternative would
be the inclusion of data from imaging in those cases
where it is available, in which case the requirement is
primarily one of data integration. Second, there may
be a role for retinal photography as an alternative to
current CVD risk approaches in resource-poor settings.
The emergence of widespread retinal photography in
the developing world through smartphone technology
and improved access for diabetic retinopathy screen-
ing may enable democratization of CVD risk stratica-
tion to this neglected population. Third, it may be used
as an ad hoc screening test. Deployment in commu-
nity optometry settings could allow identication of
people at risk of CVD who would not otherwise have
attended their primary care physician and are therefore
prompted to have further investigation.
Direct noninvasive visualization of the microvascu-
lar circulation is a unique attribute of the neurosen-
sory retina. The shared anatomical and physiologic
characteristics between retinal vessels and those of
the kidney and heart support the potential utility of
retinal assessment as a conduit to systemic vascular
disease risk stratication. The assessment of the retinal
circulation has evolved from initially being through
direct ophthalmoscopy, which is fraught with substan-
tial intra- and interobserver variability, to the intro-
duction of digital or digitized retinal photography
with greater repeatability and precision.31,32 A large
number of population-based studies have now demon-
strated signicant relationships between retinal vascu-
lar features and both cardiovascular disease outcomes,
in particular myocardial infarction and stroke, and
risk factors (age, blood pressure, smoking, presence of
diabetes mellitus).15,16,30,33
The most convincing association has been made
between risk of incident stroke and retinal vascular
morphology. The Atherosclerosis Risk in Communities
study was the rst population-based study to evaluate
the relationship between retinal vasculature and cardio-
vascular disease on a large scale and incorporate retinal
photography.15 Not only were images graded for quali-
tative features suggestive of hypertensive retinopathy,
such as cotton wool spots and arteriovenous nicking,
but, following digitization of the images, semiquantita-
tive measurement of the retinal vessel caliber also was
completed. Controlling for known risk factors, includ-
ing age, sex, diabetes, and blood features, most features
of retinopathy were indeed associated with a higher
relative risk of incident stroke, but it was also noted
that such risk increased in those with smaller arteriole-
to-venule ratios in a proportional scale. This was
reinforced by similar results among diabetic patients
in the Wisconsin Epidemiologic Study of Diabetic
Retinopathy.34 However, interestingly, a meta-analysis
incorporating six studies evaluating retinal vascu-
lar caliber and incident stroke over 5 to 12 years
concluded that wider retinal venular caliber and not
retinal arteriolar caliber predicted stroke with a pooled
hazard ratio of 1.15 (condence interval, 1.05–1.25 per
20-micron increase in caliber).5To a certain extent,
this nding persists when considering coronary heart
disease (CHD) but with an important distinction.
Retinal vascular morphology appears to be helpful only
for risk stratication in women. A similar meta-analysis
by McGeechan and colleagues4evaluating the same
six population-based cohorts revealed wider retinal
venules, and narrower arterioles predicted 5- to 14-year
risk of CHD events (namely, myocardial infarction,
coronary bypass grafting, and coronary angioplasty)
in women but not in men. The authors hypothesize that
microvascular dysfunction plays a more prominent role
in CHD risk in women.
Rather than focusing on CVD events, a number of
studies have established links between retinal vascular
characteristics and risk factors. In recent work using
over 5000 participants from the European Prospective
Investigation into Cancer-Norfolk Eye Study, Owen
et al.35 examined retinal vessel caliber and tortuosity
using the fully automated software QUARTZ. Increas-
ing arteriolar tortuosity was associated with rising age
and systolic blood pressure, whereas venular tortuos-
ity was more related to body mass index and preva-
lent type 2 diabetes mellitus. In terms of vascular
caliber, retinal venular caliber was higher in older
Downloaded from on 02/24/2020
Oculomics: An Eye on Systemic Disease TVST | Special Issue | Vol. 9 | No. 2 | Article 6 | 4
patients, smokers, and those with raised triglyceride
levels. Arterioles were narrower in older patients and
those with higher systolic blood pressure and total
The studies mentioned thus far have centered on
traditional statistical modeling techniques, such as
regression and survival analysis, to draw insights on
how retinal structure changes in CVD. However, these
hypotheses-based methods rely on clinician direction
within a narrow prespecied group of parameters. In
contrast, Poplin et al.18 argue that the plethora of infor-
mation within retinal photographs lends itself ideally to
deep learning. The team from Google Research trained
a convolutional neural network on fundus photos of
>280,00 patients from the UK Biobank and EyePACS
to predict CVD risk factors. Not only did the model
indeed predict smoking status and major cardiac events
with reasonable accuracy (area under the receiver
operating characteristic curve [AUC] of 0.71 and 0.70
respectively), it surprised many readers in its ability to
identify sex and age with high condence (AUC of 0.97
and mean absolute error of 3.26 years, respectively).
Importantly, the model performed well when externally
validated on a separate data set of Asian patients by an
independent research group.36 A further novel feature
of their work reects on the concern of the limited
interpretability inherent in deep learning systems. They
employed the deterministic technique, soft attention, to
illustrate which pixels of the image were most inu-
ential in the model’s decision. Although ophthalmic
researchers may not be surprised to see that systolic
blood pressure prediction was predominantly predicted
using the retinal vessels, it may fascinate others to learn
that the foveal appearance was instrumental in predict-
ing biological sex. These saliency maps therefore have
the potential to not only reassure us of the biological
plausibility of model decision making but also stimu-
late research into potential novel biomarkers, akin to
phenotype-rst genome-wide association studies.
Neurodegenerative Disease
There are over 9.9 million new cases of demen-
tia worldwide each year. The World Alzheimer Report
2015 estimates that the number of aected people
will double every 20 years, equating to >130 million
cases in 2050.37 In the United Kingdom, demen-
tia overtook CVD as the leading cause of death in
2016.38 Yet, despite these alarming gures, it has been
estimated that 50% to 80% of cases of the most
common form of dementia, Alzheimer disease (AD),
remain unrecognized in high-income countries, due to
the challenges in detection and diagnosis.37 Typ i c ally,
individuals suspected of having dementia are assessed
by their primary care physician using the usual combi-
nation of medical history, physical examination, and,
in some cases, investigations in conjunction with a
questionnaire regarding cognitive function. However,
not only do these have variable diagnostic accuracy,
but they also depend on an index of suspicion as
well as attendance by the individual. In addition, the
gold-standard diagnosis for some forms of demen-
tia, such as AD, relies on brain biopsy, which is not
appropriate for routine use. Evidence is growing for
less invasive investigations such as cerebrospinal uid
analysis of amyloid protein and magnetic resonance
imaging (MRI), but again, there are resource impli-
cations to using these tests at scale.3941 One imaging
protocol increasingly used for AD diagnosis, MRI–
positron emission tomography amyloid, requires a
prescan injection of a radioactive tracer and can take
over an hour to complete. This is in stark contrast to
the seconds needed for acquiring retinal imaging.
In 1986, following on from observations that people
with AD had visual decits that could not be explained
purely by cortical disease, David Hinton and colleagues
histologically examined the optic nerves of 10 patients
with AD.42 Their report was the rst to conclusively
demonstrate the reduction in retinal nerve ber layer
and retinal ganglion cell number typical of AD. This
is perhaps unsurprising given the shared embryologic
origin of the cerebral cortex and eye. Emerging initially
as diverticula in the primitive diencephalon in weeks 3
to 4 of development, the primitive optic vesicle under-
goes a series of invaginations in tandem with overly-
ing mesenchyme and surface ectoderm to result in a
partition between the layers of the retina and walls
of the forebrain bounded by the optic nerve. Accord-
ingly, the cerebral cortex and eye share many charac-
teristics, including immune privilege mediated through
a combination of physical barriers, such as the blood-
retinal barrier with its resemblance to the blood-brain
barrier, and inhibitory microenvironments. Optic nerve
morphology also mirrors that of its central nervous
system counterparts ensheathed in myelin and three
meningeal layers.
Before considering the relevant literature in this
eld, it is important to appreciate the semantic distinc-
tions within neurodegenerative disease research, which
may not be immediately obvious to visual scien-
tists and ophthalmologists. Rather than describing a
disease, the term dementia is an umbrella term refer-
ring to a constellation of symptoms secondary to
impairment of higher-order cerebral functions such as
memory, language, and problem solving. Accounting
for >60% of cases of dementia in people older than
Downloaded from on 02/24/2020
Oculomics: An Eye on Systemic Disease TVST | Special Issue | Vol. 9 | No. 2 | Article 6 | 5
65 years, AD is the most common cause of demen-
tia and is classically characterized by the deposition of
amyloid beta oligomers and neurobrillary tangles.43
In contrast, vascular dementia, the second leading
cause, is intimately linked to cerebrovascular disease
and can be subcategorized depending on the imaging
characteristics of infarcts and white matter changes.
The remaining 20% are attributable to rarer forms
such as Lewy body dementia, frontotemporal degener-
ation, and dementia associated with Parkinson disease.
A large proportion of cases may also be mixed, most
commonly AD and vascular dementia, suggesting not
only that both retinal vascular morphology and nerve
ber layer may be useful but that signicant overlap
exists in biomarker measurement.
Much of the current work bridging AD and retinal
morphology revolves around OCT, but meaningful
associations have been reported in other imaging
modalities. As one of the rst available modalities,
retinal photography in those with cognitive decline
showed changes in vessel caliber and branching indices,
but this may be accounted for by the shared risk
factors between CVD and certain forms of demen-
tia. The Rotterdam Study demonstrates evidence of
retinopathy in participants with prevalent dementia,
but its presence did not appear to confer an increased
risk of incident dementia.44 The peripheral retina may
also hold clues—patients with AD appear to have
both a higher baseline and an increase in peripheral
hard drusen over two years of follow-up.45 Anatomi-
cally, however, changes in the optic nerve and retinal
nerve ber layer would seem most plausible represen-
tatives of neurodegenerative disease. Early on, it was
noted that red free retinal nerve ber layer (RNFL)
photographs reveal a higher proportion of defects
in patients with AD, but this has not always been
consistent, perhaps owing to the limited number of
46 Trick et al.47 and Berisha et
al.48 directly sought to address previous psychophysical
work revealing inferior visual eld defects in patients
with AD, nding that the superior RNFL was prefer-
entially impaired in their cohort.
The most consistent retinal feature of AD is OCT-
measured changes in the RNFL. The RNFL comprises
the axons of retinal ganglion cells that project directly
to the lateral geniculate nucleus, and thinning has been
shown to correlate with AD and cognitive decline
in two separate meta-analyses.7,10 Individuals with
AD show reductions in the RNFL, ganglion cell–
inner plexiform layer, and overall macular volume.49
The similarity of RNFL thickness between people
with established AD and mild cognitive impairment
suggests that ganglion axonal loss occurs early in
the disease and may therefore have some predic-
tive value.50,51 In 2018, two large population-based
analyses substantially bolstered our understanding
in this eld. A study by Ko et al.9analyzing the
OCTs of >30,000 participants of the UK Biobank
study found that thinner RNFL was not only associ-
ated with lower cognitive testing scores (as measured
by the Mini Mental State Examination), but also
those in the thinner quintiles were more likely to
perform poorly on the test three years later. In the
same month, researchers from the Rotterdam Study
published their ndings that thinner RNFL was associ-
ated with an increased risk of developing dementia,
highlighting the potential role of a biomarker of early
To our knowledge, there is no published work relat-
ing to the use of deep learning in the prediction
of dementia from retinal imaging, but this is likely
to change soon given current research eorts. In a
2019 systematic review, Jo et al.52 appraised 16 studies
employing deep learning in neuroimaging (magnetic
resonance imaging) for the diagnosis of AD. It there-
fore seems likely that similar applications will pervade
the retinal sphere.
We have sought to highlight the enormous poten-
tial of leveraging retinal biomarkers for the charac-
terization of systemic disease, particularly those of
rising prevalence within the aging population. Given
the size of emerging data sets and complexity of
imaging techniques, these benets may be most eec-
tively secured through modern AI techniques, such
as deep learning.53 The role of deep learning in this
eld is likely to extend beyond simple risk predic-
tion from an individual’s retinal images. In an attempt
to combat the “black box” issue of limited inter-
pretability in deep learning, groups are leveraging
methods such as saliency maps to highlight image
regions/pixels that most contribute to the model’s
decision. Given the vast quantities of data incor-
porated within modern ophthalmic imaging, such
as volumetric OCT scanning, the identication of
novel biomarkers through these methods is attractive.
Indeed, as per the work of Poplin et al.18 on the predic-
tion of cardiovascular risk factors from retinal photog-
raphy, foveal morphology appeared to be crucial in
the model’s decision of determining sex. This ability
to generate new hypotheses is exciting and should
then be validated on a separate data set. However,
this deviation from traditional hypotheses-led research
must be approached with caution, accounting for the
Downloaded from on 02/24/2020
Oculomics: An Eye on Systemic Disease TVST | Special Issue | Vol. 9 | No. 2 | Article 6 | 6
well-known issues associated with data dredging, which
frequented early genomics research.54
A further application for deep learning in this
eld comes from segmentation. Many OCT devices
now come with preinstalled automated segmentation
software, but these are generally designed for the
identication of retinal layers, rather than disease
features, and can be complicated by error. Rather than
providing a global classication for an input image,
deep learning can be employed to label each specic
pixel. This can then be fed into a further neural
network to provide an overall classication. De Fauw
et al.55 employed this technique in OCT classica-
tion for macular diseases by designing an intermedi-
ate segmentation map, which was then analyzed by
a further model to provide a classication and triage
urgency. This cascade not only standardizes anatom-
ical output independent of the acquiring device but
also permits a degree of interpretability by the clini-
cian, who can more easily discern, for example, the
presence of subretinal uid in a diagnosis of neovas-
cular age-related macular degeneration. However, the
segmentation map has signicant value independent
of a feed-forward process. Segmentation facilitates our
ability to record change more accurately in retinal
morphology and will likely be instrumental in dynamic
risk prediction models, especially given that retinal
imaging is frequently acquired on a regular ongoing
basis. Incorporation of updated cardiovascular risk
factors so as to illustrate a trajectory of change has
shown promise over more conventional risk prediction
methods, and this principle will likely extend to retinal
Among the most crucial steps in constructing deep
learning–based models is the acquisition and curation
of a training data set of sucient magnitude. These
are typically derived from either large prospective
observational cohort studies or retrospective real-world
data. In the former, the data may be available to
researchers either through open access (e.g., MESSI-
or upon application (e.g., UK Biobank: https://www. In a recent systematic review evalu-
ating 82 studies comparing the performance of deep
learning models with health care professionals in
disease classication from medical imaging, 25 studies
leveraged data from open-access repositories.57 In the
eld of oculomics, retrospective real-world data can be
a more challenging option as the desired labels (such
as myocardial infarction within ve years) will not
align with the original purpose for capture, typically
being eye disease. Moreover, ophthalmic care is often
provided in standalone ophthalmic settings. In this
situation, researchers may consider record linkage as a
possible solution. We provide details of a case example
of this in Box 1.
In this article, we have focused on two specic
exemplars, cardiovascular and neurodegenerative
disease, but there is emerging potential for generating
novel hypotheses from linking large-scale ophthalmic
data sets with other specialties. Such is the objec-
tive of initiatives such as Health Data Research UK
(, which seeks to unite health
care data nationally to facilitate innovative discovery
with a strong underpinning of public involvement and
engagement. AI-based methods in conjunction with
interdisciplinary expertise will be crucial in tackling
the future health care challenges of common chronic
disorders of the body. Undoubtedly, oculomics will be
one of the keys to these eorts.
Box 1. Case Example: AlzEye—Linking
Ophthalmic Imaging and Systemic Disease
Labels at Scale to Provide New Insights into
Dementia (and Cardiovascular Disease)
When trying to achieve the necessary scale of data
for machine learning approaches, the use of routinely
collected data is an attractive alternative to the
high-cost, researcher-led data sets compiled through
epidemiologic studies or biobanks. One of the aims of
such an approach is to create virtual biobanks much
cheaper than otherwise possible (arguably a “biobank-
on-a-shoestring”) and which may indeed better reect
the population of interest (vs. the somewhat skewed
population that has been observed in some biobank
An example of this kind of approach is AlzEye, the
United Kingdom’s rst and largest linkage of complex
three-dimensional imaging data (fundus photographs
and retinal OCT) to systemic health diagnostic codes
for the purposes of exploring retinal ultrastruc-
tural associations and predictors of dementia and
its subtypes. AlzEye depends on the combination of
both local and nationally held data sets within the
United Kingdom’s National Health Service (NHS).
Specically, AlzEye is a pseudonymized data set
linking retinal photographs and OCT scans of all
patients older than 40 years attending Moorelds
Eye Hospital NHSFT with Hospital Episode Statis-
tics (HES), a national database consisting of all admis-
sions, emergency attendances, and outpatient appoint-
ments in England. The appropriate use and linkage of
such data depend on satisfying many criteria, includ-
ing ethical approval, data security, and governance.
Engagement with the public has been pivotal to the
approach. We surveyed 483 participants to canvass
Downloaded from on 02/24/2020
Oculomics: An Eye on Systemic Disease TVST | Special Issue | Vol. 9 | No. 2 | Article 6 | 7
public opinion on the use of eye scans for research and
the acceptability of large data sets to identify patterns
of systemic disease. Two members of the public sit on
the AlzEye working group, and information regarding
the study is outlined on the funding charity’s website.
This kind of study is complex, and the approval process
that AlzEye underwent was appropriately robust with
a number of dierent approvals required prior to the
establishment of AlzEye. Although the exact process
will vary from country to country, the processes are
likely to share similar principles, and we therefore
highlight them here. The rst stage required us to secure
a research sponsor, necessitating institutional approval
consisting of research and development, informa-
tion governance, and information technology at both
the NHS data custodian (Moorelds Eye Hospital
NHSFT) and the research institute (University College
London). Important conditions involving third-party
linkage by a “trusted third party,” robust data privacy
measures, and sucient computing infrastructure were
outlined at this stage. In AlzEye, the linkage process
is as follows: (1) images from Moorelds Eye Hospital
are pseudonymized through the removal of all identi-
ers and replacement with a unique study ID. These
are then transferred to University College London. (2)
Simultaneously, a spreadsheet of the image identiers
(date of birth, unique NHS number, sex) is securely sent
to NHS Digital, the national body overseeing the HES
data warehouse. (3) NHS Digital strips the identiers
and returns the relevant HES data with pseudonymized
study IDs to University College London, where it is
linked with corresponding images. Thus, HES data
never enter the source of imaging data (Moorelds
Eye Hospital), and conversely, identiers never enter
University College London (Fig. 1).
Prior to commencement, all research studies in the
United Kingdom require ethical approval through the
Research Ethics Service, but some specic studies may
warrant additional approvals. AlzEye was approved by
the National Health Service Research Ethics Commit-
tee in 2018. Due to the large number of patients
included (more than 250,000), the historical nature
of the data, and the advanced age and diculty in
contacting patients, it would not be feasible to obtain
consent from patients. Therefore, to use identiable
data for the linkage, a specic type of approval was
sought involving an application to the Condential
Advisory Group, who advise the UK Health Research
Authority on whether sucient justication exists to
access data without consent. In the United Kingdom,
this is known as a “Section 251 approval,” deriving
from the 2006 NHS Act, which provides provision for
this kind of application. The Health Research Author-
Figure 1. The ow of data is such that the Moorelds Eye Hospi-
tal never receives HES data and University College London does not
receive any identiers. University College London, as a trusted third
party, links images from Moorelds Eye Hospital with HES data from
NHS Digital based on a unique study ID.
ity, collating the opinions of the respective committees,
granted ultimate approval in late 2018.
Upon these approvals, applications to NHS Digital for
the procurement of HES data can then be processed.
In addition to the external approvals, NHS Digital
has its own internal approval process detailing, in
particular, the legal basis upon which data are being
accessed. When a given application is approved, it is
then presented on behalf of the applicant by NHS
Digital to the Independent Group Advising on the
Release of Data (IGARD), a committee of specialist
and lay members who assess all applications to NHS
Digital for the dissemination of condential informa-
tion. In January 2019, IGARD gave approval, remark-
ing that the AlzEye application “could be used as an
exemplar to help other researchers with their applica-
tions to the Data Access Request Service.”
Supported by a Springboard Grant from the
Moorelds Eye Charity (EK) and UK National Insti-
tute for Health Research (NIHR) Clinician Scien-
tist Award (NIHR-CS–2014-12-023; PAK). The views
expressed are those of the author and not necessarily
those of the NHS, the NIHR, or the Department of
AKD and PAK contributed equally to this work.
Disclosure: S.K. Wagner, None; D.J. Fu, None;
L. Faes, None; X. Liu, None; J. Huemer, None;
Downloaded from on 02/24/2020
Oculomics: An Eye on Systemic Disease TVST | Special Issue | Vol. 9 | No. 2 | Article 6 | 8
H. Khalid, None; D. Ferraz, None; E. Korot, Google
Health (E); C. Kelly, None; K. Balaskas, Alimera (F),
Allergan (F), Bayer (F), Heidelberg Engineering (F),
Novartis (F), TopCon (F); A.K. Denniston, None;
P.A. Keane, Heidelberg Engineering (F), Topcon (F),
Carl Zeiss Meditec (F), Haag-Streit (F), Allergan (F),
Novartis (F, S), Bayer (F, S), DeepMind (C), Optos (C)
1. Kim DH, Chaves PH, Newman AB, et al. Reti-
nal microvascular signs and disability in the
Cardiovascular Health Study. Arch Ophthalmol.
2. Wong TY, Mohamed Q, Klein R, Couper DJ. Do
retinopathy signs in non-diabetic individuals pre-
dict the subsequent risk of diabetes? Br J Ophthal-
mol. 2006;90:301–303.
3. Gunthner R, Hanssen H, Hauser C, et al. Impaired
retinal vessel dilation predicts mortality in end-
stage renal disease [published online April 1, 2019].
Circ Res.
4. McGeechan K, Liew G, Macaskill P, et al. Meta-
analysis: retinal vessel caliber and risk for coro-
nary heart disease. Ann Intern Med. 2009;151:404–
5. McGeechan K, Liew G, Macaskill P, et al. Pre-
diction of incident stroke events based on retinal
vessel caliber: a systematic review and individual-
participant meta-analysis. Am J Epidemiol. 2009;
6. Wong TY, Mitchell P. Hypertensive retinopathy. N
Engl J Med. 2004;351:2310–2317.
7. Thomson KL, Yeo JM, Waddell B, Cameron JR,
Pal S. A systematic review and meta-analysis of
retinal nerve ber layer change in dementia, using
optical coherence tomography. Alzheimers Dement
(Amst). 2015;1:136–143.
8. Mutlu U, Colijn JM, Ikram MA, et al. Association
of retinal neurodegeneration on optical coherence
tomography with dementia: a population-based
study. JAMA Neurol. 2018;75:1256–1263.
9. Ko F, Muthy ZA, Gallacher J, et al. Associ-
ation of retinal nerve ber layer thinning with
current and future cognitive decline: a study
using optical coherence tomography. JAMA Neu-
rol. 2018;75:1198–1205.
10. Coppola G, Di Renzo A, Ziccardi L, et al.
Optical coherence tomography in Alzheimer’s
disease: a meta-analysis. PLoS One. 2015;10:
11. Duke-Elder SS, ed. Volume XV Summary of Sys-
temic Ophthalmology. London, UK: Henry Kimp-
ton; 1976.
12. Fujimoto J, Swanson E. The development, com-
mercialization, and impact of optical coherence
tomography. Invest Ophthalmol Vis Sci. 2016;57:1–
13. Pontikos NWS, Faes L, Kortuem K, Fasler K, Bal-
askas K, Keane PA. Comment on: Trends in retina
specialist imaging utilization from 2012 to 2016 in
the United States Medicare fee-for-service popula-
tion. Am J Ophthalmol. In press.
14. OCT rollout in every Specsavers Announcted.
Optometry Today (2017). Available at https://
Accessed September 16, 2019.
15. Wong TY, Klein R, Couper DJ, et al. Retinal
microvascular abnormalities and incident stroke:
the Atherosclerosis Risk in Communities Study.
Lancet. 2001;358:1134–1140.
16. Wong TY, Klein R, Sharrett AR, et al. Retinal
arteriolar narrowing and risk of coronary heart
disease in men and women: the Atherosclero-
sis Risk in Communities Study. JAMA. 2002;
17. Wong TY, Klein R, Sharrett AR, et al. The preva-
lence and risk factors of retinal microvascular
abnormalities in older persons: the Cardiovascular
Health Study. Ophthalmology. 2003;110:658–666.
18. Poplin R, Varadarajan AV, Blumer K, et al. Pre-
diction of cardiovascular risk factors from retinal
fundus photographs via deep learning. Nat Biomed
Eng. 2018;2:158–164.
19. Bright R. Tabular view of the morbid appearances
in 100 cases connected with albuminous urine.
Guy’s Hosp Rep. 1836;1:380–400.
20. Gunn R. Ophthalmoscopic evidence of (1) arterial
changes associated with chronic renal diseases and
(2) of increased arterial tension. Trans Ophthalmol
Soc UK. 1892;12:124–125.
21. Keith NM,Wagener HP, Barker NW. Some dif-
ferent types of essential hypertension: their course
and their prognosis. Am J Med Sci. 1939;197:332–
22. World Health Organization. Global Health Esti-
mates 2016: Disease Burden by Cause, Age, Sex,
by Country and by Region, 2000–2016.Geneva,
Switzerland: WHO; 2018.
23. Arnett DK, Blumenthal RS, Albert MA, et al.
2019 ACC/AHA guideline on the primary pre-
vention of cardiovascular disease: executive sum-
mary: a report of the American College of Car-
diology/American Heart Association Task Force
Downloaded from on 02/24/2020
Oculomics: An Eye on Systemic Disease TVST | Special Issue | Vol. 9 | No. 2 | Article 6 | 9
on Clinical Practice Guidelines. J Am Coll Cardiol.
24. Ridker PM, Statins Cook NR.: new American
guidelines for prevention of cardiovascular disease.
Lancet. 2013;382:1762–1765.
25. Kuragaichi T, Kataoka Y, Miyakoshi C,
Miyamoto T, Sato Y. External validation of
pooled cohort equations using systolic blood
pressure intervention trial data. BMC Res Notes.
26. Enoch J, McDonald L, Jones L, Jones PR, Crabb
DP. Evaluating whether sight is the most valued
sense [published online October 3, 2019]. JAMA
27. Robson J, Dostal I, Sheikh A, et al. The NHS
Health Check in England: an evaluation of the rst
4 years. BMJ Open. 2016;6:e008840.
28. The State of the National Eye Health 2017: A Year
in Review. RNIB (2018). Available at https://www.les/APDF%20The%20
Accessed September 5, 2019.
29. Ho H, Cheung CY, Sabanayagam C, et al.
Retinopathy Signs improved prediction and reclas-
sication of cardiovascular disease risk in diabetes:
a prospective cohort study. Scientic Reports.
30. Cheung CY, Tay WT, Ikram MK, et al. Retinal
microvascular changes and risk of stroke: the Sin-
gapore Malay Eye Study. Stroke. 2013;44:2402–
31. Newsom RS, Sullivan PM, Rassam SM, Jagoe R,
Kohner EM. Retinal vessel measurement: com-
parison between observer and computer driven
methods. Graefes Arch Clin Exp Ophthalmol.
32. Rassam SM, Patel V, Brinchmann-Hansen O,
Engvold O, Kohner EM. Accurate vessel width
measurement from fundus photographs: a new
concept. Br J Ophthalmol. 1994;78:24–29.
33. Longstreth W, Larsen EKM, Klein R, et al. Asso-
ciations between Findings on cranial magnetic res-
onance imaging and retinal photography in the
elderly: the Cardiovascular Health Study. Am J
Epidemiol. 2006;165:78–84.
34. Klein R, Klein B, Moss S, Wong T, Sharrett
A. Retinal vascular caliber in persons with type
2 diabetes: the Wisconsin Epidemiological Study
of Diabetic Retinopathy: XX. Ophthalmology.
35. Owen CG, Rudnicka AR, Welikala RA, et al.
Retinal vasculometry associations with car-
diometabolic risk factors in the European
Prospective Investigation of Cancer–Norfolk
Study. Ophthalmology. 2019;126:96–106.
36. Ting DSW, Wong TY. Eyeing cardiovascular risk
factors. Nat Biomed Eng. 2018;2:140–141.
37. World Alzheimer Report 2015. The Global Impact
of Dementia: An Analysis of Prevalence, Inci-
dence, Cost and Trends. Alzheimer’s Disease Inter-
national (2015). Available at: https://www.alz. Published by
Alzheimer’s disease International, London, United
Kingdom. Accessed September 1, 2019.
38. Deaths registered in England and Wales (series
DR): 2017. Oce for National Statistics (2019).
Available at
walesseriesdr/2017. Accessed September 1, 2019.
39. Hansson O, Lehmann S, Otto M, Zetterberg H,
Lewczuk P. Advantages and disadvantages of the
use of the CSF Amyloid beta (Abeta) 42/40 ratio
in the diagnosis of Alzheimer’s disease. Alzheimers
Res Ther. 2019;11:34.
40. Struyfs H, Van Broeck B, Timmers M, et al. Diag-
nostic accuracy of cerebrospinal uid amyloid-
beta isoforms for early and dierential demen-
tia diagnosis. J Alzheimers Dis. 2015;45:813–
41. Seo EH, Park WY, Choo IH. Structural MRI and
amyloid PET imaging for prediction of conversion
to Alzheimer’s disease in patients with mild cogni-
tive impairment: a meta-analysis. Psychiatry Inves-
tig. 2017;14:205–215.
42. Hinton DR, Sadun AA, Blanks JC, Miller CA.
Optic-nerve degeneration in Alzheimer’s disease. N
Engl J Med. 1986;315:485–487.
43. Prince M, Knapp M, Guerchet M, et al. Dementia
UK. London, United Kingdom: Alzheimer’s Soci-
ety 2nd ed. 2014.
44. Schrijvers EM, Buitendijk GH, Ikram MK, et al.
Retinopathy and risk of dementia: the Rotterdam
Study. Neurology. 2012;79:365–370.
45. Csincsik L, MacGillivray TJ, Flynn E, et al.
Peripheral retinal imaging biomarkers for
Alzheimer’s disease: a pilot study. Ophthalmic
Res. 2018;59:182–192.
46. Hedges TR III, Perez Galves R, Speigelman D,
Barbas NR, Peli E, Yardley CJ. Retinal nerve ber
layer abnormalities in Alzheimer’s disease. Acta
Ophthalmol Scand. 1996;74:271–275.
47. Trick GL, Trick LR, Morris P, Wolf M. Visual
eld loss in senile dementia of the Alzheimer’s type.
Neurology. 1995;45:68–74.
48. Berisha F, Feke GT, Trempe CL, McMeel JW,
Schepens CL. Retinal abnormalities in early
Downloaded from on 02/24/2020
Oculomics: An Eye on Systemic Disease TVST |SpecialIssue|Vol.9|No.2|Article6|10
Alzheimer’s disease. Invest Ophthalmol Vis Sci.
49. Cheung CY, Chan VTT, Mok VC, Chen C, Wong
TY. Potential retinal biomarkers for dementia:
what is new? Curr Opin Neurol. 2019;32:82–91.
50. Kesler A, Vakhapova V, Korczyn AD, Naftal-
iev E, Neudorfer M. Retinal thickness in patients
with mild cognitive impairment and Alzheimer’s
disease. Clin Neurol Neurosurg. 2011;113:523–
51. Paquet C, Boissonnot M, Roger F, Dighiero P, Gil
R, Hugon J. Abnormal retinal thickness in patients
with mild cognitive impairment and Alzheimer’s
disease. Neurosci Lett. 2007;420:97–99.
52. Jo T, Nho K, Saykin AJ. Deep learning in
Alzheimer’s disease: diagnostic classication and
prognostic prediction using neuroimaging data.
Front Aging Neurosci. 2019;11:220.
53. LeCun Y, Bengio Y, Hinton G. Deep learning.
Nature. 2015;521:436–444.
54. Lee JK, Williams PD, Cheon S. Data mining
in genomics. Clin Lab Med. 2008;28:145–166,
55. De Fauw J, Ledsam JR, Romera-Paredes B,
et al. Clinically applicable deep learning for diag-
nosis and referral in retinal disease. Nat Med.
56. Teramukai S, Okuda Y, Miyazaki S, Kawamori
R, Shirayama M, Teramoto T. Dynamic pre-
diction model and risk assessment chart for
cardiovascular disease based on on-treatment
blood pressure and baseline risk factors. Hypertens
Res. 2016;39:113–118.
57. Liu X, Faes L, Kale AU, et al. A comparison of
deep learning performance against health-care pro-
fessionals in detecting diseases from medical imag-
ing: a systematic review and meta-analysis. Lancet
Digital Health. 2019;1:e271–e297.
Downloaded from on 02/24/2020
... [4][5][6][7][8][9][10][11][12][13] With the booming OCT deployment in primary care settings and people's concerns for eye health, the easily accessible, risk-free, and highresolution retinal scans are becoming an attractive alternative for screening systemic health in routine community scenarios. [14][15][16][17] The biological link between alterations in retinal layers and systemic health remains unknown. Metabolomics offers a novel opportunity to study the biological signatures behind these complex features, 18 especially considering that metabolic risk factors contribute substantially to various age-related diseases. ...
Full-text available
Background: The retina is considered a unique window to systemic health, but their biological link remains unknown. Methods: A total of 93,838 UK Biobank participants with metabolomics data were included in the study. Plasma metabolites associated with GCIPLT were identified in 7,824 participants who also underwent retinal optical coherence tomography; prospective associations of GCIPLT-associated metabolites with 12-year risk of mortality and major age-related diseases were assessed in 86,014 participants. The primary outcomes included all- and specific-cause mortality. The secondary outcomes included incident type 2 diabetes mellitus (T2DM), obstructive sleep apnea/hypopnea syndrome (OSAHS), myocardial infarction (MI), heart failure, ischemic stroke, and dementia. C-statistics and net reclassification indexes (NRIs) were calculated to evaluate the added predictive value of GCIPLT metabolites. Calibration was assessed using calibration plots. Findings: Sixteen metabolomic signatures were associated with GCIPLT (P< 0.009 [Bonferroni-corrected threshold]), and most were associated with the future risk of mortality and age-related diseases. The constructed meta-GCIPLT scores distinguished well between patients with high and low risks of mortality and morbidity, showing predictive values higher than or comparable to those of traditional risk factors (C-statistics: 0.780[0.771-0.788], T2DM; 0.725[0.707-0.743], OSAHS; 0.711[0.695-0.726], MI; 0.685[0.662-0.707], cardiovascular mortality; 0.657[0.640-0.674], heart failure; 0.638[0.636-0.660], other mortality; 0.630[0.618-0.642], all-cause mortality; 0.620[0.598-0.643], dementia; 0.614[0.593-0.634], stroke; and 0.601[0.585-0.617], cancer mortality). The NRIs confirmed the inclusion of GCIPLT metabolomic signatures to the models based on traditional risk factors resulted in significant improvements in model performance (5.18%, T2DM [P=3.86E-11]; 4.43%, dementia [P=0.003]; 4.20%, cardiovascular mortality [P=6.04E-04]; 3.73%, MI [P=1.72E-07]; 2.93%, OSAHS [P=3.13E-05]; 2.39%, all-cause mortality [P=3.89E-05]; 2.33%, stroke [P=0.049]; 2.09%, cancer mortality [P=0.039]; and 1.59%, heart failure [P=2.72E-083.07E-04]). Calibration plots showed excellent calibration between predicted risk and actual incidence in the new models. Interpretation: GCIPLT-associated plasma metabolites captured the residual risk for mortality and major systemic diseases not quantified by traditional risk factors in the general population. Incorporating GCIPLT metabolomic signatures into prediction models may assist in screening for future risks of these health outcomes.
... The eye is the only organ in which we can see the cardiovascular blood vessels and an inner blood-retinal barrier (a tight protective layer of cells and capillaries that prevent larger molecules from entering the retina) similar to the blood-brain barrier (London et al., 2013). Consequently, retinal images are also used to analyze systemic diseases such as cardiovascular disease (Wagner et al., 2020). Since the eye is more accessible than the (Palanker, 2016) (b) Raw image taken from OCT and marked by Heidelberg (2010). ...
Retinal optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA) are promising tools for the (early) diagnosis of Alzheimer's disease (AD). These non-invasive imaging techniques are cost-effective and more accessible than alternative neuroimaging tools. However, interpreting and classifying multi-slice scans produced by OCT devices is time-consuming and challenging even for trained practitioners. There are surveys on machine learning and deep learning approaches concerning the automated analysis of OCT scans for various diseases such as glaucoma. However, the current literature lacks an extensive survey on the diagnosis of Alzheimer's disease or cognitive impairment using OCT or OCTA. This has motivated us to do a comprehensive survey aimed at machine/deep learning scientists or practitioners who require an introduction to the problem. The paper contains 1) an introduction to the medical background of Alzheimer's Disease and Cognitive Impairment and their diagnosis using OCT and OCTA imaging modalities, 2) a review of various technical proposals for the problem and the sub-problems from an automated analysis perspective, 3) a systematic review of the recent deep learning studies and available OCT/OCTA datasets directly aimed at the diagnosis of Alzheimer's Disease and Cognitive Impairment. For the latter, we used Publish or Perish Software to search for the relevant studies from various sources such as Scopus, PubMed, and Web of Science. We followed the PRISMA approach to screen an initial pool of 3073 references and determined ten relevant studies (N=10, out of 3073) that directly targeted AD diagnosis. We identified the lack of open OCT/OCTA datasets (about Alzheimer's disease) as the main issue that is impeding the progress in the field.
... Oculomics, a rapidly developing field dedicated to identifying ocular biomarkers for systemic disease, is gaining attention for several cardiovascular and neurological applications [7]. A variety of retinal microvasculature features have become recognizable for their associations with systemic disease states [8]. ...
Full-text available
Sickle cell disease (SCD) exists on a phenotypic spectrum with variable genetic expressivity, making it difficult to assess an individual patient’s risk of complications at any particular point in time. Current and emerging SCD treatments, including CRISPR-based gene editing, result in a variable proportion of affected red blood cells (RBCs) still vulnerable to sickling. Clinical serological indicators of disease such as hemoglobin, indirect bilirubin, and reticulocyte count and clinical metrics including number of emergency department visits and hospitalizations over time often fall short in their ability to objectively quantify ischemic disease activity and efficacy of treatments. Clearly, better clinical biomarkers are needed. The rapidly developing field of oculomics leverages the transparent nature of the ocular tissue to directly study the retinal microvasculature in order to characterize the status of systemic diseases. In this case report, we demonstrate the ability of optical coherence tomography angiography (OCT-A) to detect and measure micro-occlusive events within the retinal capillary bed before and after RBC exchange transfusion and following CRISPR-based gene editing, as an indicator of systemic ischemic disease activity and measure of treatment efficacy. The implications of these findings are discussed.
... The retina is the only tissue of the human body that allows direct noninvasive visualization of microvasculature. The relations between various systemic diseases, including heart, and changes in ocular microvasculature has been a focus of studies in recent years [2][3][4]. Alternations in the thickness of the retina and its layers as well as changes in retinal microvasculature parameters was shown to be related to cardiovascular health [5][6][7]. Poplin et al., in 2018, developed an algorithm based on about 300,000 images of retinal fundus. ...
Full-text available
Introduction. Optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA) allowed visualization of retina and choroid to nearly the capillary level; however, the relationship between systemic macrovascular status and retinal microvascular changes is not yet known well. Aim. Our purpose was to assess the impact of retinal optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA) parameters on prediction of coronary heart disease (CHD) in acute myocardial infarction (MI) and chronic three vessel disease (3VD) groups. Methods. This observational study included 184 patients—26 in 3VD, 76 in MI and 82 in healthy participants groups. Radial scans of the macula and OCTA scans of the central macula (superficial (SCP) and deep (DCP) capillary plexuses) were performed on all participants. All participants underwent coronary angiography. Results. Patients in MI groups showed decreased parafoveal total retinal thickness as well as GCL+ retinal thickness. Outer circle total retinal thickness and GCL+ retinal thickness were lowest in the 3VD group. The MI group had thinner, while 3VD the thinnest, choroid. A decrease in choroidal thickness and vascular density could predict 3VD. Conclusions. A decrease in retinal and choroidal thickness as well as decreased vascular density in the central retinal region may predict coronary artery disease. OCT and OCTA could be a significant, safe, and noninvasive tool for the prediction of coronary artery disease.
... Oculomics is a new field for identifying biomarkers for systemic diseases using artificial intelligence and extensive eye health data [10]. Eye examinations are frequently performed for a wide range of vision-related clinical indications [11]. ...
Aims: Sarcopenia is characterized by a gradual loss of skeletal muscle mass and strength with increased adverse outcomes. Recently, large-scale epidemiological studies have demonstrated a relationship between several chronic disorders and ocular pathological conditions using an oculomics approach. We hypothesized that sarcopenia can be predicted through eye examinations, without invasive tests or radiologic evaluations in the context of predictive, preventive, and personalized medicine (PPPM/3PM). Methods: We analyzed data from the Korean National Health and Nutrition Examination Survey (KNHANES). The training set (80%, randomly selected from 2008 to 2010) data were used to construct the machine learning models. Internal (20%, randomly selected from 2008 to 2010) and external (from the KNHANES 2011) validation sets were used to assess the ability to predict sarcopenia. We included 8092 participants in the final dataset. Machine learning models (XGBoost) were trained on ophthalmological examinations and demographic factors to detect sarcopenia. Results: In the exploratory analysis, decreased levator function (odds ratio [OR], 1.41; P value <0.001), cataracts (OR, 1.31; P value = 0.013), and age-related macular degeneration (OR, 1.38; P value = 0.026) were associated with an increased risk of sarcopenia in men. In women, an increased risk of sarcopenia was associated with blepharoptosis (OR, 1.23; P value = 0.038) and cataracts (OR, 1.29; P value = 0.010). The XGBoost technique showed areas under the receiver operating characteristic curves (AUCs) of 0.746 and 0.762 in men and women, respectively. The external validation achieved AUCs of 0.751 and 0.785 for men and women, respectively. For practical and fast hands-on experience with the predictive model for practitioners who may be willing to test the whole idea of sarcopenia prediction based on oculomics data, we developed a simple web-based calculator application ( to predict the risk of sarcopenia and facilitate screening, based on the model established in this study. Conclusion: Sarcopenia is treatable before the vicious cycle of sarcopenia-related deterioration begins. Therefore, early identification of individuals at a high risk of sarcopenia is essential in the context of PPPM. Our oculomics-based approach provides an effective strategy for sarcopenia prediction. The proposed method shows promise in significantly increasing the number of patients diagnosed with sarcopenia, potentially facilitating earlier intervention. Through patient oculometric monitoring, various pathological factors related to sarcopenia can be simultaneously analyzed, and doctors can provide personalized medical services according to each cause. Further studies are needed to confirm whether such a prediction algorithm can be used in real-world clinical settings to improve the diagnosis of sarcopenia. Supplementary information: The online version contains supplementary material available at 10.1007/s13167-022-00292-3.
... Our results are in line with other reports which have demonstrated that DL algorithms can make use of retinal fundus images to predict modifiable CVD risk factors, including diabetes, hypertension, and cholesterol [25,28,30,[46][47][48] and non modifiable risk factors such as chronological ager [24]. However, like the Framingham equations, the existing algortihms are unable examine the relative contribution of each of the individual factors that comprise risk as they utilise a statistical method which imposes linearity between these individual parameters during analaysis. ...
Full-text available
Purpose: To create and evaluate the accuracy of an artificial intelligence Deep learning platform (ORAiCLE) capable of using only retinal fundus images to predict both an individuals overall 5 year cardiovascular risk (CVD) and the relative contribution of the component risk factors that comprise this risk. Methods: We used 165,907 retinal images from a database of 47,236 patient visits. Initially, each image was paired with biometric data age, ethnicity, sex, presence and duration of diabetes a HDL/LDL ratios as well as any CVD event wtihin 5 years of the retinal image acquisition. A risk score based on Framingham equations was calculated. The real CVD event rate was also determined for the individuals and overall population. Finally, ORAiCLE was trained using only age, ethnicity, sex plus retinal images. Results: Compared to Framingham-based score, ORAiCLE was up to 12% more accurate in prediciting cardiovascular event in he next 5-years, especially for the highest risk group of people. The reliability and accuracy of each of the restrictive models was suboptimal to ORAiCLE performance ,indicating that it was using data from both sets of data to derive its final results. Conclusion: Retinal photography is inexpensive and only minimal training is required to acquire them as fully automated, inexpensive camera systems are now widely available. As such, AI-based CVD risk algorithms such as ORAiCLE promise to make CV health screening more accurate, more afforadable and more accessible for all. Furthermore, ORAiCLE unique ability to assess the relative contribution of the components that comprise an individuals overall risk would inform treatment decisions based on the specific needs of an individual, thereby increasing the likelihood of positive health outcomes.
... This will ultimately improve holistic patient care beyond ophthalmology by allowing patients to be diagnosed with various conditions in a non-invasive manner, which can be done by a trained technician and automated AI analysis. Future studies should apply similar DL models to the prediction and classification of systemic features relevant to other pathologies that have both a high public health burden and a potential ophthalmic manifestation beyond cardiovascular disease and diabetes, such as Alzheimer's disease [26]. ...
While color fundus photos are used in routine clinical practice to diagnose ophthalmic conditions, evidence suggests that ocular imaging contains valuable information regarding the systemic health features of patients. These features can be identified through computer vision techniques including deep learning (DL) artificial intelligence (AI) models. We aim to construct a DL model that can predict systemic features from fundus images and to determine the optimal method of model construction for this task. Data were collected from a cohort of patients undergoing diabetic retinopathy screening between March 2020 and March 2021. Two models were created for each of 12 systemic health features based on the DenseNet201 architecture: one utilizing transfer learning with images from ImageNet and another from 35,126 fundus images. Here, 1277 fundus images were used to train the AI models. Area under the receiver operating characteristics curve (AUROC) scores were used to compare the model performance. Models utilizing the ImageNet transfer learning data were superior to those using retinal images for transfer learning (mean AUROC 0.78 vs. 0.65, p-value < 0.001). Models using ImageNet pretraining were able to predict systemic features including ethnicity (AUROC 0.93), age > 70 (AUROC 0.90), gender (AUROC 0.85), ACE inhibitor (AUROC 0.82), and ARB medication use (AUROC 0.78). We conclude that fundus images contain valuable information about the systemic characteristics of a patient. To optimize DL model performance, we recommend that even domain specific models consider using transfer learning from more generalized image sets to improve accuracy.
Full-text available
Purpose of review: In this review, we consider the challenges of creating a trusted resource for real-world data in ophthalmology, based on our experience of establishing INSIGHT, the UK's Health Data Research Hub for Eye Health and Oculomics. Recent findings: The INSIGHT Health Data Research Hub maximizes the benefits and impact of historical, patient-level UK National Health Service (NHS) electronic health record data, including images, through making it research-ready including curation and anonymisation. It is built around a shared 'north star' of enabling research for patient benefit. INSIGHT has worked to establish patient and public trust in the concept and delivery of INSIGHT, with efficient and robust governance processes that support safe and secure access to data for researchers. By linking to systemic data, there is an opportunity for discovery of novel ophthalmic biomarkers of systemic diseases ('oculomics'). Datasets that provide a representation of the whole population are an important tool to address the increasingly recognized threat of health data poverty. Summary: Enabling efficient, safe access to routinely collected clinical data is a substantial undertaking, especially when this includes imaging modalities, but provides an exceptional resource for research. Research and innovation built on inclusive real-world data is an important tool in ensuring that discoveries and technologies of the future may not only favour selected groups, but also work for all patients.
Purpose of review: Retinal microvasculature assessment has shown promise to enhance cardiovascular disease (CVD) risk stratification. Integrating artificial intelligence into retinal microvasculature analysis may increase the screening capacity of CVD risks compared with risk score calculation through blood-taking. This review summarizes recent advancements in artificial intelligence based retinal photograph analysis for CVD prediction, and suggests challenges and future prospects for translation into a clinical setting. Recent findings: Artificial intelligence based retinal microvasculature analyses potentially predict CVD risk factors (e.g. blood pressure, diabetes), direct CVD events (e.g. CVD mortality), retinal features (e.g. retinal vessel calibre) and CVD biomarkers (e.g. coronary artery calcium score). However, challenges such as handling photographs with concurrent retinal diseases, limited diverse data from other populations or clinical settings, insufficient interpretability and generalizability, concerns on cost-effectiveness and social acceptance may impede the dissemination of these artificial intelligence algorithms into clinical practice. Summary: Artificial intelligence based retinal microvasculature analysis may supplement existing CVD risk stratification approach. Although technical and socioeconomic challenges remain, we envision artificial intelligence based microvasculature analysis to have major clinical and research impacts in the future, through screening for high-risk individuals especially in less-developed areas and identifying new retinal biomarkers for CVD research.
Deep learning has seen tremendous interest in medical imaging, particularly in the use of convolutional neural networks (CNNs) for developing automated diagnostic tools. The facility of its non-invasive acquisition makes retinal fundus imaging amenable to such automated approaches. Recent work in analyzing fundus images using CNNs relies on access to massive data for training and validation - hundreds of thousands of images. However, data residency and data privacy restrictions stymie the applicability of this approach in medical settings where patient confidentiality is a mandate. Here, we showcase results for the performance of DL on small datasets to classify patient sex from fundus images - a trait thought not to be present or quantifiable in fundus images until recently. We fine-tune a Resnet-152 model whose last layer has been modified for binary classification. In several experiments, we assess performance in the small dataset context using one private (DOVS) and one public (ODIR) data source. Our models, developed using approximately 2500 fundus images, achieved test AUC scores of up to 0.72 (95% CI: [0.67, 0.77]). This corresponds to a mere 25% decrease in performance despite a nearly 1000-fold decrease in the dataset size compared to prior work in the literature. Even with a hard task like sex categorization from retinal images, we find that classification is possible with very small datasets. Additionally, we perform domain adaptation experiments between DOVS and ODIR; explore the effect of data curation on training and generalizability; and investigate model ensembling to maximize CNN classifier performance in the context of small development datasets.
Full-text available
Background: Deep learning offers considerable promise for medical diagnostics. We aimed to evaluate the diagnostic accuracy of deep learning algorithms versus health-care professionals in classifying diseases using medical imaging. Methods: In this systematic review and meta-analysis, we searched Ovid-MEDLINE, Embase, Science Citation Index, and Conference Proceedings Citation Index for studies published from Jan 1, 2012, to June 6, 2019. Studies comparing the diagnostic performance of deep learning models and health-care professionals based on medical imaging, for any disease, were included. We excluded studies that used medical waveform data graphics material or investigated the accuracy of image segmentation rather than disease classification. We extracted binary diagnostic accuracy data and constructed contingency tables to derive the outcomes of interest: sensitivity and specificity. Studies undertaking an out-of-sample external validation were included in a meta-analysis, using a unified hierarchical model. This study is registered with PROSPERO, CRD42018091176. Findings: Our search identified 31 587 studies, of which 82 (describing 147 patient cohorts) were included. 69 studies provided enough data to construct contingency tables, enabling calculation of test accuracy, with sensitivity ranging from 9·7% to 100·0% (mean 79·1%, SD 0·2) and specificity ranging from 38·9% to 100·0% (mean 88·3%, SD 0·1). An out-of-sample external validation was done in 25 studies, of which 14 made the comparison between deep learning models and health-care professionals in the same sample. Comparison of the performance between health-care professionals in these 14 studies, when restricting the analysis to the contingency table for each study reporting the highest accuracy, found a pooled sensitivity of 87·0% (95% CI 83·0-90·2) for deep learning models and 86·4% (79·9-91·0) for health-care professionals, and a pooled specificity of 92·5% (95% CI 85·1-96·4) for deep learning models and 90·5% (80·6-95·7) for health-care professionals. Interpretation: Our review found the diagnostic performance of deep learning models to be equivalent to that of health-care professionals. However, a major finding of the review is that few studies presented externally validated results or compared the performance of deep learning models and health-care professionals using the same sample. Additionally, poor reporting is prevalent in deep learning studies, which limits reliable interpretation of the reported diagnostic accuracy. New reporting standards that address specific challenges of deep learning could improve future studies, enabling greater confidence in the results of future evaluations of this promising technology. Funding: None.
Full-text available
Deep learning, a state-of-the-art machine learning approach, has shown outstanding performance over traditional machine learning in identifying intricate structures in complex high-dimensional data, especially in the domain of computer vision. The application of deep learning to early detection and automated classification of Alzheimer's disease (AD) has recently gained considerable attention, as rapid progress in neuroimaging techniques has generated large-scale multimodal neuroimaging data. A systematic review of publications using deep learning approaches and neuroimaging data for diagnostic classification of AD was performed. A PubMed and Google Scholar search was used to identify deep learning papers on AD published between January 2013 and July 2018. These papers were reviewed, evaluated, and classified by algorithm and neuroimaging type, and the findings were summarized. Of 16 studies meeting full inclusion criteria, 4 used a combination of deep learning and traditional machine learning approaches, and 12 used only deep learning approaches. The combination of traditional machine learning for classification and stacked auto-encoder (SAE) for feature selection produced accuracies of up to 98.8% for AD classification and 83.7% for prediction of conversion from mild cognitive impairment (MCI), a prodromal stage of AD, to AD. Deep learning approaches, such as convolutional neural network (CNN) or recurrent neural network (RNN), that use neuroimaging data without pre-processing for feature selection have yielded accuracies of up to 96.0% for AD classification and 84.2% for MCI conversion prediction. The best classification performance was obtained when multimodal neuroimaging and fluid biomarkers were combined. Deep learning approaches continue to improve in performance and appear to hold promise for diagnostic classification of AD using multimodal neuroimaging data. AD research that uses deep learning is still evolving, improving performance by incorporating additional hybrid data types, such as—omics data, increasing transparency with explainable approaches that add knowledge of specific disease-related features and mechanisms.
Full-text available
Objective: The risk of atherosclerotic cardiovascular disease (ASCVD) is estimated using the American College of Cardiology (ACC)/American Heart Association (AHA) Pooled Cohort Equations (PCEs). However, the accuracy of this tool remains controversial, particularly among patients who are recommended statin therapy according to the ACC/AHA guidelines. We performed external validation of PCEs among patients eligible for statin therapy using data from the systolic blood pressure intervention trial (SPRINT). Results: Our study included 4057 patients from among the 9361 patients in SPRINT. The mean patient age was 64.5 years, and the median predicted 10-year risks of ASCVD were 17.2% and 12.3% for men and women, respectively. Over a median follow-up of 3.3 years, 133 primary events (including 23 cardiovascular deaths) were noted, whereas 304 events were predicted by the PCEs. The PCEs demonstrated poor calibration (Hosmer-Lemeshow test, p < 0.001) and overestimated the probability consistently. Additionally, they showed moderate discrimination [area under the curve: 0.65 (95% confidence interval, 0.60-0.69)]. This study demonstrates that PCEs might overestimate the risk of ASCVD in patients who are recommended statin therapy.
Full-text available
The cerebrospinal fluid (CSF) biochemical markers (biomarkers) Amyloidβ 42 (Aβ42), total Tau (T-tau) and Tau phosphorylated at threonine 181 (P-tau181) have proven diagnostic accuracy for mild cognitive impairment and dementia due to Alzheimer’s Disease (AD). In an effort to improve the accuracy of an AD diagnosis, it is important to be able to distinguish between AD and other types of dementia (non-AD). The concentration ratio of Aβ42 to Aβ40 (Aβ42/40 Ratio) has been suggested to be superior to the concentration of Aβ42 alone when identifying patients with AD. This article reviews the available evidence on the use of the CSF Aβ42/40 ratio in the diagnosis of AD. Based on the body of evidence presented herein, it is the conclusion of the current working group that the CSF Aβ42/40 ratio, rather than the absolute value of CSF Aβ42, should be used when analysing CSF AD biomarkers to improve the percentage of appropriately diagnosed patients.
Full-text available
The volume and complexity of diagnostic imaging is increasing at a pace faster than the availability of human expertise to interpret it. Artificial intelligence has shown great promise in classifying two-dimensional photographs of some common diseases and typically relies on databases of millions of annotated images. Until now, the challenge of reaching the performance of expert clinicians in a real-world clinical pathway with three-dimensional diagnostic scans has remained unsolved. Here, we apply a novel deep learning architecture to a clinically heterogeneous set of three-dimensional optical coherence tomography scans from patients referred to a major eye hospital. We demonstrate performance in making a referral recommendation that reaches or exceeds that of experts on a range of sight-threatening retinal diseases after training on only 14,884 scans. Moreover, we demonstrate that the tissue segmentations produced by our architecture act as a device-independent representation; referral accuracy is maintained when using tissue segmentations from a different type of device. Our work removes previous barriers to wider clinical use without prohibitive training data requirements across multiple pathologies in a real-world setting.
Importance Sight is often considered to be the sense most valued by the general public, but there are limited empirical data to support this. This study provides empirical evidence for frequent assertions made by practitioners, researchers, and funding agencies that sight is the most valued sense. Objective To determine which senses are rated most valuable by the general public and quantify attitudes toward sight and hearing loss in particular. Design, Setting, and Participants This cross-sectional web-based survey was conducted from March to April 2016 through a market research platform and captured a heterogeneous sample of 250 UK adults ages 22 to 80 years recruited in March 2016. The data were analyzed from October to December 2018. Main Outcomes and Measures Participants were first asked to rank the 5 traditional senses (sight, hearing, touch, smell, and taste) plus 3 other senses (balance, temperature, and pain) in order of most valuable (8) to least valuable (1). Next, the fear of losing sight and hearing was investigated using a time tradeoff exercise. Participants chose between 10 years without sight/hearing vs varying amounts of perfect health (from 0-10 years). Results Of 250 participants, 141 (56.4%) were women and the mean (SD) age was 49.5 (14.6) years. Two hundred twenty participants (88%) ranked sight as their most valuable sense (mean [SD] rating, 7.8 [0.9]; 95% CI, 7.6-7.9). Hearing was ranked second (mean [SD] rating, 6.2 [1.3]; 95% CI 6.1-6.4) and balance third (mean [SD] rating, 4.9 [1.7]; 95% CI, 4.7-5.1). All 3 were ranked above the traditional senses of touch, taste, and smell (F7 = 928.4; P < .001). The time tradeoff exercise indicated that, on average, participants preferred 4.6 years (95% CI, 4.2-5.0) of perfect health over 10 years without sight and 6.8 years (95% CI, 6.5-7.2) of perfect health over 10 years without hearing (mean difference between sight and hearing, 2.2 years; P < .001). Conclusions and Relevance In a cross-sectional survey of UK adults from the general public, sight was the most valued sense, followed by hearing. These results suggest that people would on average choose 4.6 years of perfect health over 10 years of life with complete sight loss, although how this generalizes to other parts of the world is unknown.
Rationale Patients with end-stage renal disease are characterized by increased cardiovascular and all-cause mortality because of advanced remodeling of the macrovascular and microvascular beds. Objective The aim of this study was to determine whether retinal microvascular function can predict all-cause and cardiovascular mortality in patients with end-stage renal disease. Methods and Results In the multicenter prospective observational ISAR study (Risk Stratification in End-Stage Renal Disease), data on dynamic retinal vessel analysis were available in a subcohort of 214 dialysis patients (mean age, 62.6±15.0; 32% women). Microvascular dysfunction was quantified by measuring maximum arteriolar dilation and maximum venular dilation (vMax) of retinal vessels in response to flicker light stimulation. During a mean follow-up of 44 months, 55 patients died, including 25 cardiovascular and 30 noncardiovascular fatal events. vMax emerged as a strong independent predictor for all-cause mortality. In the Kaplan-Meier analysis, individuals within the lowest tertile of vMax showed significantly shorter 3-year survival rates than those within the highest tertile (66.9±5.8% versus 92.4±3.3%). Univariate and multivariate hazard ratios for all-cause mortality per SD increase of vMax were 0.62 (0.47–0.82) and 0.65 (0.47–0.91), respectively. Maximum arteriolar dilation and vMax were able to significantly predict nonfatal and fatal cardiovascular events (hazard ratio, 0.74 [0.57–0.97] and 0.78 [0.61–0.99], respectively). Conclusions Our results provide the first evidence that impaired retinal venular dilation is a strong and independent predictor of all-cause mortality in hemodialyzed end-stage renal disease patients. Dynamic retinal vessel analysis provides added value for prediction of all-cause mortality and may be a novel diagnostic tool to optimize cardiovascular risk stratification in end-stage renal disease and other high-risk cardiovascular cohorts. Clinical Trial Registration URL: . Unique identifier: NCT01152892.
Purpose of review: To summarize the current findings on clinical retinal diseases and retinal imaging changes with dementia, focusing on Alzheimer's disease. Recent findings: Studies observed that clinical retinal diseases such as age-related macular degeneration, open-angle glaucoma and diabetic retinopathy are related to dementia, but the associations are not entirely consistent. In terms of the retinal neuronal structure, multiple retinal neuronal layers are significantly thinner in Alzheimer's disease dementia, such as the parapapillary retinal nerve fiber layer (RNFL) and macular ganglion cell-inner plexiform layer (GC-IPL). Recent studies further demonstrated that macular GC-IPL and macular RNFL are also significantly thinner in the preclinical stage of Alzheimer's disease. A thinner RNFL is also associated with a significantly increased risk of developing both cognitive decline and Alzheimer's disease dementia. In addition, studies consistently showed that retinal vascular changes are associated with poorer cognitive performance, as well as prevalent and incident Alzheimer's disease dementia. Summary: The current findings support the concept that changes in the retina, particular in retinal neuronal structure and vasculature, can reflect the status of cerebral neuronal structure and vasculature, highlighting the potential role of retinal changes as biomarkers of dementia.