Validation of a deep learning-based retinal biomarker (Reti-CVD) in predicting cardiovascular disease: UK Biobank data | BMC medicine

ethics statement

This retrospective study was deemed exempt from institutional review board (IRB) review by the SingHealth Centralized Institutional Review Board (CIRB). This study adhered to the principles of the Declaration of Helsinki. Written informed consent was obtained from participants in the original studies. [10, 11].

Study population

We used clinical data and retinal photographs from the UK Biobank, a population-based prospective cohort in the United Kingdom. [12]. The UK Biobank protocol is available online [13].

We excluded (1) duplicate retinal photographs (n = 18,423), (2) those who had type 1 diabetes (n= 288), (3) those with pre-existing CVD at baseline (n= 7624), (4) low quality photographs (n= 11,115) and (5) those under 40 years old (n= 1) (Additional file 1: eFigure 1). Pre-existing CVD was defined as a prior history of coronary disease, other cardiac disease, stroke, transient ischemic attack, peripheral arterial disease, or cardiovascular surgery and patients undergoing cardiovascular procedures based on the Classification of Interventions and Procedures version 4 ( OPCS -4) [10].

A total of 48,260 participants, representing the general population with no history of CVD, were included for analysis. In addition, we also defined three risk subgroups: (1) the statin-free cohort, individuals who do not take a statin; (2) the stage 1 hypertension cohort, individuals with stage 1 hypertension and not on antihypertensive medication; and (3) the middle-aged cohort, individuals aged between 40 and 64 years at baseline (additional file 1: eFigure 1). Since the goal of Reti-CVD is the primary prevention of CVD, we focused on individuals who have a lack of awareness of risk factors associated with cardiovascular disease. In a previous publication, the authors reported, based on analyzes stratified by age, that the rates of ignorance were higher in individuals aged between 40 and 49 years and lower in individuals aged over 70 years. [14]. Therefore, our middle-aged cohort is classified on the basis of vulnerability.

The retinal photographs included in the study were taken using the Topcon 3D OCT-1000 Mark II (Topcon Corporation) between December 7, 2009 and July 21, 2010. The retinal cameras used in the training set include the automatic retinal camera AFP-210 non-mydriatic retinal camera (NIDEK Corporation, Aichi, Japan), TRC-NW8 non-mydriatic retinal camera (Topcon Corporation, Tokyo, Japan) and Nonmyd AD (Kowa Co. Ltd., Shizuoka, Japan). We have not included the Topcon 3D OCT-1000 Mark II (Topcon Corporation) in our training set.

The other variables used in this study were defined below. Pre-diabetes and diabetes were defined based on (1) medical history and (2) glucose ≥ 5.5 mmol/L. Medical histories of high cholesterol, type 1 diabetes, and antihypertensives were self-reported and collected from an initial assessment questionnaire about medical conditions. The status of smoker was also self-declared and categorized into “smoker in life” and “never”.

Definition of cardiovascular disease events

At UK Biobank, we use hospitalization and mortality data provided by National Health Service (NHS) records. The main outcome of interest in the current study reflected the outcome used in the QRISK3 risk score: fatal cardiovascular events (ICD-10 I00-99, F01, Q20-Q28, C38.0, P29, G45) [15] or nonfatal coronary disease, ischemic stroke, or transient ischemic attack (ICD-10 G45, I20-24, and I63-64) [16].

QRISK3 and borderline-QRISK3 group score calculation

For each individual, the QRISK3 score was calculated using the R package, version 3.6 [17]. QRISK3 score distribution is provided in additional file 2: eFigure 2. For comparison with groups of three Reti-CVD strata, we divided subjects into 5 groups based on QRISK3 score (%) (≥ 0 to < 5; ≥ 5 a > 10; ≥ 10 to > 15; ≥ 15 to < 20; and ≥ 20). Furthermore, as the recommended threshold for initiating statins and antihypertensive drugs from a 10-year CVD risk is 10%, we defined a “borderline risk” group as those who had a QRISK3 score between 7.5 and 10%. . Specifically, we divided subjects into 5 groups compared to the previous group of three QRISK3 risk strata as described in the NICE guidelines. Herein lies the difference in the ranges used to assess CVD events. We used 5% intervals compared to the previous 10% intervals so that a more detailed comparison could be made with Reti-CVD.

New retina-based cardiovascular disease risk stratification system ≈

Details of RetiCAC model development and earlier validations have been described elsewhere. [9]. Briefly, the RetiCAC score was defined based on a probability score derived from our binary classification deep learning algorithm (absence versus presence of coronary artery calcium [CAC]). Probability scores ranged from zero to one, with a high value indicating a high probability of the presence of CAC. First, we improve the RetiCAC algorithm using more datasets that include retinal photographs and CT scans from Korea (additional file 3: eDocument 1). Second, we proposed new cardiovascular disease risk stratification groups (ie, Reti-CVD) with optimized cut-off values ​​based on the 40th and 95th percentile of the Reti-CVD score among 48,260 participants after exclusion. Cut-off values ​​were determined to have a similar incidence rate: the Reti-CVD low-risk group was projected to have a similar incidence rate as the QRISK3 risk group from 0 to 5% (i.e., 2.5 per 1000 person-years), and the Reti-CVD moderate risk group was projected to have an incidence rate similar to that of the QRISK3 risk group of 5 to 10% (i.e., 7.0 per 1000 person-years) . We used these proposed cut-off values ​​to further stratify CVD risk in UK Biobank participants.

Statistical analysis

Analyzes were performed using P< 0.05 as significance level, Stata/MP version 14.0 for survival analysis and R version 3.4.4 for net reclassification index (NRI) estimation using R survIDINRI package [18]. Descriptive statistics are provided for all participants and also according to the three Reti-CVD risk groups.In the UK Biobank, hospitalization and mortality data were available up to 5 May 2021 at the time of analysis, and each participant was followed up to 11.4 years from the date of the initial visit. In the survival analysis, each patient was followed for up to 11.4 years (median follow-up, 11.0 years) from the initial visit date to the last follow-up date or the date of the CVD events.In all populations, the cumulative incidence rate of cardiovascular events was assessed in the three groups (low, moderate and high risk) defined by Reti-CVD using the Kaplan-Meier method. The Cox proportional hazards model was used to estimate hazard ratios (HRs) and trends in HRs and respective P-values ​​were examined by fitting a linear model for the three categories. Unadjusted HR was provided according to three Reti-CVD risk groups, and QRISK3-adjusted HR trend was provided.

To help future patients make an informed decision about statin and initiation of antihypertensive medication, we only included the borderline QRISK3 group that had a QRISK3 score between 7.5 and 10% 10-year CVD risk. In the borderline QRISK3 group, the cumulative incidence of cardiovascular event rate was evaluated in the three groups (low, moderate and high risk) according to Reti-CVD and compared with participants with QRISK3 5–7.5% and 10–12, 5%. The same analysis was repeated for the middle-aged group (40 to 64 years old).

The incremental prognostic value of Reti-CVD over QRISK3 in predicting CVD events was assessed using C statistics and continuous net reclassification index (NRI) [18]. In addition, decision curve analysis was used to compare the net benefit of the models at different thresholds over the QRISK3 model and the Reti-CVD model plus age and sex. Age and gender were included for fair comparison because the QRISK3 is based on the survival model including multiple risk factors including age, gender, smoking and comorbidities. Continuous models were presented as decision analysis curves (demonstrating the potential results of using any threshold in this model) and models with the highest net benefit were considered to perform better.

Validation of a deep learning-based retinal biomarker (Reti-CVD) in predicting cardiovascular disease: UK Biobank data | BMC medicine

Leave a Reply

Your email address will not be published.

Scroll to top