A population level Markov cohort simulation was employed to model the long-term disease progression of patients with type 1 diabetes. Long-term (i.e., micro and macrovascular) events for each arm were modeled via reductions in A1c levels. The baseline characteristics of this population cohort reflect those of the adult population (i.e., 25 years of age and older) in the Tamborlane et al. study on CGM [¹⁰]. All subjects were type 1 diabetes patients, with approximately 20 years since diagnosis, a mean age of 40 years, and a mean A1c level of 7.6% (+ or - 0.5%). A cycle length of one year was used for the Markov analysis, with a time horizon of 33 years, assuming a life expectancy of 73 years. The Markov model is represented in a decision analysis format (Figure 1), using TreeAge Pro 2009 (TreeAge Software, Williamstown, MA, USA). Continuous glucose monitoring with self-monitoring of blood glucose is compared to self-monitoring of blood glucose alone. All costs are in 2007 US dollars, and a discount rate of 3% was used for costs and QALYs.

There are many widely published and validated models, such as the CORE Diabetes Model, that project long-term diabetes outcomes [¹⁴,¹⁵]. However, we built a model targeted specifically towards the clinical benefit of CGM technology in a population with characteristics similar to the Tamborlane et al. adult type 1 diabetes study population [¹⁰]. In particular, Tamborlane et al. found a mean reduction in A1c of 0.5% over the trial time period for the adult patients using CGM technology [¹⁰]. The 0.5% reduction in A1c was used for the derivation of the four CGM risk reduction parameters in our model (Table 1). The level of detail for the calculation of input parameters in our model was not available in published CORE Diabetes Model studies. We used inputs and assumptions from the model built by the C.D.C. Cost-Effectiveness Group [¹⁶,¹⁷], other literature sources [¹⁸,¹⁹], and the expertise offered by our research team. The C.D.C. Cost-Effectiveness Group used similar modeling inputs and assumptions as were used in the CORE Diabetes Model (i.e., inputs derived from the Diabetes Control and Complications Trial (DCCT), the United Kingdom Prospective Diabetes Study (UKPDS), and other literature sources) [^14-17]. Therefore, the model we built was based on similar inputs and assumptions used to develop the CORE Diabetes Model, but tailored to serve the needs of our analysis. For more information on model inputs and assumptions please see Additional File 1.

In this model, all members of the population start with no complications. After this, the population can transition to one of six health states including retinopathy, nephropathy, neuropathy, Coronary Heart Disease (CHD), continue with diabetes and no complications, or death. From the five disease states, the population may then enter an additional seven disease states: nephropathy and CHD, neuropathy and CHD, retinopathy and CHD, neuropathy and nephropathy, blindness, end stage renal disease, lower extremity amputation and neuropathy, or death (transition probabilities shown in Table 1). Patients can develop a maximum of four concomitant chronic comorbidities in the Markov model.

As delineated in Table 1, transition probabilities are drawn from the best available estimates from the literature [^16-19]. Based on evidence from Klein et al. [¹⁸], the transition probabilities of going from nephropathy to CHD (0.022), neuropathy to CHD (0.029), and retinopathy to CHD (0.028) are equal to the estimates of going from CHD back to the respective microvascular disease states. The transition probability from neuropathy to nephropathy (0.097) is conditional and drawn directly from Wu et al [¹⁹]. When the population enters concomitant disease states such as neuropathy and nephropathy for example, they are limited to that state for the rest of the cycle. The transition back into each concomitant disease state is the complimentary probability based on mortality rates (available in Additional File 1).

The probability estimates just described show the progression of diabetes for those with an average A1c level of around 8%. CGM has been shown to reduce A1c levels by 0.5% in adult patients [¹⁰]. CGM exhibited its relative risk reduction for development of chronic comorbidity as a result of its reduction in A1c levels. Risk reduction parameters were drawn from two sources: the DCCT [²⁰] for microvascular complications, and a meta - analysis relating to macrovascular complications by Selvin et al [²¹].

Utility values for each disease state were taken from the EQ-5D catalogue by Sullivan et al (Table 1) [²²]. Each disease state begins with the unadjusted mean EQ-5D score from the population in MEPS 2000-2002 with diabetes mellitus, adjusted to reflect a mean age of 40 years. The utility calculation for each disease state also includes deductions for age by cycle length, and discounting by 3% [²³]. There are a total of 12 different utilities for each disease state. Incremental effectiveness is expressed in quality-adjusted life years (QALYs) gained.

Costs were derived from evidence published by the ADA [¹]. The annual mean cost of diabetes represents the per capita expenditures for people with diabetes at all age groups for hospital inpatient visits, nursing/residential facility visits, physician's office visits, emergency department (ED) trips, hospital outpatient visits, home health care, hospice care, podiatry care, insulin, diabetic supplies, oral agents, retail prescriptions, other supplies, and patient time [¹]. Lost wages served as a proxy for patient time. The ADA estimates that people with diabetes experience an additional 2.5 days absent compared to those without diabetes [¹]. The authors also estimated that the same population with diabetes on average earns $250 a day. They also estimate that the population aged 64 or less has approximately $625 of patient time per year for annual treatment of diabetes [¹]. The assumption for the population over 64 is one day of lost wages ($250). Other costs in the model include marginal annual costs for each disease state, such as blindness, end stage renal disease, lower extremity amputation and neuropathy, retinopathy, neuropathy, nephropathy, and CHD, along with the concomitant disease states. The marginal costs for each disease state were calculated using average length of stay in an inpatient hospital setting and the cost per medical event, estimated from the ADA [¹]. Costs per health state are delineated in Table 1. The concomitant disease states were estimated by summing the marginal cost for each disease state, with the exception of blindness, lower extremity amputation, and end stage renal disease (i.e., neuropathy and CHD, nephropathy and CHD, retinopathy and CHD, neuropathy and nephropathy, where each were calculated separately). While the summation assumption for marginal costs of each combination of disease states may overestimate the costs associated with having those disease states, the ADA does note their cost estimates are an underestimate of the societal cost attributable to diabetes [¹]. CGM costs were estimated from a diabetes technology and treatment purchasing website [²⁴]. Annual and initial costs are an average based on 3 systems, the Guardian Real - Time, Dexcom seven, and MiniMed Paradigm Real - Time system. The initial cost of CGM ($4,809) consists of the monitor, transmitter, two hours of patient time for education, and sensors for the first year. The annual costs ($4,189) thereafter include additional sensors per year, two hours of patient time for maintenance, and additional transmitters and batteries for the year. The initial CGM cost estimate is included in the zero cycle of the Markov model node CGM. The annual cost of CGM is then included in all disease states including no complications after cycle zero.

The all cause mortality rate was based on an average of all race categories (Non-Hispanic white, African-American, Hispanic, Native American, and Asian), and gender, from the C.D.C. Cost-Effectiveness group [¹⁶]. Increased mortality risks were drawn from the Early Treatment Diabetic Retinopathy Study (ETDRS) by Cusick et al [²⁵]. The tables for each mortality rate (neuropathy, nephropathy, CHD, LEA, and ESRD, and each concomitant disease state) are available in Additional File 1.

Probabilistic sensitivity analysis was performed using Monte Carlo simulation to evaluate the multivariate uncertainty in the model. The input parameters were varied simultaneously over specified ranges. Various probability distributions were chosen based on assumptions for each input parameter. The beta distribution was specified for the probability, utility, and risk reduction parameters. The Gamma distribution was specified for the cost parameters. The Monte Carlo simulation drew values for each input parameter and calculated expected cost and effectiveness for each arm of the model. This process was repeated 10,000 times to give a range of all expected cost and effectiveness values. Additionally, univariate sensitivity analysis was conducted to identify variables that had the largest impact on the model results. For the univariate sensitivity analysis we varied all parameters shown in Table 1 by +/- 15%. The parameters that had the largest impact on the model results are presented in a tornado diagram. The top ten variables from the tornado diagram were individually varied by 50% to estimate the effect on the model results.

The results for the base-case analysis are shown in Table 2. The mean total lifetime costs for SMBG were $470,583. The mean total lifetime costs for SMBG and CGM technology totaled $494,135, resulting in an incremental cost of $23,552. Lifetime effectiveness for SMBG was 10.289 QALYs. Lifetime effectiveness for SMBG with the addition of CGM technology was 10.812 QALYs, resulting in an incremental effectiveness of 0.523 QALYs. The incremental cost-effectiveness ratio (ICER) was $45,033 per QALY for CGM technology. Mortality was not directly reduced by CGM; it simply reduced the probability of entering disease states, thereby delaying the increased mortality from complications.

Results of the probabilistic sensitivity analysis are shown in Table 2 and Figure 2. The ranges given in Table 2 are 95% credible ranges for the expected cost and effectiveness. Figure 2 is a scatter plot of incremental cost-effectiveness pairs for the use of CGM with SMBG vs. SMBG only. The dashed diagonal line represents US$50,000 per QALY. Each dot represents one simulation. The ICER estimates in the southeast quadrant make up 10.66% of the simulations, and indicate that CGM is less costly and more effective, dominating SMBG. The rest of the simulations lie in the northeast quadrant with 36.96% below US$50,000/QALY. Results show that 48% of the observations are cost-effective for a willingness-to-pay of US$50,000 per QALY and 70% for a WTP of $100,000/QALY.

The univariate sensitivity analysis results are shown in Figure 3 as a tornado diagram, expressed in terms of net monetary benefit. Net monetary benefit is calculated by taking the difference in effectiveness and multiplying by society's willingness-to-pay, less the difference in costs. After identifying the ten variables with the largest impact on the model results, each was varied individually by 50%. The utility of diabetes with no complications, the annual cost of CHD, and the probability of going from diabetes with no complications to the CHD disease state, had the largest impact on the model results. The utility of diabetes with no complications was decreased by 50%, and the corresponding incremental effectiveness dramatically decreased, resulting in an ICER over US$300,000/QALY. When the utility of diabetes with no complications was increased by 50%, incremental effectiveness increased, decreasing the ICER to approximately US$30,000/QALY. The annual cost of CHD also had a large impact on the model results, and when decreased by 50%, the ICER was US$86,000/QALY. When the annual cost of CHD was increased by 50% the ICER was US$12,000/QALY. The probability of going from diabetes with no complications to the CHD disease state was decreased by 50%, estimating an ICER of approximately US$66,000/QALY. When the probability of entering the CHD disease state was increased by 50% the ICER was US$32,000/QALY. The other variables listed in the tornado diagram were also varied by 50%, but offered no meaningful impact on the model results (within the range of US$40,000/QALY to US$60,000/QALY).

a Beta distribution assumed

b Credible range of values from the 2.5th and 97.5th percentiles of the 10,000 second order Monte Carlo simulations

c Gamma distribution assumed for all cost parameters

d Beta distribution assumed for all risk reduction parameters; start age, years since diagnosis, and discount rate were not varied

*95% credible ranges based on the results from the 10,000 Monte Carlo simulations