Retinal age gap as a predictive biomarker of stroke risk

The UK biobank is a large-scale population-based prospective cohort study of 502, 656 UK residents aged 40 to 69 years who were registered with the National Health Service (NHS). The overall study protocols and data are available elsewhere [24]. Briefly, baseline assessments were performed between 2006 and 2010 in 22 assessment centres across the UK. Participants completed electronic questionnaires to provide information on socio-demographics, lifestyle, environmental exposures, medical history and cognitive functions. Physical examinations including blood pressure, heart rate, grip strength, anthropometrics and spirometry were done for all participants. Biological samples including stored blood, urine and saliva samples were collected. Follow-up of medical conditions was performed mainly through data linkages to hospital records and mortality registries.

This study was reviewed and approved by the National Information Governance Board for Health and Social Care and the NHS North West Multicenter Research Ethics Committee (11/NW/0382) and the Biobank consortium (application no. 62489). Since we used de-identified data in a public dataset, the Medical Research Ethics Committee of Guangdong Provincial People’s Hospital waived the requirements to obtain the ethical approval. The study was performed in accordance with the Declaration of Helsinki. All participants provided informed consent.

Between 2009 and 2010, ophthalmic examinations were introduced at six assessment centres across the UK [25]. The 45° non-mydriatic retinal fundus and optical coherence tomography (OCT) imaging of the optic disc and macular were captured using a spectral domain OCT for each eye (Topcon 3D OCT 1000 Mk2, Topcon Corp, Tokyo, Japan). At baseline, ophthalmic examinations were completed in 66,500 participants, resulting in a total of 131,238 fundus images.

A total of 80,169 images from 46,969 participants passed the image quality check and were included in the analysis. The characteristics of the participants stratified by the number of images passing the quality check were described in detail in Additional file 1: Table S1. Among 46,969 participants, 11,052 participants did not report any previous disease at baseline. The DL model for age prediction was constructed based on fundus images of disease-free participants. To maximize the data available, binocular images, if available, were used for training and validation. The association between the retinal age gap and stroke was investigated using images from the remaining 35,304 participants who had no history of stroke at baseline. Images from the right eye were included in the test set to predict retinal age and images from the left eye were used only if images from the right eye were not available.

The methods of retinal age prediction using DL models were described in detail in a previous study [26]. Our previous study has assessed the performance of the DL model for age prediction. The DL model accurately predicted retinal age, as reflected by a strong correlation of 0.80 (P<0.001) between predicted retinal age and chronological age, as well as an overall mean absolute error (MAE) of 3.55 years. Attention maps retrieved from the DL model for age prediction mainly highlighted areas around the retinal vessels in the fundus images.

The retinal age gap was defined as the difference between the retinal age predicted by the DL model based on fundus images and the chronological age. A positive retinal age gap indicated an ‘older’-appearing retina, while a negative one indicated a ‘younger’-appearing retina.

Stroke was ascertained by the UK Biobank Outcome Adjudication Group, and was defined by codes 430.X, 431.X, 433.X1, 434.X1, 436.X in the 9th edition of the International Classification of Diseases (ICD-9) and ICD-10 codes I60, I61, I63, and I64. Stroke events were derived from linked electronic health records, including hospital records on admissions and diagnoses from hospitals in England, Scotland and Wales, as well as cause of death obtained from national death registers. The date of the first known stroke after baseline assessment was recorded. The follow-up period for each participant was defined from the recruitment date of the UK Biobank study to 28th February 2018 (the last follow-up date), or to the date of the first known stroke, whichever came first.

Covariates in the present analyses included socio-demographic factors (baseline age, sex, ethnicity, Townsend deprivation indices [TDI], education attainment), lifestyle factors (smoking status, drinking status, physical activity level), and general health status. Baseline age and sex were obtained from central registry or self-reported questionnaires. Self-reported ethnicity was divided into white or non-white. TDI was a proxy measure of socioeconomic status based on the postcode. Education attainment was classified into college/university degree or above, or others. Smoking and drinking status were categorized as current/previous users, or never. Physical activity level was categorized into reaching moderate/vigorous/walking recommendation or not. General health status was classified as excellent/good or fair/poor. Body mass index (BMI) was calculated as the weight of an individual in kilograms divided by their height in meters squared. Obesity was defined as a BMI of 30 kg/m² or above. Diabetes mellitus was defined as any record of self-reported or doctor-diagnosed diabetes mellitus, or the use of anti-hyperglycaemic medications or insulin. Hypertension was defined as self-reported, or doctor-diagnosed hypertension, or the use of antihypertensive drugs, or an average systolic blood pressure ≥ 130 mmHg or an average diastolic blood pressure ≥ 80 mmHg.

Continuous variables were reported as means and standard deviations (SDs), while categorical variable was reported as numbers and percentages. Unpaired t-tests and Chi-square tests were performed to examine the differences of the continuous and categorical variables, respectively. The log-rank test was used for comparing the survival distributions among different retinal age gap groups. Cox proportional hazards regression models were used to estimate the effect of retinal age gap on the risk of stroke. Each variable was assessed for the proportional hazards assumption and all of them met the assumption. Retinal age gap was introduced into the models as a continuous variable (per one-year increase) and a categorical variable (quintiles), respectively. Model I adjusted for baseline age, sex, and ethnicity. Model II adjusted for all variables in model I, and also TDI, educational level, smoking status, drinking status, physical activity level, diabetes mellitus, hypertension, obesity and general health status. Logistic regression models were used to estimate the predictive value of the well-established conventional risk factor-based model (including age, gender, smoking status, history of diabetes, systolic blood pressure, and total cholesterol to HDL-cholesterol ratio) [27] and the retinal age-based model in 10-year stroke risk. Area under the receiver-operating-characteristic curve (AUC) was used to describe the discrimination of the models in predicting 10-year stroke risk.

Sensitivity analysis was performed to adjust for the age-squared term in the final models in addition to age. We also investigated whether retinal age acceleration residual (defined as the residuals from regressing predicted retinal age against chronological age) was a biomarker of stroke in the second sensitivity analysis.

A two-sided p value of < 0·05 indicated statistical significance. All analyses were performed using R (version 3.3.0, R Foundation for Statistical Computing, www.​R-project.​org, Vienna, Austria) and Stata (version 13, StataCorp, TX, USA).

A total of 35,304 participants without any stroke history at baseline were included in the analyses (mean age 56.7 ± 8.04 years, 55.9% females and 93.2% white ethnicity). Table 1 depicts the baseline characteristics of the participants overall and stratified by retinal age gap quintiles. All features were significantly different across quintiles of retinal age gap, except for a history of diabetes mellitus.

During a median follow-up of 5.83 years (interquartile range [IQR]: 5.74–5.97), a total of 282 (0.80%) participants had stroke events. Table 2 shows the baseline characteristics of participants with and without incident stroke events. Participants who experienced a stroke were more likely to be older, of male gender, less educated, less physically active, obese, and with a history of diabetes mellitus and hypertension, and with a poorer general health status.

As shown in Table 3, after adjusting for age, gender and ethnicity, each one-year increase in the retinal age gap was independently associated with a 5% increase in the risk of stroke (Hazard Ratio [HR] = 1.05, 95% confidence interval [CI]: 1.01–1.08, P = 0.006). This finding remained significant after further adjustments for other confounding factors (HR=1.04, 95% CI: 1.00–1.08, P = 0.029).

Participants in different retinal age gap quintiles had different survival distributions of incident stroke based on log-rank test (P < 0.001). After adjusting for confounding factors, participants whose retinal age gaps were in the fifth quintile had significantly higher risks of stroke compared to those whose retinal age gaps were in the first quintile (HR = 2.37, 95% CI: 1.37–4.10, P = 0.002). However, participants with retinal age gaps in the second, third and fourth quintiles had comparable risks of stroke compared to those with retinal age gaps in the first quintile. (HR = 1.16, 95% CI: 0.80–1.69, P = 0.433; HR = 1.10, 95% CI: 0.72–1.70, P = 0.660; HR = 1.29, 95% CI: 0.78–2.1, P = 0.323, respectively). A trend in the prediction accuracy of stroke across different quintiles of retinal age gaps (HR _trending = 1.17, 95% CI: 1.03–1.33, P = 0.016) was noted.

Results comparable to those of the main analysis were noted when the age-squared term was included in the final model (Additional file 1: Table S2). We also found that retinal age acceleration residual was significantly associated with incident stroke (Additional file 1: Table S3).

The predictive value of a well-established risk factor-based model (including age, gender, smoking status, history of diabetes, systolic blood pressure, and total cholesterol to HDL-cholesterol ratio) and retinal age-based model in the prediction of 10-year stroke risk was shown in Fig. 1. The AUCs of the retinal age-based model was slightly higher than that of risk factor-based model (0.676, 95% CI: 0.644–0.708; 0.662, 95% CI: 0.630–0.693), while the difference did not reach significance (p=0.511).

Future studies are still needed to improve the DL algorithm and investigate into its real-world application and extrapolate its clinical value. First of all, external validations using other datasets comprising populations of different demographic features, such as older age and different ethnicities, and even using different retinal imaging modalities as input, such as optic coherence tomography data, are needed to improve the generalizability of the model and refine the algorithm for more large-scale application. In addition, given the heterogeneity of stroke types and pathogenesis such as ischaemic stroke and hemorrhagic stroke, subgroup analyses for stroke subtypes are needed to further streamline the use of retinal age gap in the prediction of stroke and provide clues to disentangle different mechanisms of stroke prediction using retinal images by the DL algorithm.

Our findings have the potential to provide several important public health and clinical implications after future refinement and clinical proof-of-concept studies. Retinal age gap has great potential to be used as a novel screening tool for individuals at high risks of stroke. Compared with those well-established prediction tools based on classic risk factors, such as the Framingham Stroke Risk Score [44], the retinal age gap assessment is characterized by convenient, non-invasive and cost-effective features, showing an enormous potential for further applications. For example, this DL algorithm may be incorporated into mobile devices, making the assessment of the retinal age gap and risk of stroke more easily accessible. This could facilitate the early referral of patients at high risks of stroke for preventative and therapeutic interventions to reduce the burden of stroke for individuals and society as a whole.

Despite the strengths of the present study including the large sample size, the long follow-up duration, comprehensive adjustments for confounding factors and standard acquisition of fundus images, several limitations should be considered. First, no external validation was performed for the retinal age prediction model, given that we have currently no access to external datasets with long-term follow-up and enough stroke events. Secondly, the UK Biobank cohort is comprised of relatively young participants (within the age range of 40–69 years). Moreover, the quality check tends to exclude the images of participants with older age. Considering the high risk of stroke in the more elderly population, our findings may be subject to generalizability. Further studies are needed to investigate the association between the retinal age gap and incident stroke in the elderly population. Nevertheless, the limited generalizability would not affect the association between the retinal age gap and stroke [45]. Thirdly, due to the low incidence rate of stroke and the lack of data on the causes of stroke, further subgroup analyses could not be conducted. Fourthly, due to the observational design of the study, we could not infer the causal effect of the ageing features of the retina on incident stroke. Lastly, we could not completely exclude the possibility of residual confounding.