Estimated Benefits of Glycemic Control in Microvascular Complications in Type 2 Diabetes

Methods

Markov Model

A Markov model was constructed to analyze the risk for retinopathy and nephropathy in patients with type 2 diabetes. The structure of the model is shown in the (Figure 1). Estimates for the two complications were modeled separately so that they were assumed not to interact. The simulated patients progressed sequentially through increasingly severe disease states; death could occur in any disease state (Figure 1). Incidence and progression of retinopathy were defined as in the DCCT [4, 17]. Blindness was defined as corrected visual acuity of 20/200 or worse and was restricted to blindness caused by retinopathy or its sequelae (for example, macular edema). Microalbuminuria was defined as an albumin level of 30 to 300 mg/g of creatinine, proteinuria was defined as a protein level greater than 300 mg/g of creatinine, and end-stage renal disease was defined as renal disease that required dialysis or transplantation.

Amputation was not modeled because of a lack of evidence on the relative contributions of microvascular and macrovascular disease to the risk for amputation. Moreover, the same patients who were identified in our analyses as being at high risk for retinal and renal complications (and who therefore will benefit more from intensified glycemic control) will have a similarly high risk for neuropathy and its associated outcomes [18].

Our study addressed clinical risks and benefits associated with glycemic control and did not directly evaluate costs. The costs of decreasing hemoglobin A_1c levels are not well defined for patients with type 2 diabetes because glucose control can be improved by many methods. Each method has different implications for the “costs” to the patient, payers, and society. However, we present results in the form of expectation functions so that if the costs and effectiveness of a specific intervention are known, estimates of the cost per complication prevented can be calculated from our tables.

Assumptions

The construction of the model was based on the states defined in the DCCT, which provided estimates for the rates of early microvascular disease. Our review of the literature and input from diabetes experts consistently identified the Rochester, Minnesota, cohort study of diabetic patients and the Wisconsin Epidemiologic Study of Diabetic Retinopathy [3, 9-11, 15] as the best studies with which to provide estimates of the rates of progression to end-stage outcomes; thus, these studies were used in the base-case analysis.

Further literature was identified by searching the MEDLINE database search with the keyword diabetes, cross-referenced with retinopathy, blindness, visual loss, nephropathy, kidney disease and failure, renal disease and failure, blood glucose, glycemic control, mortality, and hemoglobin A1c. We reviewed the abstracts of the identified articles and obtained the articles that were relevant to the model structure. Additional literature was identified by review of references in these articles and through discussion with national experts in the field.

Patients were assumed to have no clinically detectable microvascular complications at the time of diagnosis of diabetes. Although up to 20% of patients have retinopathy at the time of diagnosis [19], we excluded this subgroup because data on the distribution of these patients across hemoglobin A_1c levels are lacking. In addition, patients who present with complications have already declared themselves to be at high risk and thus should be considered for intensive control.

We assumed that incidence and early progression of retinopathy and development of microalbuminuria were related to level of glycemic control at the rates of progression shown in the DCCT [4, 14]. Each 10% increase in hemoglobin A_1c level is accompanied by a 20% increase in the rate of developing microalbuminuria, a 56% increase in the rate of developing retinopathy, and a 64% increase in the rate of progression of retinopathy [14]. Further progression (to renal disease beyond microalbuminuria and from retinopathy to blindness) was assumed not to be related to level of glycemic control [20-25]. Data for sensitivity analyses were derived from the 95% CIs in the DCCT and the range of estimates available in the literature (Table 1).

Mortality estimates were based on U.S. data from the Department of Vital Statistics [40]. These estimates were modified to reflect the higher mortality rates in patients with type 2 diabetes [41-47]. Mortality was not adjusted for level of glycemic control. Some cohort studies show a relation between level of glycemic control and mortality, but it is not clear whether a causal relation exists [48].

Nephropathy, even at early stages, has been reported to be associated with increased rates of death, particularly death from cardiovascular disease [37, 49-52]. Much of this association may be due to confounding, but the effect persists after adjustment according to available severity and comorbidity measures [50]. The base-case model assumes that early nephropathy does not increase mortality rates. However, to evaluate the potential importance of this factor, we compared the mortality benefits of improved glycemic control by using half of or the full observed association between early renal disease and increased mortality rates.

Model Structure and Implementation

The model was run independently for each age of diabetes onset and level of glycemic control and was implemented by creating matrices of the probabilities of going from one health state to the next during a 1-year period (Figure 1). The percentage of patients developing each level of complication was tabulated annually, continuing through 90 years of age. Mortality rates were updated at 5-year intervals. This technique represents a standard implementation of a nonstationary Markov process [53, 54].

Average life expectancy was calculated by multiplying the proportion of patients dying during each year by the total number of years survived. Average time spent in each disease state was determined by dividing the total time spent in a state by the number of patients who develop that state. All calculations and modeling were performed by using Stata for Windows (Stata Corp., College Station, Texas).

Sensitivity analyses were conducted by using varying estimates at all transition points (Table 1) [54]. We included the full range of estimates in the literature in one-way sensitivity analyses. However, because some studies had extreme values for the transition probabilities (most likely because of small sample sizes or atypical patient populations), three-way sensitivity analyses were conducted by using the midpoint between the base-case and extreme estimates. We also tested the validity of the model by modifying the simulated patients to fit the characteristics of the populations of various actual studies of natural history (for example, the mean values for age, hemoglobin A_1c level, and initial prevalence of disease found in these studies were used to generate a limited set of results) and by comparing the predictions of the model (within one-way sensitivity analysis range) with the observed values from these studies.

Using the model predictions and the characteristics of a population of patients with type 2 diabetes at a large staff-model health maintenance organization [12], we examined the benefits of targeting (by hemoglobin A_1c level and age at diabetes onset) and treating a population of diabetic patients with a hypothetical intervention that improved hemoglobin A_1c levels by 2 percentage points or to a value of 7%. The population of the health maintenance organization had a mean age (±SD) at diabetes onset of 60 ± 11 years and a mean hemoglobin A_1c level of 8.2% ± 1.3%.

Results

Patient Risks and Benefits

The calculated lifetime risks for developing blindness and end-stage renal disease, with ranges from the one-way sensitivity analysis, are presented in Table 2 and Table 3. The risks for developing end-stage outcomes are highest in patients who develop diabetes at a young age and in those who have poor glycemic control; thus, the benefits of treatment to reduce hemoglobin A_1c levels are most marked in these groups. For example, in patients who develop diabetes at 45 years of age, improving glycemic control from a hemoglobin A_1c level of 9% to a level of 7% results in an estimated 2.3-percentage point decrease in lifetime risk for blindness caused by diabetic retinopathy (from 2.6% to 0.3%). In contrast, the same change in hemoglobin A_1c level in patients with diabetes onset at 65 years of age would decrease the lifetime risk for blindness by only 0.5 percentage points (Table 2). However, these patients will achieve much greater benefit when moving from poor to moderate glycemic control (lifetime risk for blindness will decrease by 1.4 percentage points if the hemoglobin A_1c level is decreased from 11% to 9%). Results for development of end-stage renal disease follow a similar pattern, but the risk is even more strongly associated with age at diabetes onset than it is with blindness (Table 3).

We also evaluated the effect that prevention of early nephropathy may have on life expectancy. Observational studies have shown that microalbuminuria and proteinuria are related to excess cardiovascular mortality in patients with type 2 diabetes. If we assume that the relation is not associated with other related factors (such as hypertension), then improving glycemic control offers substantial increases in life expectancy. For example, a 2-percentage point decrease in hemoglobin A_1c level results in a potential increase of 1.3 years in life expectancy among patients who developed diabetes at 45 years of age, an increase of 0.9 years at 55 years of age, an increase of 0.5 years at 65 years of age, and an increase of 0.2 years at 75 years of age. However, if only half of the observed excess mortality is causally related to early nephropathy or if half of the effect is attenuated by standard therapies (such as aspirin, angiotensin-converting enzyme inhibitors, therapy for hyperlipidemia, and optimal blood pressure control), substantive treatment benefits are limited to patients who developed diabetes at a younger age and those who have poor glycemic control.

Targeting High-Risk Patients

Health care systems are often interested in the potential benefits of offering an intervention (such as a disease management program) to a group of patients. We examined the effects of offering an intervention that decreases hemoglobin A_1c levels by 2 percentage points to a diabetic patient population at a large staff-model health maintenance organization [12]. If the intervention was offered to all patients with hemoglobin A_1c levels of 8% or greater, the number needed to treat (for life) to prevent one case of blindness would be 79. However, because benefit varies substantially according to level of baseline hyperglycemia and age at diabetes onset, we can further stratify patients for risk and examine the relative benefit of offering the intervention to subsets of the population.

The effects of using different hemoglobin A_1c criteria for offering the intervention are presented in Table 4 (because preventing end-stage renal disease requires many patient-years of treatment, only results for blindness are presented). By offering the intervention to the 2.1% of patients with hemoglobin A_1c levels of 12% or greater, one case of blindness would be prevented for every 25 persons treated (for life); for every 34 years of treatment, 1 year of blindness would be prevented. However, the marginal gain of decreasing the treatment threshold to a hemoglobin A_1c level of 11% is only about half; compared to a cutoff of 12%, a cutoff of 11% would require administering treatment for 67 additional patient-years for every additional year of blindness prevented (Table 4).

Table 5 shows how targeting patients on the basis of age at diabetes onset can improve the efficiency of the intervention. Indeed, we can identify a strategy whereby 84% of the potential benefit of the intervention can be achieved by intensifying treatment in only 17% of all diabetic patients (hemoglobin A_1c cutoff levels of 9% for patients who developed diabetes in their 40s, 10% for patients who developed diabetes in their 50s, 11% for patients who developed diabetes in their 60s, and 12% for patients who developed diabetes in their 70s).

Sensitivity Analyses and Validity Testing

The sensitivity analyses resulted in a range of outcomes that do not substantially affect our main conclusions. Table 6 shows the bounds of the three-way sensitivity analyses in terms of the impact that improving glycemic control has on reducing the lifetime risk for blindness. The ranges of the sensitivity analyses for end-stage renal disease follow a similar pattern. The widest range of risk was found in the patients with the youngest age at diabetes onset, and the ranges of absolute risk reduction for patients with diabetes onset at older ages were much narrower. Other assumptions that varied in the sensitivity analyses (for example, mortality rates) also had little effect on our results. The estimates of the benefits of moving from moderate to tight glycemic control are more stable than those of moving from poor to moderate glycemic control. Nevertheless, the variation in estimates seen in the sensitivity analysis do not substantially alter the conclusion that the benefit of moving from poor to moderate glycemic control is much more substantial than that of moving from moderate to tight glycemic control.

One way to test the validity of this model is to determine whether it can accurately predict the rates of microvascular disease reported in actual patient populations. We compared the outcomes predicted by our model with those observed in the UK Prospective Diabetes Study (UKPS) [55] and in various cohort studies [3, 10, 11, 30, 56]. The model accurately predicts rates of outcomes for 10 of 12 comparisons. The only exceptions are 1) the rate of early retinopathy in the UKPS (28% to 32% according to our model compared with 37% seen in the trial [55] and 2) the 10-year rate of blindness from retinopathy in the Wisconsin Epidemiologic Study (0.1% to 0.8% according to our model compared with approximately 1.1% seen in the study [30]; this rate is calculated on the basis of the estimate that one quarter of cases of blindness in the study were due to diabetic retinopathy) [15]. However, estimates of blindness from both the UKPS and Rochester study, rates of retinopathy from the Wisconsin and Rochester cohorts, and all estimates of renal disease fall within the ranges predicted by the model. The UKPS represents completely external validation. Some of the validation analyses have a circular component because the data from studies used to generate estimates were used to validate the model. However, no study other than the DCCT contributed more than one transition probability. Thus, these analyses offer evidence that the model structure and estimates are accurate, and they support the use of DCCT estimates to model outcomes in patients with type 2 diabetes.

Discussion

Diabetes is the most frequent cause of blindness and renal failure in industrialized nations [57, 58]. Improving glycemic control in patients with type 2 diabetes may substantially reduce the rates of these devastating complications. Therefore, many authors now recommend that blood glucose levels be maintained at almost-normal levels as an important goal of therapy for all patients with diabetes [59, 60]. Others have countered that little evidence suggests that improving glycemic control will strongly affect patients with type 2 diabetes [5]. To help settle this debate, there have been calls for identification of hemoglobin A_1c levels that alter the risk–benefit ratio of intensive glycemic management and for determination of which patients are most likely to benefit from improvements in glycemic control [7]. Our study gives estimates to help guide the management of type 2 diabetes.

It should be emphasized that there is only preliminary evidence that improved glycemic control in patients with type 2 diabetes leads to a reduction in microvascular complications [28]. Thus, the calculated benefits from this model depend on extrapolation from the experience with type 1 diabetes. Although some researchers question the use of estimates from type 1 diabetes to model the development of early microvascular complications in type 2 diabetes, the epidemiologic literature supports this conclusion [3, 9, 28, 31, 35, 61]. We tested the validity of this assumption by determining whether our model predicts development of these complications in observational studies of patients with type 2 diabetes.

The results of our study suggest that many patients who develop type 2 diabetes at a young age will receive substantial benefits from tight glycemic control. For many patients with type 2 diabetes, however, there may be little additional benefit once moderate glycemic control has been achieved. Some patients will be interested in further efforts to reduce their risk for advanced complications, even when treatment is intrusive. For many others, however, such interventions will not be desired. Our results can greatly aid this patient–physician decision-making process.

The results of our study also allow health care systems to estimate the expected benefits of targeting interventions to decrease hemoglobin A_1c levels. Although we did not tie specific costs to interventions, estimates of cost-effectiveness can be derived from our tables. For example, if a health care system has an intervention that will reduce hemoglobin A_1c levels by 2 percentage points and will cost $25 per patient per month, this intervention would incur up-front costs of about $32 400 per year of blindness prevented when used to treat patients who have hemoglobin A_1c levels of 10% and developed diabetes between 40 and 49 years of age. In comparison, the costs would be about $117 000 for diabetic patients who had similar A_1c levels but developed diabetes between 60 and 69 years of age. Although the greater efficiency of targeting high-risk patients for intensified glycemic control is intuitively obvious, the magnitude of this effect is somewhat startling. If an intervention that decreases hemoglobin A_1c levels by 2 percentage points is available, a strategy can be outlined that achieves more than 80% of the potential benefit (preventable patient-time spent being blind) by intensifying treatment in less than 20% of one health maintenance organization's patient population.

If early diabetic nephropathy is causally related to the increased mortality rates seen in observational studies, a major benefit of improved glycemic control is an increase in life expectancy. It remains unclear whether other therapies (such as angiotensin-converting enzyme inhibition) could negate some or most of this effect. Nevertheless, this potential increase in life expectancy, largely due to a decrease in atherosclerotic disease, serves as a reminder of the importance of macrovascular disease in type 2 diabetes. A major limitation of the literature, and thus of our model, is the inability to define the relation between macrovascular disease and glycemic control [4, 62-66]. If ongoing trials show that improved glycemic control decreases the rate of macrovascular disease, intensive control is likely to provide marked improvements in life expectancy. Alternatively, if intensified glycemic treatment regimens increase cardiovascular risk by as little as 10%, the excess mortality would overwhelm the benefit seen from prevention of microvascular complications. The final results of the UKPS [67] may help to clarify this issue.

Another limitation of our analysis is the inability to model the potential benefits that improved glycemic control may have in preventing daily glucose swings. These swings may negatively affect symptoms and outcomes. Because of insufficient information, we did not model the benefits of improved glycemic control on outcomes related in part to neurologic complications (such as amputation). As is true with any decision model, it is not possible to evaluate all potential treatment risks and benefits.

We also caution that the estimates used in our model came from studies of patient populations that were primarily of European descent. African Americans, Native Americans, and other ethnic groups, who often have a higher risk for microvascular disease, are underrepresented in these studies. Thus, the potential benefit of improved glycemic control may be underestimated for these patients. It is not possible to use this model to predict which interventions will provide the greatest benefit in ethnic minorities with type 2 diabetes. However, it remains uncertain whether age at diabetes onset and degree of hyperglycemia are largely responsible for the observed higher complication rates or whether comorbid conditions, lack of access to care, or other factors may contribute.

The limited benefit of intensive treatment in patients with diabetes onset at an older age is a demonstration of the concept of competing risks, in which the benefits from treatments that occur over the long term diminish as life expectancy decreases. An important caveat to generalizing this principle is that providers too often assume that elderly patients benefit less from all life-extending treatments or that treatment benefits in the elderly are somehow “worth” less than those in younger patients [68]. Although it is clear that patients with late-onset diabetes will benefit less from intensified glycemic control, there are other instances when, because of high short-term risk and benefit, interventions in older patients have benefits that are equal to or greater than those in younger patients.

Our study generates information that may help in the management of diabetes and inform policy and policy debate on the use of health care resources. It provides better estimates, derived by using a probability model, of which patients may be most likely to benefit from intensified glycemic control and might allow patients to make more informed decisions about treatment plans and goals that are often costly and difficult to achieve. Although many questions about the best treatment strategies for type 2 diabetes remain, our study provides evidence that targeting specific, high-risk patients for intensive glycemic control is likely to provide most of the achievable benefit while at the same time limiting the number of patients who need to be treated intensively.