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

Aims We examine whether inclusion of artificial intelligence (AI)-enabled retinal vasculometry (RV) improves existing risk algorithms for incident stroke, myocardial infarction (MI) and circulatory mortality. Methods AI-enabled retinal vessel image analysis processed images from 88 052 UK Biobank (UKB) participants (aged 40–69 years at image capture) and 7411 European Prospective Investigation into Cancer (EPIC)-Norfolk participants (aged 48–92). Retinal arteriolar and venular width, tortuosity and area were extracted. Prediction models were developed in UKB using multivariable Cox proportional hazards regression for circulatory mortality, incident stroke and MI, and externally validated in EPIC-Norfolk. Model performance was assessed using optimism adjusted calibration, C-statistics and R2 statistics. Performance of Framingham risk scores (FRS) for incident stroke and incident MI, with addition of RV to FRS, were compared with a simpler model based on RV, age, smoking status and medical history (antihypertensive/cholesterol lowering medication, diabetes, prevalent stroke/MI). Results UKB prognostic models were developed on 65 144 participants (mean age 56.8; median follow-up 7.7 years) and validated in 5862 EPIC-Norfolk participants (67.6, 9.1 years, respectively). Prediction models for circulatory mortality in men and women had optimism adjusted C-statistics and R2 statistics between 0.75–0.77 and 0.33–0.44, respectively. For incident stroke and MI, addition of RV to FRS did not improve model performance in either cohort. However, the simpler RV model performed equally or better than FRS. Conclusion RV offers an alternative predictive biomarker to traditional risk-scores for vascular health, without the need for blood sampling or blood pressure measurement. Further work is needed to examine RV in population screening to triage individuals at high-risk.

Model diagnostics (with 95% confidence intervals) from internal validation of circulatory mortality in UK Biobank (2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018). External validation in EPIC-Norfolk cohort using biomedical data from the third health check (2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)   UK Biobank eye examination occurred at baseline in a subset of participants 5 from December 2009 to July 2010 towards the latter end of recruitment in 6 UK Biobank centres. Participants attended for repeat assessment 1 to 5 years after recruitment and ocular assessments in this latter phase (August 2012-June 2013) were largely from individuals that had not undergone an ocular assessment on entry into UK Biobank. Both phases included visual acuity, autorefraction, intraocular pressure and corneal biomechanics. 5 Digital fundus photography and spectral domain OCT images were taken using the Topcon 3D-OCT 1000 Mark 2. Non-mydriatic 45° digital colour images, centred on the fovea were captured from 68,550 participants in the first phase and 19,502 from the second phase. Overlap with baseline ocular assessment was minimal. EPIC-Norfolk eye examination. Ophthalmic tests included measurement of vision, visual acuity (LogMAR acuity), and closed field auto-refraction (Humphrey model 500, Humphrey Instruments, San Leandro, California, USA), which was used to estimate axial length. Macular centred 45 digital fundus photographs were taken using a TRC-NW6S non-mydriatic retinal camera and IMAGEnet Telemedicine System (Topcon Corporation, Tokyo, Japan) with a 10 megapixel Nikon D80 camera (Nikon Corporation, Tokyo, Japan) without pharmacological dilation of the pupil.

Health outcomes
The primary outcome was circulatory mortality as defined using International Classification of Diseases

Development of circulatory mortality models in UK Biobank
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.
Throughout models were developed in UK Biobank for men and women separately, and externally validated in EPIC-Norfolk. We hypothesized that retinal vessel characteristics in relation to disease incidence, might be modified by age, smoking status, presence of CVD/diabetes and use of BP lowering BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance Supplemental material placed on this supplemental material which has been supplied by the author(s) medications. Hence, two-way interactions between retinal vasculometry and age, smoking status and selfreported use of blood pressure medication, prevalent diabetes and CVD were first examined in mutually adjusted Cox proportional hazard 7 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, R 2 statistic (representing a measure of explained variation) and calibration slope (where a slope of 1.0 is ideal). 8 The model from the bootstrapped sample was applied to the bootstrapped sample to estimate apparent performance and to the original dataset to test model performance. Optimism was estimated within each bootstrapped sample as the difference in performance parameters (C-statistic, R 2 and calibration slope) between model performance vs apparent performance. The overall (average) optimism across all bootstrapped samples was determined to adjust measures of model performance (C-statistic, R 2 and calibration slope).

External validation of circulatory mortality models in EPIC-Norfolk cohort
The original beta coefficients from the prognostic models were adjusted for shrinkage to allow for overfitting using the calibration slopes adjusted for optimism from the bootstrapped sampling. The adjusted linear predictor was then applied to the EPIC-Norfolk cohort and C-statistic, R 2 and calibration slope estimated. Calibration plots of the observed vs expected event probability by octiles of predicted risk of an event were calibrated to the average 5-year baseline survival in the EPIC-Norfolk cohort.

Framingham Risk Scores for stroke and MI in UK Biobank and EPIC-Norfolk cohorts
Framingham risk scores (FRS) for incident fatal and non-fatal stroke 9 and MI 10 were applied to UK Biobank and EPIC-Norfolk cohorts and recalibrated to baseline survival function within each cohort. 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. 10 Those reporting a history of heart attack or stroke or those with a date of event stroke or MI prior to retinal image capture were excluded from the corresponding prognostic modelling for that outcome. FRS models were also extended to include retinal vasculometry. Model development and validation followed a similar approach as described for circulatory mortality.
BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance Supplemental material placed on this supplemental material which has been supplied by the author(s)

Retinal vasculometry models for stroke and MI in UK Biobank and EPIC-Norfolk cohort
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 retinal vasculometry 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. Participants reporting diabetes or use of blood pressure lowering medications were included in stroke analyses. Participants with missing data on smoking status or self-report on medications for lowering blood pressure or lipids, or those that preferred not to report a history of heart attack or stroke were excluded from all FRS analyses (UK Biobank n= 1182 (1.8%); EPIC-Norfolk n=93 (1.6%)).
Prognostic models using retinal vasculometry included up to 26 candidate predictors in men and up to 28 in women, in the stepwise procedure based on inclusion of main effects and interactions with retinal vasculometry with p<0.2. A maximum of 16 predictors were identified by the stepwise procedure with p<0.1 in any single model. Retinal vasculometry measures excluded by the stepwise procedure were reinserted back into the model to check whether they became statistically significant. Fractional polynomial models were used to examine presence of non-linear associations but none were identified.

Sensitivity analyses
Sensitivity analyses restricted the entire model development and validation to the white ethnic group to check for systematic differences in model performance. With the EPIC-Norfolk cohort having a relatively smaller number of incident events, we assessed the external validation of models 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).   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.  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.