Introduction

Atrial fibrillation (AF) is already one of the commonest cardiac conditions and expected to further double in prevalence over the next two decades [1]. It is a leading contributor to several complications including ischemic stroke, myocardial infarction, vascular dementia and mortality [2]. Diagnosis of AF relies on capturing what is often a sporadic heart rhythm abnormality on an electrocardiogram (ECG). The likelihood of AF detection depends on aligning symptoms with ECG-based monitoring, but AF episodes themselves do not directly relate to thromboembolic outcomes [2]. Rather, AF is a thrombo-inflammatory condition closely associated with underlying vascular disease. Appropriate use of established therapy, including direct oral anticoagulants and modulators of vascular disease, have the potential for substantial public health impact, if earlier diagnosis of AF can be achieved across broader communities [1].

The retina with its vasculature and multiple layers provides a non-invasive method to assess microvascular and central nervous system alterations [3]. Novel insights can be provided by analysing widely-available optical coherence tomography (OCT) scans and colour fundus photographs (CFP) using artificial intelligence (AI) assisted segmentation, classification and predictive modelling [3,4]. Retinal vasculature parameters and retinal sublayer thickness are associated with cardiovascular disease (CVD) and specifically with AF [5,6], but prior studies have been limited by size, scope and imaging modality. We hypothesised that retinovasculature and neuroretinal indices could be a biomarker of AF, providing an easily-accessible window for detection of vascular disease status in individual patients. In this study, we analysed two large independent cohorts, drawing on data from more than 90,000 individuals to understand the links between retinal morphology and the prevalence and incidence of AF (Fig 1).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic detailing study design, retinal imaging modalities, and results summary.

AF: atrial fibrillation, FD: fractal dimension, GCIPL: ganglion cell-inner plexiform layer, INL: inner nuclear layer, UKB: UK Biobank.

https://doi.org/10.1371/journal.pdig.0001661.g001

Results

This analysis drew on data from 90,190 individuals across the two independent cohorts. The AlzEye analysis consisted of 50,651 individuals, of which 5,735 had AF (Fig 2, prevalence: 11.3%). Individuals with AF were, on average, older, more likely to be male, have hypertension and diabetes mellitus. They were also significantly more likely to have suffered a previous MACE (Table 1, 40.6% versus 10.0%, p < 0.001). The UKB analysis included 39,539 participants, of which 526 had prevalent AF at baseline (1.3%) (Fig 2). Similar to AlzEye, individuals in UKB with AF were generally older, more likely to be male, have hypertension and diabetes mellitus (Table 2). Within UKB, 10.5% of participants with AF had a previous MACE compared to 1.5% without AF (p < 0.001).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Baseline characteristics of the AlzEye cohort stratified by presence of previous major adverse cardiovascular event.

https://doi.org/10.1371/journal.pdig.0001661.t001

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Baseline characteristics of the UK Biobank.

https://doi.org/10.1371/journal.pdig.0001661.t002

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Flow chart for the AlzEye and UK Biobank cohort analyses.

MACE: Major adverse cardiovascular event.

https://doi.org/10.1371/journal.pdig.0001661.g002

Prevalent atrial fibrillation

In AlzEye, after adjustment for age, sex, ethnicity, hypertension and diabetes mellitus, individuals with prevalent AF had significantly thinner retinal venular calibre compared to individuals without AF (-0.05, 95% CI: -0.08, -0.03, p = 9.1 x 10^-5). Retinal venular fractal dimension was significantly lower in those with AF (-0.08, 95% CI: -0.11, -0.06, p = 1.1 x 10^-10), whereas arteriolar fractal dimension was higher (0.05, 95% CI: 0.02, 0.08, p = 1.7 x 10^-4). Significantly thinner retinal sublayers in the mRNFL (-0.69 microns, 95% CI: -0.91, -0.46, p = 1.9 x 10^-9), mGCIPL (-1.96 microns, 95% CI: -2.34, -1.58, p < 2.0 x 10^-16), mINL (-0.57 microns, 95% CI: -0.75, -0.38, p = 1.4 x 10^-9) and PRL (-0.93 microns, 95% CI: -1.38, -0.48, p = 5,1x10^-5) were seen among those with prevalent AF. Given the high prevalence of MACE in individuals with AF, we restricted the analysis to those with no previous MACE. Similar changes were seen in venular calibre and venular fractal dimension, mRNFL, mGCIPL and INL.

Replicating the analysis in the UKB cohort of 526 individuals with prevalent AF using the same variable adjustment (S1 Table) and additionally adjusting for alcohol consumption and smoking status, effect estimates were of the same direction as AlzEye, with smaller magnitudes of difference (Table 3). In UKB, only thinner mGCIPL was significantly different in those with AF (-1.19 microns, 95% CI: -1.73, -0.65, p = 1.5 x 10^-5, Fig 3). Sensitivity analysis excluding those with previous MACE in UKB did not result in any significant change in effect estimates. Individuals with prevalent AF had significantly thinner mGCIPL (-1.15 microns Table 4, 95% CI: -1.72, -0.58, n prevalent AF = 473, S2 Table). Similarly, excluding those with self-reported eye diseases in UKB did not alter the direction or magnitude of significant effect estimates (S3 Table).

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Association between prevalent atrial fibrillation and retinal morphology. Estimates are derived from multivariable mixed effects linear regression adjusting for age, sex, ethnicity, hypertension and diabetes mellitus; in the UK Biobank analysis, models were also adjusted for alcohol consumption and smoking status.

https://doi.org/10.1371/journal.pdig.0001661.t003

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Association between incident atrial fibrillation and retinal morphology. Estimates are derived from frailty models. Model 1 was unadjusted; Model 2 included adjustment for age, sex and ethnicity. Model 3 additionally included adjustment for hypertension, diabetes mellitus, alcohol consumption, and smoking status.

https://doi.org/10.1371/journal.pdig.0001661.t004

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Summary of regression coefficients for prevalent and incident atrial fibrillation (AF) in both the AlzEye and UK Biobank cohorts.

For prevalent AF, coefficients are derived from adjusted mixed effects linear regression. For incident atrial fibrillation, hazard ratios were derived from adjusted frailty models per standard deviation decrease. AF: atrial fibrillation, CI: confidence interval, HR: hazard ratio, mGC-IPL: macular ganglion cell-inner plexiform layer, mINL: macular inner nuclear layer, mRNFL: macular retinal nerve fibre layer, OCT: optical coherence tomography, PRL: photoreceptor layer, RPE-BM: retinal pigment epithelium-Bruch’s membrane layer.

https://doi.org/10.1371/journal.pdig.0001661.g003

Discussion

Drawing on data across two large distinct cohorts, we observed differences in retinal sublayer thickness and retinal vasculature in individuals with AF. Prevalent AF was consistently associated with a thinner mGCIPL. In the hospital-attending cohort of AlzEye, even after restricting to those without concurrent MACE, we found several differences in retinovascular indices as well as mINL thickness. Differences in the mGCIPL thickness were also evident in individuals with incident AF from the UKB longitudinal cohort. Our report, which represents to our knowledge by some magnitude the largest reported analysis of retinal morphology in people with AF, underlines the mechanistic associations of AF and structural retinal changes and, once validated by further studies could support potential prognostic utility of retinal imaging in cardiac arrhythmias.

Across both our datasets we observed a thinner mGCIPL in patients with AF. This is the final output layer of the vertebrate retina, and we demonstrate an effect size that is comparable to six years older age in the UKB analysis [7]. We hypothesize that this difference is driven predominantly through retrograde transsynaptic degeneration in those with AF, which mechanistically accounts for the mGCIPL thinning seen in individuals with multiple sclerosis and, more pertinently, following ischaemic stroke [8,9]. Patients with AF have an increased risk of silent cerebral infarction, with one European prospective cohort study (Swiss-AF) [10] of 1227 AF patients demonstrating new cerebral infarctions in 5.5% of patients within 2 years. Of note, the majority were clinically silent, which may rationalize our persistent finding of thinner mGCIPL even when excluding those with documented previous myocardial and cerebrovascular events. While the mGCIPL changes could potentially be explained with appropriate neuroimaging, we could not include such analysis in our study. Swept-source OCT and OCT angiography have been used to characterize the superficial capillary plexus of the retinal macular area in AF, showing reduced retinal capillary densities and perfusion as well as thinner mGCIPL and RNFL thickness when compared to age- and sex matched controls [6,10]. Furthermore, they also show a dependence of the retinal vascular parameters with echocardiographic measurements of the left atrium. The mGCIPL changes as shown in both of our cohorts are in line with a recently published study exploring the influence of AF on the progression of glaucoma, a neurodegenerative disease [5]. The particular sensitivity of the mGCIPL to systemic pathophysiological processes is further supported by an analysis of over 42,000 UK Biobank participants that GCIPL explains approximately double the variance attributable to systemic factors, thus being the superior inner retinal biomarker [7]. To contextualise these effect sizes, the -1.96 µm mGCIPL thinning observed in the AlzEye cohort represents 37.8% of the mGCIPL standard deviation and is equivalent to approximately 13 years of age-related thinning, and approximately 4 times the mGCIPL thinning associated with daily alcohol consumption. The -1.19 µm difference in the UKB cohort similarly represents 22.9% of the standard deviation and is equivalent to approximately 8 years of aging. The combination of retrograde transsynaptic neurodegeneration from AF-associated cerebrovascular injury and direct ischaemic susceptibility of the high-metabolic-demand ganglion cell layer provides a dual mechanistic basis for the mGCIPL as a hypothetically plausible and reproducible retinal correlate of AF. Given the associations between thinner mGCIPL and AF, future research could validate our results by linking the retinal anatomic changes with neuroimaging to establish this relationship between silent cerebral infarcts, AF and retinal sublayer differences.

In the AlzEye dataset we found significantly thinner mINL in patients with AF, even those without MACE. The INL, a retinal sublayer containing the cell bodies of horizontal cells, bipolar cells, amacrine cells and interplexiform neurons is typically preserved in individuals with neurological diseases, such as Alzheimer disease and multiple sclerosis, a finding which has been attributed to the protective synaptic barrier for retrograde transsynaptic degeneration [11,12]. Our observation may instead be explained through a microvascular pathogenesis within the inner retina. The INL receives its blood supply through the deep capillary plexus of the macular vasculature, a watershed zone susceptible to subclinical ischemia in the form of retinal ischaemic perivascular lesions. These focal thinnings of the mINL have been increasingly reported with observations in participants with coexisting cardiovascular disease [13], single subcortical infarcts [14], carotid artery disease [14,15], and AF [16]. Novel high resolution OCTs in individuals affected by AF would improve visualisation of the INL morphology and characterize microvascular dysfunction of the inner retina.

The analysis of retinovascular fractal dimensions and vascular density reveal significant alterations in patients with conditions such as minor stroke, congestive heart failure, certain retinal vascular diseases, type 2 diabetes, sleep apnea, and macrovascular cardiac disease, positioning these metrics as potential biomarkers for population screening [17–20]. In our study, the AlzEye cohort demonstrated lower venular fractal dimension and, to a lesser extent, greater arteriolar fractal dimension in individuals with prevalent AF, even after MACE adjustment. Furthermore, lower arteriolar and venular fractal dimension were observed in the AlzEye group with incident AF. However, similar to the previously noted INL thinning, these distinct findings could not be replicated in the UKB cohort. This discrepancy may stem from differences in sample size, inherent population variance, and cohort characteristics, as UKB participants are generally healthier than the UK population [21], whereas the AlzEye dataset comprises hospital patients with a higher burden of systemic and ocular diseases [22]. Consequently, further investigations leveraging longitudinal data, ideally capturing progression from paroxysmal to persistent AF, are essential to clarify these associations.

While the observed retinovascular alterations and retinal sublayer thickness changes are consistent with our pathomechanistic hypothesis of a clinical association with AF, their context is paramount. The diagnosis of AF currently relies on cardiac monitoring, clinical presentation, and cardiovascular risk profiling. The limitations of ECG monitoring over a prolonged period is complicated by the fact that photoplethysmography-based monitoring is constrained by evidence quality and the requirement for physician ECG confirmation before diagnosis [1]. Retinal imaging in comparison, captures structural end-organ sequelae of AF that persist independently of arrhythmia burden at any single timepoint. We propose that these retinal findings are not a substitute for this paradigm, but rather a complementary tool. From an oculomic perspective, they highlight the complex systemic nature of AF and, if confirmed in subsequent studies assessing their predictive potential, could guide clinicians to pursue more extensive cardiac investigations, such as implantable loop recorders, in clinically equivocal cases.

Notwithstanding strengths of our report including the large sample size across two cohorts, multimodal retinal imaging, and ethnic diversity of the AlzEye cohort, there are limitations to highlight. We relied on self-report and/or hospital admissions data for identifying individuals with AF, which may be prone to misclassification bias. A 2024 manual validation study however showed high PPV for an ICD-10 code for AF (PPV 99.1%) but more modest sensitivity (77.5%) [23], echoing findings from a 2013 US-based systematic review across 16 studies (median PPV, 89%, sensitivity 79%) [24]. Because paroxysmal and asymptomatic AF will likely not be included in hospital coding or self-reports, some individuals classified as controls, and in the incident analysis, may in fact have had undiagnosed AF at the time of retinal imaging. Thus, while we can be confident of our cases, controls may be affected by AF (even more so given the diagnostic challenge in identifying asymptomatic AF) suggesting our effect measures to underestimate the true association.Thus, while we can be confident of our cases, controls may be affected by AF (even more so given the diagnostic challenge in identifying asymptomatic AF) suggesting our effect measures to underestimate the true association. While we sought to replicate our findings across two large but disparate cohorts, neither is truly representative of the broader population. Participants of UKB engage in fewer adverse health behaviour and have greater life expectancy than the general UK population [21]. In contrast, the hospital cohort of AlzEye represents one of the most ethnically diverse cohorts in the Western Hemisphere and is skewed towards those with greater levels of socioeconomic deprivation [22]. Replication however of our findings across both cohorts bolsters the argument of a true association. In the incident AF analysis, approximately 20% of individuals developed AF within 24 months of retinal imaging. Given that AF is typically paroxysmal in its early stages, our findings may be more applicable to those with prevalent rather than incident AF. Finally, as with other observational health data analyses, there may be factors which we could not account for but may plausibly influence both AF risk and retinal structure (residual confounding).

This study demonstrates that individuals with prevalent and incident AF exhibit thinner mGCIPL and mINL, alongside differences in vascular indices, with a similar but less consistent pattern observed for incident AF. The observed mGCIPL thinning may result from microvascular events associated with AF that affect the visual pathways and retina. In contrast, the novel finding of mINL thinning could be interpreted in the context of underlying subclinical ischemia reflecting systemic cardiovascular compromise. These findings, once validated in subsequent studies, highlight the potential for non-invasive retinal imaging to provide early detection and risk stratification of AF, as well as an easily-accessible opportunity to understand AF pathophysiology and directly assess end-organ damage. Given the widespread availability of retinal imaging in primary care optometrist stores across developed nations, further attention is warranted to explore the role of retina imaging for earlier diagnosis of AF, and the opportunity to target therapy addressing underlying vascular disease to reduce the large patient and healthcare burdens associated with AF.

Materials and methods

Supporting information

Acknowledgments

The authors thank Polly Rawlinson for project management, Curtiss Green, Jonathan McKee, and Muqbool Khan for information governance expertise and Alice Selway-Clarke for information technology support. We also acknowledge the support of Tony Ko and Reza Jafari from Topcon Healthcare for support with the use of the TABS software. This research has been conducted using the UK Biobank Resource under Application Number 2112.