Clinical Outcomes and Cost-Effectiveness of Strategies for Managing People at High Risk for Diabetes

Context A previous Markov modelbased analysis estimated that use of the Diabetes Prevention Program diet and exercise intervention to forestall diabetes in high-risk people would be cost-effective from a societal perspective. Contribution Using a validated model designed to be more complete and realistic than previous models, the authors estimated that the intervention would cost society about $62600 per quality-adjusted life-year saved. It would be cost-saving if the annual cost of the intervention decreased from $672 to $100. Implications This model suggests that the Diabetes Prevention Program intervention costs more per quality-adjusted life-year saved than previously estimated, and health plans and insurers may consider it too expensive to cover. The Editors Recent randomized, controlled studies have shown that diabetes can be prevented or delayed in high-risk individuals by intensive lifestyle modification programs (1, 2) or glucose-lowering drugs (2-4). For example, in the Diabetes Prevention Program (DPP), the relative reductions in the 2.8-year incidence of diabetes were 58% in the lifestyle modification group and 31% in the metformin group (2). This raises hopes of substantially reducing the morbidity, mortality, and cost of this important disease. However, the trial was too short to observe the effects on microvascular or macrovascular outcomes, and the programs cost several hundred dollars a year (5). These findings generate obvious questions: What are the long-term effects of trying to prevent diabetes in high-risk people? Does lifestyle modification truly prevent or just postpone diabetes? Is such a prevention program cost-effective? What is the best strategy? A previous analysis has suggested that lifestyle modification would be cost-effective over 75 years from a societal perspective (6). We used a more thorough, clinically realistic, and independently validated model to estimate the short- and intermediate-term health and economic effects of different prevention programs for high-risk individuals and health plans, as well as for society. Methods We conducted the analysis by using the Archimedes model, which has been described elsewhere (7-9). Briefly, it is a simulation model written at a relatively high level of anatomic, physiologic, clinical, and administrative detail. It uses object-oriented programming to create in the model objects that correspond to objects in reality, one-to-one. Among the hundreds of objects are people, pancreases, cells, plasma glucose levels, coronary arteries, plaque, chest pain, emergency departments, electrocardiograms, aspirin, and angioplasties. Helpful analogies might be a flight simulator (in which the objects include the plane and its wings, airports, runways, buildings, and the wind), or the SimCity computer game. In the Archimedes model, each individual is simulated down to the level of hepatic glucose production, insulin resistance, -cell fatigue, and similar biological variables. The core of the model is a set of differential equations that represent the anatomy and physiology pertinent to diseases and their complications. Currently, the model includes diabetes, congestive heart failure, coronary artery disease, stroke, hypertension, and asthma in a single integrated model. The structure and equations of the model pertinent to diabetes and its complications are described elsewhere (8, 9). The Appendix and a technical report available through our Web site (10) describe additional aspects of the model and its validations that are pertinent to this analysis. Calculations are performed by using a distributed computing network. Clinical Events The model includes the biological variables and outcomes relevant to diabetes and its complications. Examples are basal hepatic glucose production; insulin amount; insulin resistance; fasting plasma glucose; hemoglobin A1c (HbA1c); 2-hour oral glucose tolerance; high-density lipoprotein (HDL) cholesterol, low-density lipoprotein (LDL) cholesterol, and total cholesterol; triglycerides; systolic and diastolic blood pressures and their determinants (for example, cardiac output, arterial compliance, peripheral resistance); weight and body mass index (BMI); stenosis of coronary arteries; retinopathy (assessed by the Early Treatment of Diabetic Retinopathy scale); urine protein; creatinine; peripheral neuropathy; foot ulcers of varying degrees of severity; and amputations. The use of differential equations preserves the continuous nature of biological variables as well as the interactions between them. Clinical outcomes are defined in terms of the underlying variables, as occurs in reality. For example, a person is said to have diabetes if his or her fasting plasma glucose level exceeds 6.9375 mmol/L (125 mg/dL) or results on a 2-hour oral glucose tolerance test exceed 11.0445 mmol/L (199 mg/dL). This enables the model to incorporate different definitions and changes in definitions. The model is continuous: Biological variables are changing and interacting continuously, the natural histories and severity of conditions progress smoothly, any clinical event can occur at any time, and the timing of events is as condensed or drawn out as occurs in reality. The model also includes a detailed representation of the processes and logistics of clinical care and their related costs. Interventions, both to prevent diabetes and to manage it when it occurs, are modeled at the level of the underlying biology. Pertinent to this analysis is that in the model, diet and exercise reduce weight (2); reduce blood pressure (11); improve LDL cholesterol, HDL cholesterol, and total cholesterol levels (12); and decrease fasting plasma glucose levels (2). The effects of metformin in the model are to reduce fasting plasma glucose and 2-hour oral glucose tolerance test results (2) (by reducing basal hepatic glucose production), decrease LDL cholesterol and triglyceride levels (13), and retard weight gain. Data used to build the model were derived from basic physiologic studies, surveys, epidemiologic studies, and clinical trials using methods described in the technical report (10). Every variable in the model is estimated from at least 1 empirical source; no variables are simply assumed. We identified specific sources by searching MEDLINE from 1970 to 28 February 2005 and by consulting textbooks and clinical experts. Because the model includes scores of continuously valued, interacting variables, it does not have simplified states, transitions, or events at discrete time intervals that can be tabulated, as is commonly done for a Markov-type model. The equations themselves are in the technical report (10). For nonmathematical readers, we have calculated annualized rates of change of representative biological variables and annualized rates of occurrence of representative clinical events, and compared them with rates for comparable events observed in epidemiologic studies and clinical trials. The Appendix reports those results. Costs The DPP measured the direct medical costs of delivering the lifestyle and metformin interventions (for example, personnel, health education materials, medications, and laboratory tests). Compared with the placebo group's costs, costs in the lifestyle group were $1356 more per person in the first year, with approximately $672 in annual costs thereafter; for the metformin group, costs were $977 in the first year and averaged $742 per year thereafter (5). Following the completion of the DPP, metformin became generic. When this is considered, the cost of the metformin program is reduced to about $780 for 3 years, or about $260 a year. In the DPP study, costs apply to the year 2000. To calculate the routine costs of providing health care to high-risk people before they develop diabetes, as well as to people with diabetes and its complications, the model includes a detailed mathematical representation of a health care system, including such elements as facilities, personnel, tests and treatments, protocols, and provider behaviors. For the base-case analysis, we obtained itemized costs from Kaiser Permanente, a nonprofit, group-practice, integrated managed care organization that provides comprehensive care (with no deductibles or copayments). The facilities, personnel, protocols, and costs in the model are based on that organization's records, at the level of detail at which actual accounts are kept (for example, 37 different kinds of office visits). The model calculates costs by keeping track of the occurrence of every event that has cost implications and adding them up. The costs assigned to any event or item were calculated by Kaiser Permanente's cost-accounting department using micro-costing methods (14), and they represent the real costs to the organization, not charges, reimbursements, or diagnosis-related groups. Because costs vary from setting to setting, the implications of different cost structures are examined in the sensitivity analysis. Calculation of costs applies to the year 2000. Indirect costs, such as lost time from work and decreased productivity, are included in the cost-effectiveness analysis through the Quality of Well-Being Index (14). We calculated the effects of lifestyle and metformin interventions on quality of life. For people who do not yet have diabetes, we used utility weights reported for the participants of the DPP study (15). For people who have diabetes and its complications, we used the results of a published survey by Coffey and colleagues (16). Both surveys used the Quality of Well-Being Index. The decrements in quality of life were assumed to be additive for people who have 2 or more complications, with a limit that quality of life could not be less than 0. Use of an additive rule biases the calculation of cost/quality-adjusted life-year (QALY) in favor of a prevention program, making the program appear more cost-effective than would occur if a multiplicative model were used. We discuss the pot