Short-term clinical effects of the pioglitazone and placebo treatment regimens from PROactive were used to project long-term outcomes using a modified version of the extensively published and validated CORE Diabetes Model 1920.

The CORE Diabetes Model is a computer simulation model developed to determine the long-term health outcomes and economic consequences of interventions in diabetes. Disease progression is modelled via 15 inter-dependent semi-Markov sub-models that simulate progression of disease related complications (angina, MI, congestive heart failure, stroke, peripheral vascular disease, diabetic retinopathy, macula oedema, cataract, hypoglycaemia, ketoacidosis, lactic acidosis, nephropathy and end-stage renal disease, neuropathy, foot ulcer, amputation and non-specific mortality). Each sub-model uses time, state and diabetes type-dependent probabilities derived from published sources. Using the CORE Diabetes Model diabetes management strategies can be compared in different patient populations in a variety of realistic clinical and economic settings.

In the current study, short-term (0 – 3 years) data from PROactive was used as a basis for long-term projections using the CORE Diabetes Model adapted to include clinical input data from PROactive 21. For a more detailed account of the PROactive long-term simulation model describing the adaptation of the original CORE Diabetes Model, the transition probabilities applied and the utility values accounted for disease states interested readers are encouraged to view the publication of Valentine et al. 21.

In brief, to incorporate the PROactive clinical and adverse event endpoint data into the CORE Diabetes Model, a number of new complication sub-models were developed or modified. The overall architecture of the modified CORE Diabetes Model remained analogous to the original version, with all active sub-models running in parallel to simulate the progression of disease and the development of diabetes-related complications. The sub-models included in the final PROactive long-term simulation model were; acute coronary syndrome, percutaneous coronary intervention (PCI), coronary artery bypass graft (CABG), bypass surgery/revascularization of the leg, hospital admission for heart failure, non-serious heart failure, oedema, myocardial infarction, transient ischemic attack (TIA), stroke, photocoagulation, severe vision loss, nephropathy, peripheral vascular disease, diabetic foot, amputation, and cataract (Figure 1).

Event rates in the placebo arm were calculated directly from the annual hazard rates observed over months 0–36 of PROactive assuming constant risk 21. To calculate event rates in the pioglitazone arm hazard ratios were applied to the event rates from the placebo arm in line with the relative risk observed for each event during PROactive. To capture the statistical uncertainty in the trial data, hazard ratios in the modelling simulations were randomly sampled from a lognormal distribution generated from μ and σ values derived directly from the trial data. Constant risk was assumed for application of hazard ratios to calculate event rates in the pioglitazone arm for all endpoints with the exception of oedema, where hazard ratios were sampled from separate distributions for the first and subsequent years of the simulation to capture the decrease in risk after the first year of treatment with pioglitazone (see Additional file 1).

All-cause and cardiovascular mortality rates were derived directly from the PROactive Study for years 1–3 of the simulation, with rates subsequently doubling every 10 years 22. Event rates for following years were calculated by applying relative risk adjustments from the United Kingdom Prospective Diabetes Study (UKPDS) and Framingham data for each additional life-year gained (i.e. as the patient gets older, his/her risk of experiencing an event increases) 2324252627. In all sub-models, the occurrence of events resulted in the accounting of event costs and, where applicable, subsequent state costs as well as assignment of the appropriate disutility values.

The modified CORE Diabetes Model estimated the impact of treatment on life expectancy, quality-adjusted life expectancy (based on CODE-2 utilities), diabetes-related complications (cumulative event rates), direct medical costs and cost-effectiveness ratios over patient lifetimes, in line with specifications for health economic evaluations and under the assumption that treatment allocation continues 2829.

A simulation cohort of patients was defined with baseline demographics, and complications representative of the two treatment arms from PROactive. Long-term outcomes were calculated 1,000 times in the model using a simulated population of 1,000 patients, in order to capture the effects of random variation between individual patients. At baseline the simulation cohort had a mean age of 61.8 years, duration of diabetes 10 years, and a mean HbA1c level of 8.1% (Table 1). For the purposes of this analysis patients were assumed to remain on the same treatment regimen for the duration of the simulation (35 years or death).

The use of angiotensin converting enzyme inhibitors (ACE inhibitors) has an important influence on renal outcomes and in the simulation cohort this was set to 62.8% based on data from PROactive. Use of aspirin and statin medication has an impact on cardiovascular events, and in line with the PROactive study cohort 73.1% and 42.9% of the simulation cohort were assigned to receive these treatments respectively. Therefore, risk adjustment for the use of aspirins or statins was disabled in the analysis as the influence of these agents was already taken into account, along with the impact of ACE inhibitors, in the cardiovascular event rates taken from PROactive. It is relevant to note here that a survey of prescription drug costs for diabetic patients in Germany reported that approximately 45% of patients with type 2 diabetes were treated with ACE inhibitors and 25% with statins in 2004 30. Additional settings for patient management parameters (e.g. screening for renal disease) were set in line with the standard of care patients in the PROactive study were receiving, with all patients receiving regular screening and foot check-ups.

For the base case simulation, the clinical effects associated with the pioglitazone and placebo treatment regimens were applied as observed during PROactive. Treatment effects on HbA1c were applied separately in simulation years 1, 2 and 3, and in subsequent years, based on the findings from PROactive (for long-term projection). Changes in HbA1c and other parameters for pioglitazone and placebo regimens were applied as summarized in Table 2 and based on the statistical report compiled by Nottingham Clinical Research Limited for the PROactive study group. After year three of PROactive the between treatment group difference in HbA1c was -0.5% for pioglitazone versus placebo. The long-term progression of all of these clinical parameters subsequently followed the patterns previously described by Palmer et al. in their description of the CORE Diabetes Model 19.

Direct medical costs were expressed in 2005 Euro (€) and taken from Germany specific sources where possible (Table 3). Unit costs retrieved from published sources that were not expressed in 2005 € values were inflated using indices from the German Federal Statistics Office http://www.destatis.de. Direct medical costs were calculated as the sum of treatment costs (based on data from PROactive), patient management costs and the cost of complications. The annual costs of study medication were accounted based on data from PROactive and a mean annual cost of €736.15 per patient for pioglitazone and zero for placebo. Calculations were based on daily costs of €1.29, €1.93 and €2.31 for treatment with 15 mg, 30 mg and 45 mg pioglitazone and zero cost for placebo.

For the base case analysis, health state utilities for the events reported in PROactive were derived wherever possible from the CODE-2 study 31. Where PROactive events were not taken into consideration in the CODE-2 formula, no substitute values were used (as this could have produced erroneous findings). All other quality of life utilities used in the base case simulation have been described previously by Palmer et al. 19. Sensitivity analyses were performed to investigate the impact of including additional quality of life disutilities on the base case findings. A list of all PROactive model specific utilities incorporated into the model is given in Additional file 2.

Discounting was applied to costs, life expectancy and quality-adjusted life expectancy at an annual rate of 5.0% in line with current recommendations for the German setting 32. To capture all relevant long-term complications, a lifetime horizon of 35 years was used in the analysis. The analysis was conducted from a third-party healthcare payer perspective in Germany.

Sensitivity analyses were performed to investigate the impact of key input variables on the projected outcomes of the analysis. Typically these were univariate sensitivity analyses where the base case parameters were left unchanged except for the parameters specifically called out in the descriptions here. In the base case, changes in HbA1c were simulated using data from PROactive in the first three years and thereafter (years 4+) an annual increase in HbA1c of 0.15% was assumed in line with observations from UKPDS 33. To investigate the impact of potential benefits with respect to β-cell function a sensitivity analysis was performed where the annual HbA1c creep of 0.15% was reduced to 0.1% for years 4+. There were very few hypoglycaemic events observed in the PROactive trial but to investigate the potential impact of these on the overall outcomes a sensitivity analysis was performed where all events were associated with a cost of €2,555 in line with data from Holstein et al. 34. The impact of time horizon was investigated by varying this between 10 and 35 years. Similarly, the impact of discount rates was assessed by varying this between zero and 10% in line with recommendations of the Hanover consensus 29. Treatment related assumptions were investigated by varying the impact of pioglitazone treatment on HbA1c such that both pioglitazone and placebo arms simulated a change in HbA1c of -0.3% corresponding to the PROactive placebo arm. The influence of risk adjustment for age on the event rates taken from PROactive was investigated by performing an analysis where no risk adjustment for age was applied during the simulation (risk adjustment factors all set to 1). In order to examine the effect of uncertainty on costs accounted in the model, sensitivity analyses were conducted varying the unit cost of pioglitazone treatment by +/- 20% and then also by varying all other costs (complication and management) by the same margins.

In the base case scenario quality of life utilities were taken from CODE-2, and for many events such as oedema and heart failure this resulted in no disutilities being applied due to the absence of data. To investigate the influence of including quality of life disutilities not included in the CODE-2 formula a number of sensitivity analyses were run to include additional quality of life disutilities as previously described 21. In brief, this involved repeating the base case analysis using the CORE default method of quality-adjusted life expectancy estimation, applying an event disutility of -0.01 for oedema and a follow up disutility of zero (as the condition is typically short-lived), applying an event disutility to hospitalization for heart failure of -0.121 and a follow up disutility of -0.181, based on the UKPDS 35, or applying an event disutility of -0.0605 for non-serious heart failure and a follow up disutility of zero (due to the absence of published data). Utility values for oedema and non-serious heart failure were based on assumptions.

A "worst case" scenario was also included in the sensitivity analysis where event disutilities for oedema (as described), hospitalization for heart failure (by applying an event disutility of -0.121 and a follow up disutility of -0.181) and revascularization of the leg (by applying an event disutility of -0.059 and a follow up disutility of zero, assuming a disutility comparable to that reported for CABG) from the post-hoc study of Lescol Intervention Prevention Study were applied 36.

The health economic analysis was performed using a non-parametric bootstrapping approach in which the progression of diabetes was simulated in 1,000 patients through the model 1,000 times to calculate the mean and standard deviation of costs, life expectancy and quality-adjusted life expectancy using second order Monte Carlo simulation 37. Mean results from each of the 1,000 iterations were used to create a scatter plot, comparing the differences in costs and outcomes for pioglitazone and placebo treatment regimens. These values were in turn used to generate a cost-effectiveness acceptability curve over a range of willingness to pay values in the German setting.

The current health-economic analysis indicated that based on clinical findings for PROactive and long-term projections with a modified version of the CORE Diabetes Model, treatment with pioglitazone was associated with improvements in life expectancy and quality-adjusted life expectancy compared to placebo. Mean life expectancy (discounted by 5.0% per annum) increased by 0.172 years with pioglitazone and after adjustment for quality of life an improvement of 0.120 quality-adjusted life years (QALYs) was projected versus placebo (Table 4). When no discounting was applied to the long-term outcomes, mean life expectancy in the pioglitazone treatment arm was 0.406 years longer than in the placebo arm.

Over patient lifetimes, treatment with pioglitazone was associated with higher direct medical costs than the placebo regimen (Table 4). Direct costs increased by €1,599 with pioglitazone compared to placebo. This increase was largely due to increased treatment costs (€27,210 versus €24,348), whilst complication-related costs decreased by €1,263 (Table 5). Treatment with pioglitazone was associated with a reduced cost for stroke events by €1,487 and an increase in costs by €274 for hospitalization for heart failure compared to placebo (although mortality rates from heart failure did not differ between groups).

Estimation of incremental cost-effectiveness ratios (ICER) for pioglitazone versus placebo treatment produced values of €9,281 per life year gained and, taking quality of life into account, €13,294 per QALY gained (Table 4). The values from the 1,000 means (each from 1,000 patients) of incremental costs and incremental effectiveness (in terms of quality-adjusted life expectancy) were used to generate a scatter plot on the cost-effectiveness plane. This analysis shows that the majority of points were in the upper right quadrant of the plane, indicating increased effectiveness and increased costs associated with pioglitazone treatment over placebo. These values were then used to create a cost-effectiveness acceptability curve by assessing what proportion of values fell below set willingness to pay values (Figure 2). The analysis demonstrated that, with a willingness to pay of €50,000 per QALY in the German setting, there was a high probability (78.2%) that pioglitazone would be cost-effective (Figure 3).

Sensitivity analysis showed that the results were most sensitive to variation in the time horizon and assumptions on the duration of the benefits of pioglitazone treatment observed in the trial, whilst the exclusion of improved β-cell function from the base case scenario was shown to bias against pioglitazone (see Additional file 3).