Artificial intelligence enabled retinal vasculometry for prediction of circulatory mortality, myocardial infarction and stroke

UK Biobank

is a prospective cohort study for which baseline biomedical and physical assessments were carried out 2006-2013, in 502,682 adults aged 40-69 years recruited from 22 UK centres.²⁴ Ocular assessments occurred during the latter phase (2009-2013; 7 centres) and included visual acuity, autorefraction, digital fundus photography with the Topcon 3D-OCT 1000 Mark 2.²¹ Non-mydriatic 45° digital colour images, centred on the fovea were available for 88,052 participants.

EPIC-Norfolk

was the UK component of the European Prospective Investigation into Cancer (EPIC) study.^25,26 Here we focus on data from the 3^rd clinical follow-up (2004-2011)²⁷ on 8,603 participants aged 48-92 years who underwent a biomedical and eye examination similar to that of UK Biobank (see Supplementary Material for further details).¹⁶

Health outcomes

The primary outcome was circulatory mortality as defined using International Classification of Diseases (ICD) (ICD-10 codes I00-I99 and ICD9 390-459) coded death registry data from the Office for National Statistics and the Health and Social Care Information Centre (now NHS Digital) for England and Wales, and the Information Services Department for Scotland, provided information on date and cause(s) of death to 31^st January 2018 for UK Biobank and 31^st March 2018 for EPIC-Norfolk. Incident MI and stroke events after retinal image capture were based on medical records linkage with hospital diagnoses of non-fatal events, supplemented with participant health and lifestyle questionnaire data from repeat surveys in UK Biobank and EPIC-Norfolk (2012-2018). ICD-10 codes I21-I25 (or ICD-9 codes 410, 411,412 429.79) were used for fatal and non-fatal MI; and ICD-10 codes I60,61,63,64 (or ICD-9 codes 430, 431,434,436) for ischaemic and haemorrhagic stroke (see Algorithmically - define health outcomes https://www.ukbiobank.ac.uk/enable-your-research/about-our-data/health-related-outcomes-data).

AI-enabled retinal image processing

A validated, fully automated AI-enabled system (QUARTZ)^18-20 extracted thousands of measures of retinal vessel width, tortuosity and area from the whole retinal image. Supervised machine learning techniques were used within QUARTZ; with a support vector machine used to create an image quality score¹⁸ and deep learning was used to develop an algorithm to distinguish between arterioles and venules.¹⁹ QUARTZ measures of width (microns²⁸), total vessel area (mm²), tortuosity (arbitrary units),^16,29 and variance of widths along a vessel segment, were averaged for each image (weighted by the length of each vessel segment), separately for arterioles and venules. Person level averages were obtained by averaging across right and left eyes.

Statistical Analysis

Statistical analyses were carried out using STATA software (version 16, StataCorp LP, College Station, TX). Retinal vessel widths and area showed normal distributions, tortuosity required log-transformation and within-vessel-width-variance required inverse square-root transformation to normalize distributions. Models were developed in UK Biobank for men and women separately throughout, and externally validated in EPIC-Norfolk. We hypothesized that retinal vessel characteristics in relation to circulatory mortality might be modified by age, smoking status, presence of CVD/diabetes and use of BP lowering medications. Hence, two-way interactions between RV and age, smoking status and self-reported use of blood pressure medication, prevalent diabetes and CVD were first examined in mutually adjusted Cox proportional hazard³⁰ models for circulatory mortality. Interaction terms with p values <0.2 were then included along with main effects in Cox regressions models using backward elimination (p value set to 0.1). Bootstrapping with 100 replications was used for internal validation to adjust model performance measures for optimism, including Harrel’s C-statistic for discrimination, R² statistic (representing a measure of explained variation)³¹ and calibration slope (where a slope of 1.0 is ideal).³² The original beta coefficients adjusted for shrinkage were applied to EPIC-Norfolk to estimate C-statistic, R² and calibration slopes. Model performance was graphically assessed from plots of the observed probability of event at 5 years by deciles of predicted risk at 5 years in UK Biobank and by octiles in EPIC-Norfolk.

Framingham risk scores (FRS) for incident fatal and non-fatal stroke²² and MI²³ were applied to UK Biobank and EPIC-Norfolk cohorts and recalibrated to baseline survival function within each cohort according to the 5 year survival rates. Following FRS criteria, participants reporting use of cholesterol lowering medications, diabetes or missing data on total or HDL cholesterol were excluded from all MI analyses.²³ FRS models were also extended to include RV. Alternative models for incident fatal and non-fatal stroke and MI using age, smoking status, medical history (self-reported history of heart attack, stroke or diabetes and use of blood pressure lowering medications) and RV only were developed in UK Biobank following the same approach as for circulatory mortality. A medical history of MI did not preclude inclusion in models for incident stroke events, and vice-versa.

Sensitivity analyses restricted model development and validation to white ethnicity. Using EPIC-Norfolk, external validation was extended to a broader spectrum of incident cerebrovascular disease (ICD10 I60-69; ICD 9 430-438) and incident ischaemic heart disease (ICD10 I20-I25; ICD9 410-414). We follow TRIPOD guidelines for reporting of model development and validation.³³

Circulatory mortality

64,144 UK Biobank participants with 327 circulatory deaths and 5862 EPIC-Norfolk participants with 201 circulatory deaths were included. In men, arteriolar and venular width, tortuosity and width-variance were identified as statistically significant predictors of circulatory mortality. In women, arteriolar and venular area and width, venular tortuosity and venular width-variation contributed to risk prediction. RV effects on circulatory mortality were modified by smoking status, blood pressure medications and history of MI. In men and women, optimism adjusted C- (0.75 to 0.77) and R² (0.33 to 0.44) statistics in UK Biobank and EPIC-Norfolk, were reasonably high (Table 2, Table S1 for full model diagnostics; Tables S2 and S3 for regression coefficients). In UK Biobank men, predicted risks were closely aligned with observed risks. A similar picture emerged for EPIC-Norfolk men cohort, with about double the risk of circulatory mortality, half the numbers of events and being about a decade older at retinal image capture had (Figure 2). UK Biobank women showed a wide separation of risk groups and close alignment of predicted and observed risks even, at low risks (<0.5%). Calibration plots for EPIC-Norfolk women were less clear due to the lower number of events available, hence 95% CI around predictions were wider (Figure 2).

• Download figure
• Open in new tab

Figure 2:

Observed risk of outcome at 5 years by deciles of predicted risk in UK Biobank and eighths of predicted risk in EPIC-Norfolk. Prediction models based on age, smoking, medical history with retinal vasculometry only.

Figure 2 footnote:

Top row: Circulatory mortality

Middle row: Incident stroke

Bottom row Incident MI

Vertical lines around symbols are 95% confidence intervals. Dotted line represents perfect calibration.

The scale of the vertical and horizontal axes is a probability e.g., 0.1 equates to a 10% risk of event by 5 years. Circulatory mortality codes: ICD10: I00-I69, ICD9: 390-359

Incident stroke codes: ICD10: I60,I61,I63,I64, ICD9: 430,431,434,436

Incident MI codes: ICD10: I21-I25, ICD9: 410,411,412,429.79

Incident stroke

63,839 UK Biobank participants with 446 incident strokes and 5,708 EPIC-Norfolk participants with 211 incident stroke events after retinal image capture were included (Figure S1). In UK-Biobank, FRS C-statistic was 0.74 in men and 0.74 in women (Table 2) with lower values in EPIC-Norfolk; approximately one third of the variation in stroke-risk incidence was explained by R² (less so in EPIC-Norfolk men). Observed risks were more aligned with predicted risks in men than in women (Figure S2). Addition of RV to FRS did not improve model performance statistics overall (Table S4; Figure S2).

Models based on age, smoking status, medical history and RV showed similar performance to FRS with C-statistic of 0.73 in men and 0.75 in women and marginally improved R² values in UK Biobank; (Table 2; full model diagnostics Tables S4). As for FRS, performance metrics were lower in EPIC-Norfolk. Multivariable models (Tables S2-S3) showed venular and arteriolar tortuosity and width were predictors of stroke in men and women and additionally venular/arteriolar area in women with some modification by smoking status, blood pressure medications and history of MI. Calibration plots showed risk predictions closer to the 45-degree line particularly at lower levels of predicted risk in women (Figure S2).

Incident MI

45,734 UK Biobank participants with 393 incident MI and 4,062 in EPIC-Norfolk with 265 incident MI after retinal image capture were included (Figure S1). In UK Biobank, FRS C-statistics were 0.71 in men and 0.76 in women with approximately one quarter (24%) of the variation in MI risk explained by FRS in men and 35% in women (Table 2). In EPIC-Norfolk, with approximately 5x the risk of MI, performance statistics were lower. Calibration plots for FRS showed better alignment of observed and predicted risks in men compared with women (Figure S3). Addition of RV to FRS did not improve model performance overall (Table S5; Figure S3). Compared to FRS alone a simpler model based on age, smoking status, medical history and RV performed marginally less well in men and women in both cohorts. (Tables 2 and S5; Figure S3). Multivariable models for MI using RV (Table S2-S3) showed arteriolar and venular width, venular width variability and arteriolar area were predictors in men, whereas for women venular tortuosity, venular/arteriolar area and venular width variability were predictors. RV effects were modified by smoking status.

Cases in top quintile of risk scores

For circulatory mortality models based on age, smoking status medical history and RV captured between 52-65% of cases of circulatory mortality in the top quintile of the risk score distribution (Table 3). For incident stroke, RV based models compared with FRS captured about 5% more cases in UK Biobank men and 8% more cases in UK Biobank women and 3% more EPIC-Norfolk men in the top quintile of risk scores (Table 3) but 1.8% fewer EPIC-Norfolk women. However, for MI, FRS captured more cases of MI in the top quintile of risk. Considering stroke and MI scores combined, the simpler RV models captured more cases in the top quintile than FRS for UK Biobank men and women, and similar proportions in EPIC-Norfolk men and women.

Sensitivity analyses

Restricting model development and validation to those of white ethnicity did not materially alter model performance for any of the models presented. FRS and all RV models for stroke showed systematically improved external validation for outcomes based on inclusion of all incident cerebrovascular disease in EPIC-Norfolk (Table S4 far right-hand column and Figure S4), especially in women. In contrast, for all incident ischaemic heart disease in EPIC-Norfolk, performance of FRS and RV models remained remarkably unchanged in men but marginally improved in women (Table S5 far right-hand column and Figure S5).

Comparisons with other studies

Prospective associations have been largely based on retinal vessel width with mortality,¹³ incident stroke,^{13 13} and with CHD (in women, not men),^11,34 from restricted measurement areas of the retina.^11,35 Measurements are often not automated, requiring operator involvement, which limits application to large populations. In agreement with others, our models show that both arteriolar and venular vasculometry contribute to risk prediction,^11,36 and this aligns with our previous work.^16,29,37 Seidelmann reported that narrower central retinal artery and wider central retinal vein equivalent dimensions offered significant additional information to equations for incident atherosclerotic CVD risk,¹¹ especially in women, but C-statistics were modest (between 0.55-0.57) compared with the much higher levels in the current study (i.e., between 0.70-0.77). Our RV models generally performed better in women and may indicate that microvascular dysfunction contributes more to CHD pathogenesis in women than in men, as they have smaller coronary arteries exhibiting more diffuse ‘non-obstructive’ atherosclerosis,³⁸ with a larger burden of coronary microvascular disease,³⁹ leading to higher morbidity and mortality.⁴⁰ A recent study using the UK Biobank data source in fewer participants (54,813 vs 65,144 in this study), showed that retinal vessel density and fractal dimensions (extracted from the entire image after deep learning vessel segmentation without distinction between arterioles and venules) were associated with other health outcomes, including overall mortality, hypertension and congestive heart failure, but did not report on risk predictionperformance.⁴¹ Moreover, there was no consistent evidence of associations with incident circulatory disease, and cerebrovascular disease and associations with incident MI were null.⁴¹ Another study in a sub-set of UK Biobank participants (n=5,663) with both retinal and cardiovascular magnetic resonance imaging used deep learning /AI approaches to estimate structural cardiac indices as intermediaries for predicting MI.⁴² However, given their approach, specific retinal features of importance remain unclear.

European Systematic Coronary Risk Evaluation (SCORE) CVD risk score⁸, QRISK3 risk score⁹ and the American College of Cardiology/American Heart Association CVD risk score algorithms have already been evaluated in UK Biobank. The published C-statistics for these three risk scores were 0.77, 0.74 and 0.74 respectively, with 95%CI that overlap with values for the simpler RV model presented in this study. However, the novel C-statistics for circulatory mortality reported in this study are higher. Our approach of focussing on the retinal microvasculature as a key prognostic marker of incident cardiovascular outcomes and circulatory mortality is supported by saliency maps presented in a study using end-to-end AI of retinal images to estimate the extent of coronary artery calcium scores in cross-sectional associations,⁴³ with C-statistics for incident CVD varying between 0.68 to 0.76. Our model using RV together with easily attainable data including age, smoking status, sex and a brief medical history, is simple, non-invasive and exhibits performance that is comparable, or even better than, current risk algorithms, including end-to-end AI approaches.

Strengths and limitations

Model development in UK Biobank provided a large sample size and number of prospective events. QUARTZ successfully processed a high percentage (∼80%) of retinal images captured by non-experts providing ‘vasculomic’ indices of vascular health. External validation in an older higher risk cohort (EPIC-Norfolk) replicated the findings, and models were also robust to inclusion of a wider spectrum of cerebrovascular and ischaemic heart disease events.

UK Biobank and EPIC-Norfolk are ‘healthy’ cohorts with relatively low event rates compared with other geographically similar middle-aged cohorts.⁴⁴ Prevalence of current smoking was very low in UK Biobank (6%) and limited the ability to examine interactions with RV. Although we did not find limiting the analysis to those of white ethnicity materially altered the results, the proportion of non-white participants in UK Biobank is low. RV may relate to microvascular endothelial function elsewhere in the body and may underpin the causal pathways behind prognostic models, which may differ with ethnicity. Confirmation of model performance in other cohorts with higher CVD rates and in different (especially non-white) ethnic groups would be informative.

Implications and Conclusions

Retinal imaging is established within clinic and hospital eye care and in optometric practices in the US and UK. AI-enabled vasculometry risk prediction is fully automated, low cost, non-invasive and has the potential for reaching a higher proportion of the population in the community because of ‘high street’ availability and because blood sampling or sphygmomanometry are not needed. RV is a microvascular marker, hence offers better prediction for circulatory mortality and stroke compared with MI which is more macrovascular, expect perhaps in women. In the general population it could be used as a non-contact form of systemic vascular health check, to triage those at medium-high risk of circulatory mortality for further clinical risk assessment and appropriate intervention. Having a further low cost, accessible, non-invasive screening test in the community to encourage clinical risk assessment uptake in the community (in addition to current screening approaches), would help to prolong disease-free status in an ever-ageing population with increasing co-morbidities, and assist with minimising healthcare costs associated with lifelong vascular diseases.