Retinal venular tortuosity and fractal dimension predict incident retinopathy in adults with type 2 diabetes: the Edinburgh Type 2 Diabetes Study

The Edinburgh Type 2 Diabetes Study (ET2DS) is a longitudinal cohort of older men and women based in Lothian, Scotland, designed to investigate the role of risk factors for vascular complications of type 2 diabetes. Methods have been previously reported in detail including in Price et al (2008) [10]. In brief, recruitment occurred in 2006–2007 when participants aged 60–75 years were randomly selected within sex and 5-year age bands from the Lothian Diabetes Register, with the final cohort shown to be largely representative of the target population of all older men and women with type 2 diabetes residing in Lothian [11]. All people within the Lothian Diabetes Register have a diabetes diagnosis according to WHO criteria, but in order to be enrolled in the ET2DS participants had to meet certain criteria to ensure a robust diagnosis of type 2 diabetes [12]. These criteria included that participants had to be taking oral glucose-lowering medications and/or insulin, or, if participants managed their diabetes through diet control methods alone, had to have an HbA_1c measure of >48 mmol/mol (6.5%) at the baseline clinic. Further investigation of clinical records was undertaken if there was concern of the diabetes status of a potential participant.

Ethical permission was granted by the Lothian Medical Research Ethics Committee at baseline and follow-up, and all participants gave written informed consent at recruitment and prior to subsequent clinic attendance. Data were collected at designated research clinics as well as through linkage to routine medical and death records, at baseline and during follow-up.

Baseline fasting blood samples were used to measure total serum cholesterol, HDL-cholesterol and HbA_1c, and early morning urine samples to measure creatinine and albumin. Height, weight, and systolic and diastolic brachial blood BPs were measured using established standard operating procedures by trained researchers, and self-administered questionnaires were used to collect data on diabetes history and treatment, smoking habits, medications and comorbidities. For cardiovascular and cerebrovascular events, answers from baseline questionnaires on medical history, the WHO chest pain and Edinburgh Claudication questionnaires, and results from a standard 12-lead ECG were combined with routine hospital discharge data [11] to define macrovascular disease (one or more of myocardial infarction, angina, stroke or transient ischaemic attack).

To determine retinopathy status at baseline and obtain retinal photographs to assess quantitative traits, all study participants were invited to an eye appointment within 3 weeks of their original baseline clinic appointment [13]. Mydriasis was achieved using 1% tropicamide drops and standard seven-field non-stereoscopic retinal colour photographs were taken of both eyes at 35° using a high-resolution TOPCON TRC-50FX digital retinal camera (Topcon Optical Company, Tokyo, Japan). The images were taken by a single, specially trained medical photographer. Retinopathy grading was undertaken by two trained optometrists, working independently, using all seven fields of both eyes and a predefined protocol using Early Treatment Diabetic Retinopathy Study (ETDRS) criteria [14]. An ETDRS grade of ≥20 in any field of either eye was considered as presence of diabetic retinopathy at baseline.

To determine incident retinopathy, results of retinopathy grading by the Scottish Diabetic Retinopathy Screening Programme (Scottish DRS) were requested, which operates an annual screening programme for all people in Scotland with diabetes. We were able to request all screening data through the 10-year follow-up period, which helped to reduce bias due to attrition. After pupil dilation using 1% tropicamide, images were taken of both eyes with a TOPCON NW8 fundus camera with Nikon D700 digital backs at a 45–50° angle. Photography and grading were performed by specially trained and accredited medical photographers using a grading scheme that ranges from R0 (no diabetic retinopathy) to R4 (proliferative diabetic retinopathy) [15]. Grade R1 is considered mild, R2 is observable, R3 is referable and R4 is proliferative. For this analysis, incident retinopathy was considered any grade R1 or above during the follow-up period. It is not uncommon for a person that has received a grade of R1 to then revert to R0 at another grading because of artefacts in the imaging, healing of microaneurysms without further progression of retinopathy or other reasons. Because of this, if a participant received a grade of R1 but no higher during the follow-up period, a grade of R1 at a minimum of two consecutive screenings was required for them to be considered to have incident retinopathy. A participant was also determined to have incident retinopathy if there were laser photocoagulation scars present (an indication of prior treatment for diabetic retinopathy).

Retinal vessel traits were measured at baseline by a single researcher (E. Sandoval Garcia), trained specially by an expert senior researcher (T. J. MacGillivray), using the Vascular Assessment and Measurement Platform for Images of the Retina (VAMPIRE) software (version 3.1.0, Universities of Edinburgh and Dundee, UK) [16]. The right eye image was analysed, unless unsuitable because of poor image quality leading to hazy or obscured views of the retinal vasculature, in which case the left eye image was used. Measures of inter- and intragrader agreement were made with involvement of a third, specially trained researcher (R. B. Forster). Measurements were taken using the optic disc-centred field.

To analyse an image, the user uploads it to the software upon which the boundary of the optic disc and the position of the fovea are detected automatically. The user can adjust these features if necessary (e.g. the optic disc in some patients may be unclear because of pathology or the fovea may have been masked by poor illumination at image acquisition). The software then automatically generates a map of the retinal vasculature and attempts to classify vessels as arterioles or venules, which can be edited by the user if some vessels are labelled incorrectly. A final measurement step is undertaken to generate quantitative traits.

The traits assessed for this analysis include vessel calibre, or width, measured as summary indices (central retinal arterial equivalent [CRAE] and central retinal venular equivalent [CRVE]), tortuosity, which describes the degree of curvature of a vessel (i.e. how much the vessels twist and turn, arterial and venular separately), and total fractal dimension, which describes how the vascular pattern fills a two-dimensional space and is thus a measure of complexity or sparsity of the vascular network. For CRAE and CRVE, the software identifies the six widest arterioles and venules that cross a standardised region called zone B [16]. Tortuosity is calculated using the six widest arteriole and venule vessels crossing another standardised region, zone C [17]. Fractal dimension was measured by multifractal analysis using the generalised sand box method, which generates an estimation of fractal dimension of several scales from zone C of a binary map [18, 19].

All variables were checked for normality and the following were transformed prior to analysis: duration of diabetes, albumin/creatinine ratio (ACR), tortuosity and fractal dimension. All of these except fractal dimension were log transformed using the natural log. Fractal dimension, which was only slightly skewed, was modified using rank transformation. In addition, fractal dimension was standardised for use in regression modelling because it is a continuous, unitless measure that falls between 1 and 2, and can be difficult to interpret in an unstandardised form. Values were standardised using the mean and SD to produce a scaled variable with a mean of 0 and SD of 1 (values maintain the same relationship to one another as the unscaled variable). The resultant OR can be interpreted as the odds given an increase of one SD.

To evaluate reliability and agreement of the measured retinal traits, interclass correlation coefficients (ICC) were used to measure intra- and intergrader agreement. A two-way mixed effects design was used to evaluate the mean of two graders or two measurements for consistency [20]. For baseline measurements, 3% (n = 30) of participants were used for the analysis of inter- and intragrader reliability.

Variables of interest were initially evaluated using Welch’s unpaired two-sample t test for continuous variables, or Wilcoxon test if assumptions were not met, and Pearson’s χ² test with Yate’s continuity correction for categorical variables.

Logistic regression analyses were used to evaluate the relationship between retinal vessel traits and incident retinopathy, and the model was built by adding blocks of covariates to better understand the impact of the variables within the model. First, an unadjusted model was generated, then age and sex were controlled for, followed by cardiometabolic risk factors (BMI, smoking status, systolic BP, and total cholesterol:HDL-cholesterol ratio), then the addition of diabetic risk factors (HbA_1c, duration of diabetes and diabetic treatment type) and finally vascular disease risk factors were added (a composite history of macrovascular disease as well as ACR). When measuring vessel width, arterial and venular vessels were evaluated separately, as CRAE or CRVE, but the corresponding value for the opposite measure was added as a covariate in the final model because of the large amount of shared variance. If there was no evidence of an association at the unadjusted or age- and sex-adjusted stages, the analysis was halted to avoid the risk of findings through multiple testing. Only cases with no missing data were included in logistic regression, and we planned to investigate any covariates with missing data exceeding 3.5% of total cases.

Subgroup analysis was planned to evaluate the different severities of retinopathy gradings. Grading groups would include mild retinopathy (R1), moderate (R2 and R3) and proliferative retinopathy (R4). However, the very small number of incident retinopathy cases between R2 and R4 meant that such analysis was not undertaken. Ethnicity subgroups were not evaluated as the ET2DS consists of 98% white British participants, reflecting the general population at baseline.

Initially, a Cox proportional hazards regression model was considered, but the proportional hazards assumption was violated and could not be resolved, so logistic regression was the preferred approach.

If any vessel traits were found to be independently associated with incident retinopathy they would be evaluated, in combination with the most highly predictive risk factors (HbA_1c, systolic BP and ACR) [2, 21‐24], using Harrell’s concordance statistic or C statistic to see if the model was improved, which is used to evaluate the discriminative ability of the model. The models would also be evaluated using the likelihood ratio test and Akaike’s information criterion (AIC).

All final models were evaluated for concerns with multicollinearity and linearity. Goodness of fit was evaluated using the Hosmer–Lemeshow test, residuals were inspected for outliers and influential cases, and interaction with predictor variables was investigated.

A two-sided p value ≤0.05 was used to indicate evidence of statistical significance. Logistic regression results were presented as OR and 95% CI, and all statistical analyses were performed using R version 3.5.1 [25].

In the unadjusted model, decreased CRAE (i.e. narrower arterioles) was associated with incident retinopathy (OR 0.93; 95% CI 0.87, 0.99; p = 0.028). This relationship was maintained after multivariable adjustment for age and sex, as well as after adding cardiometabolic and diabetes-related risk factors, with little change in the point estimate at each step. However, the relationship lost statistical significance after vascular disease history and CRVE were incorporated (OR 0.95; 95% CI 0.87, 1.03; p = 0.212).

In unadjusted and multivariable-adjusted models, evidence for an association between decreased CRVE (i.e. narrower venules) and incident retinopathy was weak (unadjusted OR 0.95; 95% CI 0.91, 1.00; p = 0.058; OR after adjustment for age, sex, cardiometabolic risk factors 0.95; 95% CI 0.91, 1.00; p = 0.048). The association was not evident after further adjustment for diabetes-related risk factors, vascular risk factors and arterial width.

Arterial tortuosity was not associated with incident retinopathy (unadjusted OR 0.99; 95% CI 0.81, 1.22; p = 0.946). There was evidence of a relatively strong association between increased venular tortuosity (i.e. more twisted venules) and incident retinopathy, OR 1.43 (95% CI 1.11, 1.84; p = 0.005) and this was maintained after each block of covariates was added (full multivariable-adjusted OR 1.51; 95% CI 1.15, 1.98; p = 0.003).

Despite little evidence of an association between decreased fractal dimension (i.e. a sparser vascular network) and incident retinopathy in initial models (unadjusted OR 0.80; 95% CI 0.63, 1.01; p = 0.059; age and sex-adjusted OR 0.79; 95% CI 0.62, 1.00; p = 0.050), evidence of an association was evident after adjustment for cardiometabolic and diabetes risk factors (OR 0.76; CI 0.60, 0.98; p = 0.033) and in full multivariable analysis (OR 0.75; 95% CI 0.58, 0.96; p = 0.017).

Venular tortuosity and fractal dimension were added, separately, to a model that contained the most highly cited risk factors for diabetic retinopathy (HbA_1c, ACR and systolic BP). For venular tortuosity the discriminative ability of the resultant model improved from 0.624 to 0.640, based on the C statistic, and was confirmed by the Likelihood ratio test, p = 0.013 and AIC measure, which decreased from 487.65 to 483.43, showing a statistically significant improvement in the model. However, when fractal dimension was added, the C statistic decreased from 0.625 to 0.621 (p value 0.048 for likelihood ratio test), indicating the addition did not improve the model. Values are shown in Table 3.

Model fit was evaluated by comparing the C statistic and AIC between the unadjusted model and the fully adjusted model and the Hosmer–Lemeshow goodness of fit test was used to look for evidence of poor model fit. All models demonstrated good fit, and when evaluating residuals there was no evidence of strong outliers or influential cases, and no indication of interaction.