We (C. Sabanayagam and T. Y. Wong) performed a systematic search of the MEDLINE (PubMed), and EMBASE databases up to December 2014. Conference proceedings and reference lists of selected articles were also manually scanned to identify possible additional studies. The following terms were used for the MEDLINE search: (exp retinal diseases/, retinopathy.tw., (retina or retinal).tw., microvessel.mp. or microvascular.tw., vessel.mp. or vascular.tw., arteriole.mp. or arteriolar.tw., venule.mp. or venular.tw.) and (diabetes.mp. or Diabetes Mellitus, Type 2/ or Diabetes Mellitus/) and (exp epidemiology/, exp epidemiologic studies/, incidence/, exp prognosis/, predict$.mp., prognos$.tw., risk.tw.). Similar search terms were used for EMBASE. We restricted our search to English language and human studies. Studies were included if they were carried out in general populations, included measurement of retinal vascular calibre and had documented diabetes incidence. Studies were excluded if they were cross-sectional or if relevant data were unavailable. We conducted our systematic review as proposed by the Meta-analysis of Observational Studies in Epidemiology (MOOSE) guidelines .
We invited the principal investigators of the selected studies to participate in this collaborative project who upon acceptance were requested to provide individual participant data on retinal vascular calibre measurements, incident diabetes and baseline measurements of traditional risk factors included in the Framingham Risk Score , such as age, sex, race/ethnicity, current smoking and drinking status, height, weight, systolic and diastolic BP, blood glucose, serum total cholesterol, history of cardiovascular disease (CVD), diabetes and hypertension, and use of glucose-lowering and antihypertensive medication. Since family history was available only in a subset of participants and measures of physical activity were highly variable across the studies, these variables were not considered for inclusion in the analyses. For the present analysis, we excluded individuals with prevalent diabetes or pre-existing CVD or those with missing data on covariates included in the multivariable model. We included those who had values for retinal vascular calibres and information on incident diabetes and covariates.
We identified 1,400 studies with our initial search strategy (Fig. 1). We then identified 16 studies that had assessed retinal calibres and diabetes in a general population. After screening these studies based on inclusion and exclusion criteria, ten studies were excluded due to the cross-sectional study design or multiple publications from the same study samples, leaving six studies eligible for the meta-analysis. One study (Rotterdam Study)  did not have data available at the time of analysis; therefore, five studies were included in the current meta-analysis: the Atherosclerosis Risk in Communities (ARIC, USA) study , the Australian Diabetes, Obesity and Lifestyle (AusDiab, Australia) study , the Beaver Dam Eye Study (BDES, USA) , the Blue Mountains Eye Study (BMES, Australia)  and the Multi-ethnic Study of Atherosclerosis (MESA, USA) . Investigators from the five studies agreed to provide data for the individual participant-level meta-analysis.
Measurement of retinal vascular calibre
Measurement of retinal vessel calibre was similar in each study, with slight variations. Briefly, for participants of each study, either digital or 35 mm retinal photographs centred on the optic disc and macula for a single or both eyes were taken at baseline. Stereoscopic 30°retinal photographs were taken using a Zeiss FF3 camera (Carl Zeiss Meditech, Jena, Germany) after pupil dilation in BDES and BMES; 45° nonmydriatic retinal photographs were taken using a Canon CR6-45NM camera (Canon, Tokyo, Japan) in ARIC and MESA; and a Canon CR45UAF camera and 45° fields were used in AusDiab. In all studies except ARIC, a pre-specified eye, usually the right eye, was used for the measurement and analysis of retinal vascular calibre. Measurements from the left eye were used when measurements from the right eye could not be used. In ARIC, retinal photographs were taken of one randomly selected eye.
Photographs from the three US studies were graded for retinal vascular measurements at the Ocular Epidemiology Reading Center at the University of Wisconsin, Madison using a computer-assisted software program (Interactive Vessel Analysis [IVAN], University of Wisconsin, Madison, WI, USA) that measured the six largest arterioles and venules. Photographs from the two Australian studies were graded at the Centre for Eye Research Australia, University of Melbourne using a computer-assisted vessel measurement system (Retinal Analysis; Optimate, Madison, WI, USA) that measured all vessels larger than 45 μm in diameter [21, 27]. Trained graders masked to participant characteristics performed retinal vascular calibre measurements following a standard protocol in both centres. Briefly, for each image, all arterioles and venules coursing through an area of 0.5–1.0 disc diameter from the optic disc margin were measured and summarised as central retinal arteriolar equivalent (CRAE) and central retinal venular retinal equivalent (CRVE) using the Parr-Hubbard formula  in ARIC and BDES, and Knudtson’s revised formula  in MESA. These equivalents represent the average retinal vessel of calibres of the eye. In AusDiab and BMES, estimates obtained from the Parr–Hubbard formula using the Howard program were converted into estimates equivalent to IVAN measures using a newer formula termed the Parr–Hubbard–Knudtson formula . Reproducibility of retinal vessel measurements were high for both CRAE and CRVE with intragrader and intergrader correlation coefficients ranging from 0.78 to 0.99 [28, 31].
Ascertainment of diabetes
We defined incident diabetes as first occurrence at any follow-up examination. Diabetes definitions used by the individual studies were based on plasma or serum glucose levels and defined as fasting glucose level ≥7.0 mmol/l, or 2 h glucose or casual glucose ≥11.1 mmol/l, or HbA1c >6.5% (47.5 mmol/mol) or the use of diabetic medications with a previous physician’s diagnosis. In ARIC, as retinal photographs were taken at visit three, this visit was considered as baseline for ascertainment of incident diabetes. After excluding individuals with prevalent diabetes at visit three, incident diabetes was defined as individuals with diabetes at the next visit after 3 years (visit four) or during the subsequent annual follow-up telephone calls from 4 to 16 years. Similarly, in MESA, as retinal photographs were taken at visit two, this visit was considered as the baseline for ascertaining incident diabetes at visits three and four.
Assessment of covariates
Information on covariates was collected from a combination of standardised interview and examination procedures in all studies. Data on demographic factors (age, sex, race/ethnicity), lifestyle (smoking, alcohol consumption), personal history of diseases and medication were collected using standardised questionnaires, and information on height, weight and BP measurements were from clinic examinations in each study. In all studies, a mean of two BP measurements was used as the BP value for the participants except for BMES in which BP was defined based on a single measurement. Ethnicity was categorised into white participants, black participants, and Mexican American/Asian/other participants. All studies included middle-aged and elderly participants except AusDiab, which included a wider age range (25–87 years).
Both CRAE and CRVE were normally distributed and the pooled within-study SD was approximately 20 μm for both arteriolar and venular calibres. We estimated the HR associated with each 20 μm decrease in retinal arteriolar and each 20 μm increase in retinal venular calibre in three models. Model 1 adjusted for age, sex and race/ethnicity (white participants, black participants, Mexican American/Asian/other participants). Model 2 included variables from model 1 plus CVD factors including current smoking status (no, yes), current drinking status (no, yes), BMI (kg/m2), systolic BP (mmHg), hypertension (no, yes) and serum total cholesterol level (mmol/l). Model 3 included variables from model 2 plus baseline fasting glucose levels (mmol/l). We estimated these separately for each study using a discrete time proportional hazards models. We pooled the log HR estimates of the different studies by random effects meta-analysis and displayed them in forest plots . The extent of heterogeneity between studies was evaluated with the inconsistency I
2 statistic . Validity of the above two-stage analyses results (with estimates of association calculated separately within each study, followed by pooling of the association estimates from different studies) was assessed by fitting a one-stage multilevel discrete time (Cox regression) random effect model using pooled individual data of all studies, with a random intercept to account for the study-specific effect. This one-stage approach was also used to investigate exposure-covariate interactions because this approach provides a flexible way of examining individual-level interactions . We performed subgroup analyses after interaction tests to detect evidence of any differences in hazards across levels of prespecified characteristics including age, sex, ethnicity, current smoking, BMI and hypertension status.
To test the robustness of our results, we performed several sensitivity analyses. First, we used an approach of leaving out one study at a time, calculating the pooled effect of the remaining studies and comparing the results with the combined effect based on all the studies. Second, to exclude the possibility that undetected diabetes or impaired fasting glucose had caused retinal microvascular changes, we repeated the analysis after excluding participants with follow-up of 5 years or less (n = 5,587 excluded). Third, we repeated the analysis after excluding participants with impaired fasting glucose at baseline (n = 4,683 excluded), defined as a fasting plasma glucose level of 5.6 mmol/l to <7.0 mmol/l. Fourth, we repeated the analysis excluding individuals with incident diabetes defined exclusively by self-report (diabetes detected by annual telephone follow-up in ARIC; n = 1,439 excluded). All analyses were performed using SAS version 9.3 (SAS Institute, Cary, NC, USA) and Review Manager 5 (The Nordic Cochrane Centre, The Cochrane Collaboration, Copenhagen, Denmark).