Diabetic Retinopathy Predicts All-Cause Mortality and Cardiovascular Events in Both Type 1 and 2 Diabetes: Meta-analysis of observational studies

The electronic databases (beginning in 1950 until July 2010) Medline and Embase were searched for the medical subject headings (MeSH) “Diabetic Retinopathy” and “mortality” or “cardiovascular disease” to identify observational studies that report the incidence of all-cause mortality, and fatal and nonfatal CV events in diabetic patients whose DR was evaluated (regardless of the language).

Studies were considered eligible for inclusion if they fulfilled the following inclusion criteria:

1. presented original data of prospective, observational studies;

2. evaluated the presence of DR, defined as any degree and/or severity according to well-validated scales, such as the Early Treatment Diabetic Retinopathy Study severity scale (8), in type 2 or type 1 diabetic patients; and

3. reported all-cause mortality and/or fatal or nonfatal CV events.

When studies reported more than one outcome separately, only all-cause mortality data were included. Fatal and nonfatal CV events were defined as a positive medical history of a CV event, including death due to CV disease and/or any of the following: myocardial infarction, angina pectoris, coronary artery bypass graft, ischemic changes on a conventional 12-lead electrocardiogram, transient ischemic attack, nonfatal stroke, or lower leg amputation.

Data were extracted independently by two investigators with an agreement value of κ = 96%. Disagreements were resolved by a third author.

Extracted data included the clinical characteristics of participants, study design, and follow-up, assessment of DR, and the number of participants who had the outcome according to DR status. Numeric data reported in the articles were used. In the few studies not reporting these data, risk estimates were calculated from the survival curves.

The Newcastle-Ottawa Scale for assessing quality of nonrandomized studies in meta-analysis was used (9).

An overall odds ratio (OR) was calculated to assess the predictive value of DR for all-cause mortality and/or CV events (composite outcome). DR was evaluated either as “any degree” or “advanced DR” compared with the group without DR in analyses stratified by type of diabetes (type 1 and type 2). Advanced DR was defined as the most severe category of DR described: proliferative DR, severe nonproliferative DR, sight-threatening DR, or any combination of these categories.

The Cochran χ² and the I² tests were used to evaluate heterogeneity among studies and a threshold value of P = 0.10 was considered significant (10). The risk estimates were obtained with random-effects meta-analysis because a significant heterogeneity was found among studies in preliminary models.

Meta-regression analyses were used to investigate potential sources of heterogeneity. The factors investigated were age, male sex, proportion of smokers, A1C test, and follow-up period (previously chosen based on their biologic relevance). In a meta-regression analysis of type 1 diabetes, only follow-up and A1C tests were included because the number of studies limited a broader evaluation of potential confounders. A sensitivity analysis including only all-cause mortality as an end point was performed to minimize the influence of CV risk factors as potential confounders.

Receiver operating characteristic (ROC) curves were constructed to obtain the pooled estimates of sensitivity and specificity of DR for the development of outcomes. The average likelihood ratio of the positive and negative test result was calculated. For practical clinical purposes, the post-test probability of the outcome was estimated using the Bayes normogram, considering the pretest probability as the mean proportion of the event in the included population (11).

The possibility of publication bias was evaluated using a funnel plot of a trial’s effect size against the SE. Funnel plot asymmetry was evaluated by the Begg and Egger tests. The trim-and-fill computation was used to estimate the effect of publication bias (12).

All statistical analyses were performed using Stata 11.0 software (StataCorp, College Station, TX).

We identified 3,180 studies in the database searches (Fig. 1). Among these studies, 3,149 were excluded based on title and abstract, leaving 31 studies for further evaluation. Of these, 11 studies were excluded after full-text evaluation (Supplementary Table 1). A total of 20 studies, comprising 19,234 patients, fulfilled our inclusion criteria and were included.

Table 1 summarizes the characteristics of the included studies. A total of 17 studies with a mean follow-up of 8.95 years evaluated type 2 diabetic patients (n = 14,896; mean age, 58.4 years) and four studies with a mean follow-up of 12.37 years evaluated type 1 diabetic patients (n = 4,438; mean age, 32.7 years).

In accordance with the Newcastle-Ottawa quality assessment scale for cohort studies, all studies achieved at least six stars, indicating an overall good quality (Table 2). Funnel plots and the Egger regression test suggested a borderline significant asymmetry in the analysis of type 2 diabetes (P = 0.10). However, the trim-and-fill computation revealed that there were no missing trials, indicating that the publication bias did not interfere with the interpretation of the results. There were no publication biases in analysis of type 1 diabetes in both tests (P = 0.21).

In the pooled analysis of the 17 included studies, the OR for all-cause mortality and/or CV events of the presence of DR was 2.34 (95% CI 1.96–2.80) compared with patients without DR (Fig. 2A). All but three studies reported an increased risk for events. However, a significant heterogeneity among the individual estimates was evident when the magnitude of the association was evaluated (I² = 62.7%, P < 0.001).

The overall sensitivity of any DR for the composite outcome was 43% (95% CI 36–50), and the specificity was 75% (71–80). The positive likelihood ratio was 1.78 (1.57–2.00) and the negative likelihood ratio was 0.75 (0.68–0.82). The post-test probability of the composite outcome for patients who had any degree of DR was 34% considering a pretest probability of 23% (Fig. 3A).

In an exploratory attempt to identify the sources of heterogeneity among trials, we undertook a meta-regression analysis considering as covariates baseline age, proportion of men, proportion of smokers, A1C test, and follow-up. None of them explained the among-studies variance (overall P = 0.80).

The ORs for the composite outcome were also adjusted for CV risk factors (age, sex, diabetes duration, blood pressure, smoking status, albuminuria, and lipid profile) in 12 of the included studies. A meta-analysis was performed taking these adjusted ORs into account, even though these studies did not include the same covariates. The presence of DR still predicted events in this adjusted meta-analysis (OR 1.61 [95% CI 1.32–1.90]).

When advanced DR was evaluated (10 reports), the OR for all-cause mortality and/or CV events was 4.22 (95% CI 2.81–6.33) compared with patients without retinopathy (I² = 63.0%, P = 0.004). The overall sensitivity of advanced DR for the combined outcome was 19% (13–28), and the specificity was 94% (91–96). The positive likelihood ratio was 3.64 (2.52–5.26) and the negative likelihood ratio was 0.84 (0.77–0.92). The post-test probability of an event occurring in an individual with advanced DR was 48% by using the Bayes normogram and considering a pretest probability of 23% (Fig. 3C).

In the pooled analysis of four included studies, the OR for all-cause mortality and/or CV events of any DR was 4.10 (95% CI 1.50–11.18) compared with individuals without retinopathy (Fig. 2). A significant heterogeneity among trials was present (I² = 86.9%, P < 0.001).

The overall sensitivity of any DR for the composite outcome was 70% (95% CI 30–93), and the specificity was 60% (29–85). The corresponding positive likelihood ratio was 1.80 (1.20–2.70) and the negative likelihood ratio was 0.47 (0.22–1.04; Fig. 3B). The post-test probability of an event occurring in a patient with DR was 12% by using the Bayes normogram and considering a pretest probability of 7.5%.

In a meta-regression analysis including the follow-up period and baseline A1C test as possible sources of heterogeneity, the follow-up of the studies explained almost 100% of the differences among the studies. Therefore, a sensitive analysis including only the three studies with follow-up >10 years was performed, demonstrating an OR of DR for the composite outcome of 6.45 (95% CI 3.91–10.85) with heterogeneity of 0% (P = 0.90).

Three studies also presented an adjusted OR of the risk conferred by DR. The presence of any DR still predicted all-cause mortality and CV events when a meta-analysis using these adjusted data were performed (OR 1.58 [95% CI 1.33–1.89]).

Three reports also included evaluation of advanced DR. The OR for all-cause mortality and/or CV events of advanced DR was 7.00 (95% CI 2.22–20.0) compared with patients without DR (I² = 85.0%, P < 0.001). Because only three studies were included in this analysis, the ROC curve could not be constructed.

In the sensitivity analyses evaluating only the eight studies that reported data on all-cause mortality, the presence of DR increased more than twofold the risk for death (OR 2.41 [95% CI 1.87–3.10]) in patients with type 2 diabetes. The risk for death was also greater in patients with type 1 diabetes and DR compared with those without DR, considering three included studies (OR 3.65 [1.05–12.66]; Fig. 2B).

These meta-analyses of cohort studies showed that the presence of any degree of DR or advanced DR was associated with an increased risk for all-cause mortality and CV events (fatal and nonfatal) in both type 2 and type 1 diabetic patients. Data from the current study show that funduscopy can be a practical tool to identify patients who are at increased risk for adverse outcomes, as demonstrated by the influence of DR in the post-test probability of having an event.

Our literature search was extensive; we tested for and found no evidence of publication bias. The quality of original studies was checked according to the Newcastle scale statement and most of the studies fulfilled all components. We also are aware that publication bias and the quality issues of individual studies may still exist despite our best efforts to conduct a comprehensive search and the lack of statistical evidence of existence of bias. Another possible limitation is that our adjusted meta-analyses undertaken were not ideal because the authors of the original studies used different statistical models. However, we must take into account that the potential confounders (covariates) were chosen in each original study according to sample characteristics.

The development of DR has been associated with known risk factors for CV disease such as hypertension, hyperglycemia, and albuminuria (2). Therefore, the association of DR with CV events would be expected, and one might think that DR is just a marker of a worse clinical status. However, in the meta-regression analyses, CV confounders did not explain the between-studies variance, suggesting that these risk factors do not fully account for the observed association. In parallel to that, our results did not change when additional meta-analyses were performed, including only the studies where the risk determined by DR was adjusted for possible confounders in both type 2 and type 1 diabetes.

An important aspect that should be considered is that the therapeutic goals in diabetes care have changed over time. The intensification of glycemic control in patients with type 1 diabetes became widespread after the publication of the Diabetes Control and Complications Trial study in 1993 (13). In patients with type 2 diabetes, a significant reduction in LDL cholesterol levels and blood pressure were observed since 1990 (14). These issues might have influenced mortality rates as well as the effects of risk factors in our meta-analysis because the timeline of included studies ranged from the 1960s to 2008. However, despite changes in goals to be achieved in patients with diabetes, the mortality rate was comparable during the periods of 1971 through 1986 and 1988 through 2000 (15). In parallel to that, the timelines of included studies all overlapped.

Highlighting the importance of DR as a possible novel marker for events in patients with diabetes is that DR was strongly associated with all-cause mortality, beyond its association with CV events. In fact, CV disease was the cause of death in less than 35% of individuals in most of the included studies (4,16), suggesting an additional mechanism by which the presence of DR increases death. In this sense, the presence of autonomic neuropathy could be a possible link between DR and CV events. Indeed, it was recently demonstrated that autonomous deregulation could lead to alterations in blood pressure and cardiac rhythm, which were associated with DR (17,18).

In conclusion, the presence of DR was associated with an increased risk for all-cause mortality and CV events in both type 2 and type 1 diabetes. Further studies are needed to understand the role of the inclusion of DR in mortality prediction scores as well as to understand the link between death/CV events and DR, especially the all-cause mortality DR association.

This study was supported in part by Conselho Nacional de Desenvolvimento Científico e Tecnológico grant 576627/2008-9, Coordenação de Aperfeiçoamento de Pessoal de Nível Superior PNPD 03021/09-2, and Fundo de Incentivo a Pesquisa–Hospital de Clínicas de Porto Alegre, Brazil.

No potential conflicts of interest relevant to this article were reported.

C.K.K. originated and designed the study, acquired data, analyzed statistics, analyzed and interpreted data, drafted the manuscript, and critically revised the manuscript for important intellectual content. T.C.R. acquired data, analyzed and interpreted data, and critically revised the manuscript for important intellectual content. L.H.C. and J.L.G. analyzed and interpreted data and critically revised the manuscript for important intellectual content. M.J.A. originated and designed the study, analyzed and interpreted data, and critically revised the manuscript for important intellectual content.