Screening for Type 2 Diabetes Mellitus: A Cost-Effectiveness Analysis

No randomized, controlled trial of screening for diabetes has been conducted. In the absence of direct evidence, cost-effectiveness models may provide guidance about preferred screening strategies.

In the absence of direct evidence, researchers have applied mathematical models of diabetes progression to the issue of screening. One analysis found that the cost per quality-adjusted life-year (QALY) gained by universal diabetes screening was lower for younger than for older people: $13 376 at age 25 to 34 years, increasing to $116 908 at age 55 to 64 years (16). This conclusion followed from the model's focus on the provision of glycemic control after screening to prevent microvascular complications. The analysis did not consider treatments to reduce the risks for complications of cardiovascular disease (CVD).
More recent research suggests that the benefits of CVD risk reduction may be substantial for people with diabetes. The Hypertension Optimal Treatment (HOT) trial found that the optimal blood pressure target is lower for people with hypertension and diabetes than for people with hypertension without diabetes (14). Other research supports the finding that intensive control of hypertension is beneficial among people with diabetes (15,(17)(18)(19). Be-cause the benefit may be greater for older people (at greater risk for CVD), the conclusion of the previous analysis, that diabetes screening is most cost-effective among younger people, needs to be reconsidered.
We performed a new cost-effectiveness analysis to compare universal diabetes screening (universal screening) and diabetes screening targeted to patients with hypertension (targeted screening). When an updated version of the model used in the previous analysis that includes benefits from intensive treatment of hypertension was applied, we estimated the incremental cost-effectiveness of these 2 strategies for people in different age groups. Our analysis considers a one-time opportunistic screening for men and women of all races and ethnicities.

The Model
We used a Markov model of diabetes disease progression to simulate lifetime diabetes-related health care costs and QALYs for people with diabetes (Appendix Figure 1, available at www.annals.org). Demographic characteristics of the simulated cohort are based on 1997 population estimates projected from the 1990 U.S. Census and data on the distribution of people with diabetes by hypertension, cholesterol level, and smoking status (20). As people progress through the simulation model from the onset of diabetes to death, they can develop 5 types of complications: nephropathy, neuropathy, retinopathy, coronary heart disease (CHD), and stroke. People can die of some of these complications or of other causes. The model includes transition probabilities between disease stages on each of the 5 complication paths. The basic model structure has been described previously (16,21). Key model parameters are presented in the Appendix Tables 1 to 10 (available at www.annals.org).
To incorporate screening into the model, we first added a screening module in which some patients with diabetes are identified earlier than they would usually have been in the absence of screening. Second, we made assumptions about the transition probabilities between disease stages from the onset of diabetes to the time of usual clinical diagnosis of diabetes on the basis of the knowledge that progression is relatively slow during this period (15). After clinical diagnosis, disease progression depends on the number of years after normal diagnosis. *Diabetes onset; †Diabetes diagnosis by screening; ‡Diabetes diagnosed clinically.

Context
In 2003, the U.S. Preventive Services Task Force recommended screening for type 2 diabetes in adults with hypertension or hyperlipidemia. The economic implications of this recommendation are unclear.

Contribution
Diabetes screening for 55-year-old hypertensive persons would cost the U.S. health care system $34 375 per quality-adjusted life-year gained. Expanding screening to all adults regardless of the presence of hypertension would cost an additional $360 966 per quality-adjusted life-year.

Implications
The cost-effectiveness of targeting diabetes screening to hypertensive adults older than 55 years of age is similar to the cost-effectiveness of many accepted health care interventions. Universal diabetes screening is far more costly.

-The Editors
Article Screening for Type 2 Diabetes Mellitus: A Cost-Effectiveness Analysis Screening allows for earlier diagnosis, which in turn allows for earlier treatment interventions, such as intensive glycemic control and intensive hypertension control. These interventions decrease the transition probabilities, thereby delaying or preventing progression to diabetes complications.
Costs are incurred for screening and diagnostic testing; standard glycemic control and, if the person is hypertensive, standard hypertension control; interventions (intensive glycemic control and, if the person is hypertensive, intensive hypertension control); and complications over the remaining lifetime of each person with diabetes. The sum of these costs and the model's estimate of the expected QALYs for each screening strategy are used to calculate the incremental cost-effectiveness ratio of screening relative to no screening. We discounted future costs and QALYs at a 3% annual rate. Costs are measured in 1997 U.S. dollars.

Interventions
We assumed that, in the absence of screening, diabetes would be diagnosed 10 years after its onset (15). With one-time opportunistic screening, diabetes would be diagnosed on average 5 years after onset and therefore patients would begin treatment 5 years earlier. After diabetes diagnosis, all patients are treated with intensive glycemic control and, if they have hypertension, with intensive hypertension control.
With targeted screening, only people with hypertension are screened. Those who screen positive and receive a diagnosis of diabetes begin intensive glycemic control and intensive hypertensive control 5 years earlier than they would in the absence of screening. With universal screening, all people, regardless of hypertension status, are screened. Those who screen positive and receive a diagnosis of diabetes begin intensive glycemic control 5 years earlier than in the absence of screening and begin intensive hypertension control 5 years earlier if they have hypertension.
We defined hypertension as a blood pressure of 140/90 mm Hg or higher. We assumed that 19% of people age 25 to 44 years, 47% of people age 45 to 64 years, and 60% of people age 65 to 74 years had hypertension and therefore were included in targeted screening (20).
Treatment of hypertension is modeled as standard (with a target diastolic blood pressure of 90 mm Hg) or intensive (with a target diastolic blood pressure of 80 mm Hg), as in the HOT trial (14). All persons with hypertension receive standard hypertension treatment until they receive a diagnosis of diabetes, after which they receive intensive hypertension treatment. The incremental cost of intensive hypertension control relative to standard control is $149 per year.
In the HOT trial, the relative risk reduction for CHD events (fatal and nonfatal myocardial infarction) was 51%, and the relative risk reduction for stroke was about 30%. Although neither of these separate relative risk reductions was statistically significant, the relative risk reduction (51%) for the aggregate outcome of major CVD events was statistically significant (P ϭ 0.005). We initially modeled the relative risk reduction for CHD events for inten- sive hypertension control to be 51%, with no risk reduction for stroke. We conducted a sensitivity analysis that included a 30% relative risk reduction for stroke on the basis of other studies showing that intensive hypertension control reduces risk among people with diabetes (17, 19). Model estimates of the effects of glycemic control are based on the United Kingdom Prospective Diabetes Study (UKPDS), a 10-year randomized, controlled trial of intensive versus conventional glycemic control (12). On the basis of the UKPDS, the reduction in hemoglobin A 1c from intensive glycemic treatment is modeled as slowing the progression of microvascular complications (12). The incremental cost of intensive glycemic control (relative to standard control) ranges from $900 to $1100 per year, depending on the number of years since diagnosis.

Screening and Diagnostic Tests
The Figure illustrates the screening and diagnostic testing process and shows where costs are incurred.

Screening Tests
We assume a one-time opportunistic screening during a regularly scheduled physician office visit. The model as-sumes screening by a fasting capillary blood glucose (CBG) test (22) and an extra 10 minutes over the usual 15 minutes for the physician visit, incurring a cost of $24.40 per person screened. Costs for the CBG test are derived from the Medicare Clinical Diagnostic Laboratory Fee Schedule (23); physician visit costs are derived from Relative Values for Physicians (24). Table 1 shows prevalence (from the National Health and Nutrition Examination Survey III data tape), CBG sensitivity and specificity values (22) with exact data points clarified via personal communication (Rolka DB, 18 January 2002), and the number needed to screen to detect one previously undiagnosed person with diabetes by sex, hypertension status, and age.

Diagnostic Tests
All people who screen (true or false) positive receive a diagnostic test, the fasting plasma glucose test, which is repeated if the result is positive. Because 2 consecutive elevated fasting plasma glucose test results define diabetes (11), we assume that this strategy has 100% sensitivity and 100% specificity. Diagnostic testing costs $8.32 per test ($5.32 for test processing plus $3.00 for blood drawing [23]). Table 1 reports the number of diagnostic tests needed to identify one previously undiagnosed diabetes case.

Analyses
Diabetes complications, life-years, and QALYs are calculated for each true case of undiagnosed diabetes in the given population. We calculated change in life-years, change in QALYs, and change in costs for diabetes-related care for people with diabetes, as well as costs for screening per person screened. Future medical costs are not calculated for those without diabetes because their care does not change with screening. However, the analysis does include the cost of screening them. "Base-case" analyses are performed by using the model's standard parameter values (Appendix Tables 1 to 10, available at www.annals.org).
To examine the variability of the cost-effectiveness ratios associated with screening, we conducted one-way sensitivity analyses for people screened at age 55 years to investigate the effect of key parameter values and assumptions. We also conducted a probabilistic sensitivity analysis in which 129 critical parameters were simultaneously varied over probability distributions on the basis of published 95% CIs or other reasonable ranges. We used the logistic normal distribution for most parameters (25) but used uniform and triangular distributions when appropriate (Appendix Tables 1 to 10, available at www.annals .org). We computed cost-effectiveness results for each of 1000 iterations for both targeted and universal screening of people age 55 years by using @Risk software (Palisade Corp., Newfield, New York) and examined the distribution of cost-effectiveness ratios across iterations.

Role of the Funding Sources
This study was supported by the Agency for Healthcare Research and Quality. Development of the cost-effectiveness model was supported by the Centers for Disease Control and Prevention. Staff of the Agency for Healthcare Research and Quality reviewed the study and provided comments on drafts of the manuscript. Staff of the Centers for Disease Control and Prevention participated in the development of the model and contributed to the manuscript. The authors were responsible for deciding to submit the manuscript for publication. Table 2 shows the cost-effectiveness analysis comparing targeted screening with no screening for people of different ages. Diabetes complication incidence, life-years, and QALYs are reported for each case of diabetes in the population screened. Change in life-years, change in QALYs, and costs of diabetes-related care for those with diabetes, as well as costs of screening, are reported per person eligible for screening. Compared with no screening, targeted screening leads to earlier initiation of intensive glycemic and hypertension treatment and a longer lifetime. It also increases costs. The increase in total incremental costs per person screened is somewhat greater for those who are younger than for those who are older. Incremental QALYs for persons with diabetes generally increase with age, primarily because of a reduction in CHD incidence. The cost-effectiveness ratios for targeted screening are lower in older people.

Universal Screening
Compared with no screening, universal screening increases lifetime costs at all ages ( Table 3). The increased costs are attributable primarily to increased treatment and intervention (including earlier intensive glycemic and hypertension control) for those who are identified through screening. The incremental total costs increase slightly from $331 per person eligible for screening at age 35 years to $479 per person eligible at age 55 years, before declining to $92 per person eligible at age 75 years.
Universal screening also adds QALYs over the lifetime of previously unidentified people with diabetes. The incremental cost-effectiveness ratios for universal screening compared with no screening are generally high and decrease with age.

Universal versus Targeted Screening
The cost-effectiveness ratios in Tables 2 and 3 show that targeted screening is more cost-effective than universal screening at every age when each alternative is compared with no screening. This finding suggests that policymakers would want to adopt targeted screening before universal screening. Then, the next relevant question is, given targeted screening, how cost-effective is it to move to universal screening by adding screening of people without hypertension to the people with hypertension already included in targeted screening? Table 4 shows the cost-effectiveness ratios for targeted versus no screening and for universal versus targeted screening. Relative to targeted screening, universal screening has very high cost-effectiveness ratios, which increase with age. This implies that screening people without hypertension is much less cost-effective than screening those with hypertension.

Sensitivity Analyses
We performed sensitivity analyses for 55-year-old persons ( Table 5); the same pattern of results holds for other ages. In the base-case analysis, the cost-effectiveness ratio In the base-case analysis, people receive intensive glycemic control after receiving a diagnosis of diabetes. Intensive glycemic control is expensive, costing $900 to $1100 per year more than standard glycemic control. In a sensitivity analysis, we assumed that people 55 years of age at screening receive lifetime standard glycemic control after diagnosis, when standard control is based on the conventional treatment arm of the UKPDS (12). The incremental cost-effectiveness ratios for both targeted and universal screening are cut in half. In another sensitivity analysis, we assumed that those screened with diabetes incurred no extra cost for intensive glycemic control during their first 5 years of treatment. This reduced the cost-effectiveness ratio for targeted screening by more than 50%, even more than the reduction associated with lifetime standard glycemic control.
If screening were to lead to diagnosis 2 or 8 years earlier than no screening (that is, 8 or 2 years after onset), the incremental cost-effectiveness ratios would be modestly different from what they are in the base-case analysis, in which screening leads to diagnosis 5 years earlier.
We found that if the sensitivity and specificity of the CBG screen test were based on values associated with random (Ͻ8 hour postprandial time) rather than fasting (Ն8 hours postprandial time) testing, the incremental costeffectiveness ratios would be only slightly higher (Ͻ1%).
The base-case analysis assumed that intensive hypertension control reduces the relative risk for CHD by 51% relative to standard hypertension control on the basis of HOT trial findings. If medication adherence is lower or if the effects of intensive hypertension control are more moderate than they were in the HOT trial, resulting in only a  25% risk reduction, the incremental cost-effectiveness ratios would increase substantially for both targeted ($119 262) and universal ($411 623) screening. As expected, the cost-effectiveness of screening is highly sensitive to the effects of intensive hypertension control. Previous research suggests that intensive hypertension control reduces the risk for stroke (17,19). In a sensitivity analysis, we assumed that intensive hypertension control leads to a 30% relative risk reduction for stroke (the not statistically significant relative risk reduction for stroke reported for the HOT trial), in addition to the risk reduction for CHD. The incremental cost-effectiveness ratios decline modestly.
The prevalence of undiagnosed diabetes may have changed since the National Health and Nutrition Examination Survey III. We reduced and increased all prevalence values by 1 SD; these analyses produced only negligible differences from the base-case cost-effectiveness ratios.
We prepared histograms of cost-effectiveness ratios re-sulting from the probabilistic sensitivity analyses (Appendix Figure 2, available at www.annals.org). Targeted screening analysis resulted in cost-effectiveness ratios with a median of $34 229 per QALY. Ninety-five percent of costeffectiveness ratios were between $21 594 and $76 099 per QALY. The universal screening analysis resulted in a median cost-effectiveness ratio of $371 324 per QALY when compared with targeted screening. Ninety-five percent of cost-effectiveness ratios were between $275 518 and $541 216 per QALY.

DISCUSSION
We found that, at every age, diabetes screening targeted to people with hypertension is more cost-effective than universal screening. In addition, we found that, taking into consideration a reduction in CHD events from earlier treatment of hypertension, both universal and targeted screening are more cost-effective for people at 55, 65, and 75 years of age than for people at 35 and 45 years of *All incremental cost-effectiveness ratios are calculated for persons age 50 years at diabetes onset and 55 years at screening. CHD ϭ coronary heart disease events; NHANES III ϭ National Health and Nutrition Examination Survey III; QALY ϭ quality-adjusted life-year. † Sensitivity and specificity of capillary blood glucose test screening is based on test results Յ120 mg/dL and both Ն8 and Ͻ8 hours postprandial time, as reported in Figure  2 of Rolka 2001 (22).
age. The most cost-effective approach to one-time diabetes screening is to target people with hypertension between ages 55 and 75 years. In this analysis, the benefit of screening comes predominantly from reducing CHD events by intensive control of hypertension rather than from reducing microvascular complications, such as end-stage renal disease or blindness, by intensive glycemic control. Among people at low risk for CHD events (for example, people in their thirties), the benefit of screening derives predominantly from decreasing end-stage renal disease, but it must be purchased at the high cost of intensive glycemic control. Among people at higher risk for CHD events (for example, people in their fifties and sixties), the benefit of intensive control of hypertension is greater and can be purchased less expensively. The benefits of intensive control of hypertension are also realized sooner than are the benefits of intensive glycemic control (15).
Our findings differ dramatically from those of a previous cost-effectiveness analysis (16). Our model modifies the previous model in several ways. First, we allow people with hypertension and diabetes to receive intensive hypertension control. Second, intensive glycemic control produces smaller reductions in diabetes complications in our model. Our assumptions about risk reduction from intensive glycemic control are based on UKPDS results (12) that were not available at the time of the previous analysis. Because intensive glycemic control leads to smaller effects on diabetes complications in our model, cost-effectiveness ratios for universal screening are higher than those in the previous report. Third, the earlier model assumed that people with diabetes would receive standard glycemic control after diagnosis. In our analysis, we assumed that people with diabetes would receive intensive glycemic control after diagnosis. Our sensitivity analyses show that cost-effectiveness ratios are substantially higher with intensive glycemic control than with standard control; the previous model produced similar results.
Our findings are consistent with modeling studies showing that people with diabetes are at highest risk for eventually developing microvascular complications if they are relatively young or have highly elevated glycemic levels (26,27). People with diabetes identified by screening usually have mildly to moderately elevated glycemic levels; intensive glycemic control to reduce hyperglycemia may be less beneficial for these people than for those with higher glycemic levels (27). Our findings also are consistent with studies showing that much of the cost and burden of diabetes is attributable to CVD complications, outcomes affected by intensive hypertension control (21, 28 -33).
Our conclusions could change if future research provides better and different evidence on model parameters. If, for example, intensive glycemic control during the preclinical phase of diabetes was shown to have a large effect on subsequent diabetes complications, then all of the costeffectiveness ratios would become more favorable. The HOT trial was a subgroup analysis; if other research shows that treatment of hypertension among people with diabetes should not differ from treatment of those without diabetes, our cost-effectiveness ratios for targeted screening would be too favorable. Similarly, if poor adherence to antihypertensive medications reduces the effectiveness of intensive hypertension treatment, the cost-effectiveness ratios will be less favorable. Further evidence can be incorporated within the model by changing model parameters.
We did not consider screening people without hypertension but with other CHD risk factors for diabetes, such as dyslipidemia or tobacco use. Compared with evidence on treatment for hypertension, there is less evidence that treatment for these risk factors should be different in people with and without diabetes (15). If future research shows that knowing a patient has diabetes affects treatment for lipid and tobacco disorders, then our analysis would need amending. People with dyslipidemia whose cardiovascular risk crosses a lipid treatment threshold with the diagnosis of diabetes might especially benefit from earlier diabetes diagnosis and earlier lipid treatment. Future models could examine the cost-effectiveness of diabetes screening for people in this group.
Our results do not contradict other analyses of the beneficial effects or cost-effectiveness of intensive glycemic control or intensive hypertension control after clinical diagnosis (21,29,30). This issue is distinct from the issue of screening. For screening, we assumed that everyone would receive intensive glycemic control and intensive hypertension control after diagnosis. The screening comparison is between starting these treatments a few years earlier and starting them after clinical diagnosis.
We did not examine the cost-effectiveness of screening to detect and treat impaired glucose tolerance or impaired fasting glucose levels. Although new research shows that intensive treatment can reduce the development of diabetes (34,35), cost-effectiveness models examining this question will need to make assumptions about the effect of reductions in diabetes incidence on various diabetes complications. We also did not examine the effect of periodic, rather than one-time, screening. For longer time intervals between screenings, the cost-effectiveness would be similar to one-time screening. For shorter intervals, the cost-effectiveness ratios would be higher.
As we had no randomized, controlled trial of screening for diabetes, we extrapolated much of the input data on various benefits of screening from studies of people whose diabetes was detected clinically. The longest follow-up is 10 years (12).
The study's strengths are the model's use of the most recent and highest-quality data on benefits and costs and our ability to carry out several sensitivity analyses, all of which gave similar results. Unlike researchers using previous models, we could model the macrovascular benefits of screening.
This study has important implications for screening for diabetes. Although universal screening achieves greater overall benefit than targeted screening, the cost of the additional benefit is high. A more efficient strategy is targeted screening of people with hypertension between the ages of 55 and 75 years, with intensive hypertension control for people detected with diabetes. This strategy provides most of the benefits of universal screening at much less cost. Other Time from diabetes onset to screening, y 5 Unif(4.00, 6.00) Assumed Time from onset to diabetes diagnosis, y 10 Unif(9, 12) 15 * Logn(a,b) ϭ lognormal distribution with mean a, lower bound of 95 CI b; Triang(a, b, c) ϭ triangular distribution with minimum a, mode b, maximum c; Unif(a,b) ϭ uniform distribution with minimum a, maximum b. † These values were applied in all model runs unless otherwise specified (in one-way and probabilistic sensitivity analyses). ‡ The distributions from which parameter values were randomly sampled in the probabilistic sensitivity analyses. The ranges for parameters without published variability data followed these guidelines: screening sensitivity and specificity values vary by Ϯ2%; costs more than $300 vary by Ϯ15%; costs less than $300, time durations, transition probabilities, hazard rates, and quality-of-life values vary by Ϯ25%; and the discounting factor varies from 2% to 5%. Relevant limits were applied to all ranges (e.g., quality of life and probabilities must be between 0 and 1).