• E-mail
• Print
• Comment
• Font Size
• Digg
• del.icio.us
• Discuss article

The CORE Diabetes Model: Projecting Long-Term Clinical Outcomes, Costs and Cost-Effectiveness of Interventions in Diabetes Mellitus (Types 1 and 2) to Support Clinical and Reimbursement Decision- Making

Posted on: Wednesday, 10 November 2004, 03:00 CST

Key words: Complications - Cost-effectiveness - Costs - Diabetes - Health economics - Modelling

SUMMARY

Objectives: We have developed an Internet-based, interactive computer model to determine the longterm health outcomes and economic consequences of implementing different treatment policies or interventions in type 1 and type 2 diabetes mellitus. The model projects outcomes for populations, taking into account baseline cohort characteristics and past history of complications, current and future diabetes management and concomitant medications, screening strategies and changes in physiological parameters over time. The development of complications, life expectancy, quality- adjusted life expectancy and total costs within populations can be calculated.

Methods: The model is based on a series of sub-models that simulate important complications of diabetes (cardiovascular disease, eye disease, hypoglycaemia, nephropathy, neuropathy, foot ulcer, amputation, stroke, ketoacidosis, lactic acidosis and mortality). Each sub-model is a Markov model using Monte Carlo simulation incorporating time, state, time-in state, and diabetes type-dependent probabilities derived from published sources. Analyses can be performed on cohorts with type 1 or type 2 diabetes. Cohorts, defined in terms of age, gender, baseline risk factors and pre-existing complications, can be modified or new cohorts defined by the user. Economic and clinical data in the model can be edited, thus ensuring adaptability by allowing the inclusion of new data as they become available; creation of country- or provider-specific versions of the model; and allowing the investigation of new hypotheses.

Conclusions: The CORE Diabetes Model allows the calculation of long-term outcomes, based on the best data currently available. Diabetes management strategies can be compared in different patient populations in a variety of realistic clinical settings, allowing the identification of efficient diabetes management strategies.

Introduction

For a number of years, mathematical modelling has been widely used to assess long-term clinical outcomes and provide economic evaluations of healthcare interventions. It provides an alternative to large-scale long-term clinical trials which would, as well as proving logistically difficult, be very expensive and take years to produce results. The value of modelling has become more widely accepted as evidenced by the recommendation for an economic analysis to accompany submissions on new interventions from regulatory bodies in a number of countries around the world, including Australia1, Belgium2, Canada3, Germany4, Italy5, the UK6 and the USA7,8. In addition, the recent publication of principles of good practice for decision analytical modelling in healthcare evaluations by a task force from the International Society for Pharmacoeconomic and Outcomes Research (ISPOR) provides evidence of the vital role of modelling for decisionmakers in the public and private sectors9.

At present, the direct costs of treating diabetes and its complications are huge, accounting for approximately 12% of total healthcare expenditure in the USA10. A recent report from the USA indicated that type 2 diabetes is associated with annual per patient costs of around $1,700 for men and $2,100 for women11. Perhaps more alarming is the prospect that the worldwide incidence of type 2 diabetes is increasing12-15. The majority of the costs associated with diabetes are attributed to the complications of the disease, including micro- and macrovascular complications, hypertension and atherogenic serum lipid profiles16-18. Healthcare providers need accurate models of diabetes and its complications to assess the merits of new and existing interventions competing for the same limited healthcare budgets. There is also evidence that the incidence of type 1 diabetes, which occurs in genetically predisposed persons as a consequence of the immune-mediated destruction of pancreatic islet beta-cells that secrete insulin19, is increasing in some European countries and is clinically manifest in younger patient age groups in others20-23. However, the recent development of potential preventative measures for type 1 diabetes24- 26 have introduced a fresh need for accurate modelling strategies to predict the long-term outcomes and costs associated with any such programmes.

As a result, several modelling frameworks to simulate diabetes and its complications have been published over the last five years27- 30. Previous models have simulated diabetes, including the complications of retinopathy, nephropathy, neuropathy and cardiovascular disease (CVD), and in some cases the models integrated the effects of treatment on these complications in the simulation. The CORE Diabetes Model covers the range of diabetes complications and has been designed to simulate their development and the effect of new and existing interventions on clinical and cost outcomes in a realistic and transparent way. It is an interactive computer model, designed as a policy analysis tool, which determines the long-term clinical and health economic outcomes in patients with type 1 or type 2 diabetes. The model projects outcomes for populations, taking into account baseline cohort characteristics and past history of complications, current and future diabetes management and concomitant medications, screening strategies, and changes in physiological parameters over time. The development of complications, life expectancy, quality-adjusted life expectancy and total costs within populations can be calculated. This paper documents the architecture, assumptions, features and capabilities of the CORE Diabetes Model corresponding to guidelines published by the ISPOR task force9, to provide transparency for users of the model and readers of analyses based on the model.

Methods

CORE Diabetes Model Overview

The model is a multilayer internet application linked to a mathematical calculation model and structured query language (SQL) database sited on a central server. Online access to the CORE Diabetes Model is available under licence from CORE - Center for Outcomes Research, the developers of the model. The structure is based on four separate elements: the user interface, the input databases, the data processor, and the output databases (Figure 1).

User Interface

The user interface is the means by which users can enter customised settings to specifically define an analysis and through which the results of any given analysis are presented (Figure 1). It was designed as a user-friendly, menu-driven interface programmed in hypertext markup language (HTML) and allows the model to be available to multiple users around the world via the Internet, regardless of their own computer operating system and/or settings. Via the interface, users can define an analysis in terms of the cohort characteristics (or use a pre-existing cohort saved in the CORE Diabetes Model); set treatment definitions and country- specific treatment pathways; input details of costs and clinical data (such as event rates); specify the type of analysis to be performed (for example, open or closed cohort analysis, budget impact analysis, or sensitivity analysis); select appropriate risk engines (such as the United Kingdom Prevention of Diabetes Study [UKPDS] or Framingham for cardiovascular complications); and set preferences for the presentation of results. After an analysis has been performed, results in tabular and graphical format are presented via the user interface.

Figure 1. Overview of the CORE Diabetes Model structure

Input Databases

The information specified by the user is stored, together with pre-existing data from published sources, in the input databases of the CORE Diabetes Model. The input databases consist of a cohort database, a clinical database, a treatment database and an economics database, and are used as the basis for the calculations required to perform each simulation (Figure 1).

Cohort Database

The cohort database defines all properties of a patient cohort that can be set in the CORE Diabetes Model, including patient demographics (such as age, gender, ethnic group and duration of diabetes), baseline risk factors (glycosylated haemoglobin [HbA^sub 1c^] level, systolic blood pressure [SBP], total cholesterol [T- CHOL], high-density lipoprotein cholesterol [HDL], low-density lipoprotein cholesterol [LDL], triglycerides [TRIG], body-mass index [BMI], number of cigarettes smoked per day and alcohol consumption), and baseline existing complications in the simulated cohort (cardiovascular [myocardial infarction, angina, peripheral vascular disease (PVD), stroke, congestive heart failure (CHF), atrial fibrillation and left ventricular hypertrophy (LVH)], renal [microalbuminuria, gross proteinuria and end-stage renal disease (ESRD)], retinopathy [background or proliferative diabetic retinopathy (BDR or PDR) and severe vision loss (SVL)], macular oedema, cataract, foot ulcer [uninfected or infected ulcer, gangrene, healed ulcer and history of amputation] and neuropathy complications)31. Once cohort settings have been defined for a simulation, each generated patient is defined according to the settings entere\d by the user. Each virtual patient is simulated through the model. The probability of developing a complication, progressing to the next level of severity of disease or dying is compared to a random number generated from a uniform distribution (between 0 and 1) for any given event. If the number drawn is less than or equal to that of the probability for the event, the event is deemed to have occurred. The factors defined for the cohort group have an influence on long-term clinical and economic outcomes32. While the baseline characteristics of the cohort are defined at the start of the simulation, risk factors (including age, duration of diabetes, BMI, HbA^sub 1c^, blood pressure and lipid profile) and complication history are updated at the end of each cycle throughout the simulation, thus accounting for natural patho-physiological progression and subsequent changing (usually increasing) risk of complications with increasing duration of diabetes.

Clinical Database

The clinical database stores all of the medical and epidemiological data that is directly implemented in the model to calculate clinical outcomes. It is a set of probabilities and risk adjustment factors for disease progression and the occurrence of acute events or complications, which are based on physiological parameters and/or patient states and characteristics. The probabilities and risk adjustment factors in the CORE Diabetes Model are taken from published sources. However, users are able to modify this information as well as enter new data as it becomes available.

Treatment Database

The treatment database stores data on treatment pathways, treatment effect and the change in each physiological parameter in the simulation as a consequence of treatment and/or patient management. From the first year of simulation, the model determines the progression of each physiological parameter over time based on equations or other data from published sources stored in the treatment database. In addition to the clinical effects of treatments, all the relevant costs associated with a treatment (medications, monitoring, investigations, consultations) are assigned in the treatment cost group.

Economics Database

The economics database of the model defines the data that are used to perform economic analysis, including direct costs (costs of patient treatment and medications, consultations, investigations etc. for acute events and long-term complications), and indirect costs (based on the human capital approach33), discount rates (separated for clinical and cost outcomes), and quality-of-life data (utilities/disutilities associated with disease states and acute events). Treatment costs are stored in the treatment database because any given treatment in different countries, although the same in terms of effect on the patient, may be associated with different costs and resource consumption. This avoids any confusion involved in linking the same treatments with different (country- specific) costs.

Data Processor/Mathematical Engine

Information stored in the input databases forms the basis of the calculations required to run each simulation and performed by the data processor, which was programmed in C++ (Microsoft Visual Studio 6.0, Enterprise Edition). The CORE Diabetes Model utilises standard decision analysis techniques to process data. A combination of Markov model structure and Monte Carlo simulation using tracker variables has been implemented to capture the long-term and progressive nature of diabetes and its complications34,35. Markov models are ideally suited to representing recursive processes, or series of events or increasingly severe disease states that occur over time34. However, one potential drawback of the Markov approach is that it requires distinct disease states to be defined, each of which are mutually exclusive. In the real-life situation patients can develop complications in different body systems simultaneously, with the development of one complication having an influence on the rate of development of other complications. This problem is overcome in the CORE Diabetes Model using Monte Carlo analysis with tracker variables to overcome the memory-less properties of the Markov model. It allows interaction between different complication sub- models. Each sub-model runs simultaneously and in parallel, thereby allowing patients to develop multiple complications within each Markov cycle and over the simulation period. Progression of one or more complications influences transition probabilities in other sub- models (where a relevant link has been established).

The CORE Diabetes Model uses first and secondorder Monte Carlo simulations with or without distributions on input parameters such as baseline risk factors or transition probabilities36,37. Nonparametric bootstrap methods are used to evaluate uncertainty in cost-effectiveness outcomes measured38. Each probability in the model is simulated using a first-order Monte Carlo approach to represent sampling uncertainty. After 1,000 simulations of 1,000 non- identical patients, 1,000 bootstrap samples are drawn and the joint distribution of mean incremental costs and mean effectiveness gained is evaluated. The percentage of the joint distributions that fall within a cost-effective range is calculated, and an acceptability curve generated. As Monte Carlo simulations are based on random numbers generated by the computer, this approach reduces the noise due to the random number generator. This results in a much more accurate estimation of the 'true' mean than when only based on one sample, equivalent to repeating the same clinical trial several times. The nonparametric bootstrap method is used to estimate the joint distribution of mean incremental costs and mean effectiveness gained. The results obtained provide additional information on the uncertainty surrounding cost-effectiveness analyses performed using the model.

Closed and Open Cohort Simulations

The CORE Diabetes Model allows simulation of both closed and open cohorts. A closed cohort simulation involves the definition of a cohort with certain baseline characteristics which is simulated until the user-defined time horizon is reached, or until all patients in the cohort have died. An open cohort simulation incorporates population dynamics, whereby a prevalent cohort is defined at baseline, and patients both leave the simulated population (by dying) or enter the simulation (by adding incident patients to the diabetes population in question). In contrast to the closed cohort analysis, in which the population size always decreases due to death, the simulated population may actually increase in the open cohort simulation if the diabetes incidence rate is higher than the annual mortality rate within the population.

Budget Impact Analysis

While economic evidence is undoubtedly useful to purchasers, it does not address the issue of affordability. Healthcare purchasers are concerned not just with maximising efficiency but also with the more simplistic goal of remaining within annual budgets, which can be achieved through budget impact analysis39. Many countries require budget impact analyses to be included in the health economic dossier for new drug approvals40,41, and the CORE Diabetes Model is able to perform budget impact analyses over short- and long-term time horizons. Based on the open cohort simulation, users can specify the percentage of patients receiving various treatments in addition to the input data required for open cohort simulation (prevalent and incident cohort characteristics). This specification can be different in each scenario, allowing the comparison of current practice versus the introduction of a new product. Gradual introduction and increasing market share of a new product or intervention over time can be simulated, and the budget impact calculated. The market shares of various treatments are defined at baseline and the changes over time defined for up to 50 years. Therefore budget impact simulations produce results such as total budget (including treatment and complications costs) for a population of diabetes patients over a specified time horizon, but also the breakdown per complication and for treatment-related costs.

CORE Diabetes Model Sub-models

There are 15 sub-models in the CORE Diabetes Model, each of which simulates different complications associated with diabetes (Figure 2). Each sub-model is a Markov model using time-, state-, time-in- state and diabetes type-dependent probabilities (where appropriate and available) to simulate the progress of patients through different states. The sub-models simulate the following complications: angina, cataract, congestive heart failure, foot ulcer and amputation, hypoglycaemia, ketoacidosis, lactic acidosis, macular oedema, myocardial infarction, nephropathy, neuropathy, peripheral vascular disease, retinopathy, stroke and non-specific mortality. For each cycle the order in which the sub-models run changes randomly. The use of Monte Carlo simulation with tracker variables allows interaction between sub-models in order to simulate accurately the relationship between the development and progression of multiple complications within individual patients. An example of this is the increased risk of developing cardiovascular disease with increasing severity of renal disease42. When each simulated patient is determined to have developed any given level of complication, the value of the tracker variable that records the onset of this complication changes from 0 to 1, and the risk of developing other complications, dying, or transiting to other disease states is adjusted accordingly. There is one additional sub-model which is not related to complications. It is the type 2 diabetes treatment sequence sub-model and simulates changes in the treatment pathway over time due to treatment failure and/or side effects of treatments to control hyperglycaemia in simulati\ons of type 2 diabetes.

Myocardial Infarction

The myocardial infarction (MI) sub-model in the CORE Diabetes Model is made up of three states: No history of MI, History of MI and Death following MI. The probabilities for transitions between states in the submodel and risk adjustments (based on patients characteristics or treatments) are the same for patients with type 1 and type 2 diabetes. Users of the CORE Diabetes Model can select whether to use the Framingham risk function or the UKPDS risk engine to calculate probabilities of MI. Calculations using the Framingham risk functions are based on a proportional hazard regression model published by D'Agostino et al. which uses published data from the Framingham study43. The regression model predicts the probability of any cardiovascular events including angina, MI and coronary death. It is used to calculate separately the probability of initial and subsequent events for male and female patients. The risk model from the UKPDS risk engine is diabetes-specific and takes into account HbA^sub 1c^, SBP, lipid levels (total cholesterol to HDL ratio), age, sex, smoking status, race, and time since diagnosis of diabetes to calculate the probability of the first MI44.

Figure 2. CORE Diabetes Model Flow Diagram, Note: States within each of the sub-models shown in this figure can be found in the online appendices at www.thecenter.ch/cdmappendices. ACEI = Angiotensin converting enzyme inhibitor; MI = Myocardial infarction, CHF = Congestive heart failure; PVD = Peripheral vascular disease; ulc = Ulcer, Non-spec = Non-specific.

The probability of a first MI is calculated using either the Framingham risk function or the UKPDS risk engine, as previously described. The risk of recurrent MI is assumed to be the same for both males and females, is indexed by year after first MI, and is based on data from Sweden published by Herlitz et al. in 1996(45). The probability of death following MI is dependent on time after the event and is taken from several published sources. The probability of sudden death following myocardial infarction is approximately 0.393 for male patients and 0.364 for female patients46. For death within 12 months of MI, separate probabilities for male and female patients and for first and recurrent MIs are indexed by age. These probabilities are based on data from the DIGAMI study47 and have been adjusted according to age and gender in the CORE Diabetes Model45. The probabilities of death differ depending on whether patients receive conventional or intensive insulin therapy following MI48. Intensive insulin therapy included multiple daily blood glucose measurements and close medical supervision.

Risk adjustments in the MI sub-model can be made for glycaemic control, aspirin use, ACEI and statin treatment, and renal function, and are applied separately to the risks of first MI, recurrent MI, sudden death, death within 12 months and long-term death following MI. The default risk adjustments in the CORE Diabetes Model are outlined in Appendix 1 and are applied equally to type 1 and type 2 diabetes unless otherwise specified. The effects of age and gender are already taken into account by the risk models used to calculate the incident MI.

Angina

There are two states in the angina sub-model: No angina and History of angina. Transition probabilities in the sub-model are the same for patients with type 1 and type 2 diabetes. The probability of developing angina calculation is based on a proportional hazard regression model published by D'Agostino et al., taking into account data from the Framingham study43. The regression model predicts the probability of any coronary heart disease including MI, angina or sudden/non-sudden coronary death. This probability is subsequently multiplied by values for the proportion of patients with angina (out of all patients with events] to calculate the probability of developing angina. The default settings for the proportion of patients developing angina are 0.42 and 0.621 for male and female patients respectively43. Further details on the regression model are provided in the section describing the myocardial infarction sub- model.

Users can also set risk multipliers for cardiovascular disease associated with the onset of renal disease, as the development of microalbuminuria, gross proteinuria and ESRD are all markers of increased risk of cardiovascular disease49.

Congestive Heart Failure

The congestive heart failure (CHF) sub-model in the CORE Diabetes Model is made up of three states: No congestive heart failure, History of congestive heart failure and Death following congestive heart failure. A logistic regression model based on data from the Framingham study was used to generate the risk profile for CHF50, which takes into account age, left ventricular hypertrophy, heart rate, SBP, pre-existing congestive heart disease, valve disease and diabetes.

The probability of death following CHF events is indexed by time and age for males and females, and is based on the publication by Kalon et al. (Appendix 2)51. These data are comparable to survival curves published by Levy et al. in 2002(52).

The risk adjustment based on HbA^sub 1c^ level of the probability of suffering a CHF event can be made separately for patients with type 1 and type 2 diabetes. For type 1 diabetes patients, the default setting is 1 (no effect) until clinical data become available that support a link between improved glycaemic control and cardiovascular disease in type 1 diabetes. A 16% risk change for a 1% point change in HbA^sub 1c^ is the default setting for type 2 diabetes patients, based on UKPDS 3553. The CORE Diabetes Model also has the facility to adjust the risk of CHF and death following CHF according to aspirin, ACEI and statin treatment, and ethnic group. The default settings for these risk adjustments is no effect, with the exception of the risk adjustment for ACEI treatment on the probability of death following CHF. Based on data from the HOPE study, a risk reduction of 24% is associated with ACEI treatment for male and female patients with type 1 and type 2 diabetes54.

Stroke

The stroke sub-model is made up of three different states: No history of stroke, History of stroke and Death following stroke. Patients who have a stroke switch to (or remain in) the History of stroke state if they survive for more than 12 months following the event. Transition probabilities and risk adjustments are the same for patients with type 1 and type 2 diabetes in the stroke sub- model.

As in the MI sub-model, probabilities for the occurrence of stroke events can be calculated in two different ways in the CORE Diabetes Model, using either the Framingham risk function or the UKPDS risk engine (as selected by the user). Calculations using the Framingham risk functions are based on a proportional hazard regression model published by Wolf et al. which estimates the probability of stroke for patients between the ages of 55 and 84 years with no history of stroke55. This risk model takes into account age, SBP, antihypertensive medication, smoking status and the presence of other cardiovascular disease, atrial fibrillation and left ventricular hypertrophy.

The UKPDS risk engine provides an approach to assess the risk of first stroke using data from 4,540 patients of both sexes with type 2 diabetes56. The risk model is diabetes-specific and takes into account duration of diabetes, age, sex, smoking status, SBP, total cholesterol to HDL ratio, and presence of atrial fibrillation.

The probabilities of recurrent stroke in the CORE Diabetes Model are based on published data from the Rochester population reported by Petty et al., and are indexed by time since first stroke57. Transition probabilities for death within 30 days following stroke are different for male and female patients and for first and recurrent stroke, and are based on the same population57. The probability of death within 12 months of a stroke is based on data from the Rochester population published by Sprafka et al.58 Probabilities are different for male and female patients and are indexed by age. Transition probabilities for death due to long-term complications after stroke are based on the same report58.

Risk adjustments for the probability of first stroke are made according to aspirin use, ACEI and statin treatment, and renal function. It should be noted that the UKPDS risk function takes HbA]c level into account when calculating the risk of stroke. Risk multipliers to adjust for the effect of aspirin treatment (primary prevention) on first stroke probabilities are set to no effect59,60. The risk of first stroke is adjusted dependent on co-existing renal disease based on the data published by Valmadrid et al. in 2000 and indexed by severity (microalbuminuria, gross proteinuria and ESRD)42.

Adjustments to the risk of recurrent stroke are made according to aspirin and ACEI treatment, and renal function. The relative risk reduction associated with aspirin treatment is set to 27%59,60 and that associated with ACEI treatment is set to 33%54. Adjustment according to renal function is the same for recurrent stroke as for first stroke.

Peripheral Vascular Disease

The peripheral vascular disease (PVD) sub-model in the CORE Diabetes Model is made up of two states: IVo PVD and PVD. The probabilities for transitions between states in the sub-model and risk adjustments are the same for patients with type 1 and type 2 diabetes. To generate the risk profile for PVD (specifically, intermittent claudication), a logistic regression model based on data from the Framingham study was used61 which takes into account the risk associated with gender, age, blood pressure, hypertension, diabetes, smoking, cholesterol and concomitant heart disease.

Probability adjustments are made only for HbA^sub 1c^ levels and are only applied to patients with type 2 diabetes. HbA^sub 1c^ adjustment is based on data from th\e UKPDS and is indexed over change from a comparator value62.

Neuropathy

The neuropathy sub-model is a simple two state model: No neuropathy and Neuropathy. The baseline prevalence of neuropathy is defined separately for type 1 and type 2 diabetes and in both cases is indexed by duration of diabetes. Prevalences are based on the Diabetes Control and Complications Trial (DCCT) for type 1 diabetes and on the study by Partenen et al. for type 2 diabetes18,63. Transition probabilities for the onset of neuropathy for patients with type 1 diabetes are also based on DCCT data and are indexed by duration of diabetes and whether patients are receiving conventional or intensive insulin therapy63. For type 2 diabetes, transition probabilities were derived from the data published by Partanen et al.18

Risk adjustments are made for glycaemic control. For type 1 diabetes, risk adjustments are made based on HbA^sub 1c^ and by conventional or intensive therapy, based on DCCT data63. In type 2 diabetes analyses, risk adjustments for change in HbA^sub 1c^ and SBP were derived from the UKPDS43,53,56,64. Risk adjustments for ethnic group were made based on US epidemiological data65-67.

Foot Ulcer

The foot ulcer sub-model in the CORE Diabetes Model is made up of nine states: No foot ulcer, Uninfected ulcer, Infected ulcer, Healed ulcer, Uninfected recurrent ulcer, Infected recurrent ulcer, Gangrene, History of amputation, and Death (increased risk of death with amputation/infection). The probabilities for transitions between states in the sub-model and risk adjustments are the same for both type 1 and type 2 diabetes.

The baseline distribution of patients between the Uninfected ulcer, Infected ulcer, Healed ulcer and Healed ulcer, History of amputation states can be defined by the user. The probability of developing a foot ulcer and going on to have a lower extremity amputation is linked to PVD and neuropathy (Figure 2)68. Based on data published by Tennvall and Apelqvist, probabilities for the development of foot ulcer are indexed by whether patients are at low, moderate or high risk68. Low-risk patients have no history of PVD or neuropathy, moderate-risk patients have a history either of neuropathy or PVD, and high risk is defined as having history of both neuropathy and PVD.

Retinopathy

The retinopathy sub-model is made up of states based on disease status (No retinopathy, BDR, PDR and SVL), as well as screening, diagnosis and treatment.

For simulations of type 1 diabetes, transition probabilities in the retinopathy sub-model are based on published data from the DCCT69. These data describe the progression to BDR and PDR in the DCCT study cohort. Transition probabilities are indexed according to whether patients are receiving intensive or conventional insulin therapy and by duration of retinopathy (0-9 years was associated with the probabilities of primary prevention; 9+ years is associated with the probabilities of secondary intervention)69. Probabilities take into account screening rates, detection rates and laser treatment.

Type 1 diabetes transition probabilities in the retinopathy sub- model are adjusted according to patient HbA^sub 1c^ levels, SBP, concomitant ACEI and treatment. Risk adjustment for HbA^sub 1c^ levels is based on data from the DCCT and affects the transitions to BDR and from BDR to PDR70,73. The HbA^sub 1c^ adjustment for progression from PDR to SVL (with and without laser therapy) is set to no effect in the CORE Diabetes Model, as currently there are no clinical data currently available on this progression. The risk adjustment for SBP is based on DCCT data and is summarised in Appendix 3(71). Adjustment for ACEI treatment is based on the data published by Malik et al. showing reductions in the progression to PDR associated with an ACEI72. Adjustments for the progression to BDR and to SVL are set to no effect as no clinical data are currently available on these transitions.

Type 2 diabetes transition probabilities for the retinopathy sub- model are based on rates published from the Wisconsin Epidemiologic Study of Diabetic Retinopathy (WESDR)73,74 and are indexed by the duration of diabetes. The same probabilities for the progression to SVL were used as with type 1 diabetes (although the model is structured such that separate probabilities can be set for type 1 and type 2 diabetes). Risk adjustments for HbA^sub 1c^ level were derived from the UKPDS (Appendix 4)62. Risk adjustments for SBP are based on data from the WESDR study73-74. The influence of ACEI treatment on the risk of progression to BDR and PDR is based on data published by Chaturvedi et al. from the EURODIAB study75.

Risk adjustment of the progression of retinopathy according to ethnic group is the same for type 1 and type 2 diabetes. Risk multipliers for non-Caucasian patients were based on US epidemiological data published in 1995 by the National Institute of Diabetes and Digestive and Kidney Diseases and providing risk adjustments for African-American, Native American and Hispanic patients65-67.

Macular Oedema

The macular oedema sub-model is made up of three states: No macular oedema, Macular oedema and SVL. The probability of developing macular oedema and progression to SVL in patients with type 1 diabetes is based on data on long-term complications reported from the DCCT study cohort71. Transition probabilities are indexed by duration of diabetes and use of conventional or intensive glycaemic control. For type 2 diabetes, transition probabilities are based on the US data published by Javitt et al. and are also indexed by duration of diabetes73.

Risk adjustments for the onset of macular oedema in patients with type 1 diabetes are made for HbA^sub 1c^ level and SBP. HbA^sub 1c^ risk adjustment is based on DCCT data71 and SBP adjustment is made according to data from the UKPDS64. For type 2 diabetes patients, risk multipliers for HbA^sub 1c^. adjustment are taken from the UKPDS 50 publication62 and for SBP are based on UKPDS 3664. Risk adjustments for ethnic group for the onset of macular oedema are the same for type 1 and type 2 diabetes and are based on US epidemiological data65-67. The transition probabilities for the progression from macular oedema to SVL take into account screening rates, the effectiveness of screening procedures and laser treatment.

Cataract

The cataract sub-model consists of three states: No cataracts, First cataract with operation and second cataract with operation. For type 1 diabetes, the default probabilities for the incidence of first and subsequent cataract are taken from a study in diabetes outpatients in the UK published by Janghorbani et al.76 The model has the facility to index cataract transition probabilities by year if required. Data from the UKPDS was used to set the default probability of first cataract in both male and female patients with type 2 diabetes77. Probabilities for developing subsequent cataracts are the same as those for patients with type 1 diabetes76.

The default risk adjustment setting for HbAu level is no effect for patients with type 1 diabetes (no clinical data available). Risk adjustment for patients with type 2 diabetes is based on data from the UKPDS and is the same for first and recurrent cataracts53.

Nephropathy

The nephropathy sub-model in the CORE Diabetes Model is made up of seven different states based on disease status (No renal complications, Micro-albuminuria, Gross proteinuria, ESRD (either Haemodialysis, Peritoneal dialysis or Kidney transplant) and Death following ESRD. Treatment for the prevention or delay of progression of nephropathy other than glycaemic control and blood pressure are also considered (e.g. ACEI or angiotensin receptor blocker treatment). In the model, microalbuminuria is defined as urinary excretion of 30-300 mg albumin per 24 hours and gross proteinuria is > 300 mg albumin per 24 hours.

For type 1 diabetes, data on the cumulative incidence of progression to microalbuminuria and gross proteinuria over time were taken from the DCCT to calculate transition probabilities for patients receiving conventional or intensive insulin therapy78. Data on the onset of microalbuminuria were taken from both the primary prevention (mean baseline duration of diabetes 2.6 years) and secondary intervention (mean baseline duration of diabetes 8.75 years) cohorts in the DCCT. Cumulative incidence data on the progression to gross proteinuria was available only in the secondary intervention cohort. These data were used to derive transition probabilities for the progression from microalbuminuria to gross proteinuria, using the assumption that progression to gross proteinuria can only occur from a microalbuminuria state (Appendix 6).

Probabilities for the progression from gross proteinuria to ESRD in the CORE Diabetes Model are based on cumulative incidence data for type 2 diabetes patients in the Rochester population (Appendix 6)79. The probabilities for progression to ESRD are stored in tables in the CORE Diabetes Model, indexed by the duration of proteinuria.

In the default settings, the probability of progression from ESRD to death is dependent on treatment and ethnic group as published by Wolfe et al.80 in 1999, but the model can incorporate data from other sources such as the US Renal Data System (USRDS) or the UK Renal Registry81,82. Separate probability tables are stored in the CORE Diabetes Model for each treatment (haemodialysis, peritoneal dialysis and kidney transplant) indexed by ethnic group (Caucasian and African-American) and duration of disease. These probabilities are applied to both type 1 and type 2 diabetes patients.

Risk adjustments for type 1 diabetes transition probabilities in the nephropathy sub-model are made for HbA^sub 1c^ level71,79,80, SBP and ACEI treatment (Appendices 3-5). Improved glycaemic control has been demonstrated to improve survival in ESRD patients83. Separate HbA^sub 1c^ risk adjustments for th\e transition from ESRD to death for each treatment (haemodialysis, peritoneal dialysis and kidney transplant) can be made in the model, but as no specific clinical data are currently available these are all set to the same value (Appendix 4). Adjustments for SBP are also indexed using a comparator value and are based on data from the DCCT and Rochester populations71,79 as outlined in Appendix 3. Risk adjustment for ACEI treatment is based on a meta-analysis published by Kshirsagar et al.83

Transition probabilities for patients with type 2 diabetes were calculated from clinical data published by Ravid et al.79, Lewis et al.80, the Rochester population84 and Wolfe et al.85 Probabilities for the onset of microalbuminuria, progression to gross proteinuria and progression to ESRD are different for patients receiving and not receiving ACEI treatment in the CORE Diabetes Model84-86.

Risk adjustments for HbA^sub 1c^ levels are based on data from the UKPDS 34 publication and are outlined in Appendix 4(87). Similarly, risk adjustments for SBP are given in Appendix 3 and are also taken from the UKPDS88. Transition probabilities are given separately for type 2 diabetes patients receiving and not receiving ACEI treatment.

Risk adjustment for the progression of nephropathy in different ethnic groups is the same for type 1 and type 2 diabetes and was taken from published US data65-67.

Hypoglycaemia

The hypoglycaemia sub-model in the CORE Diabetes Model is made up of two states: Alive and Death (due to hypoglycaemia). Patients in the Alive state can experience a major hypoglycaemic event (generally defined as a hypoglycaemic event requiring third party medical intervention), which may result in death (switch to Death state) or survival (remain in the Alive state). For patients with type 1 diabetes, the probability of hypoglycaemic events is dependent on the type and intensity of insulin therapy that patients are receiving. The CORE Diabetes Model uses two methods of calculating transitions probabilities in the hypoglycaemia sub- model based on HbA^sub 1c^ levels and on age.

For the HbA^sub 1c^-based method, hypoglycaemic event probabilities are evaluated over a range of HbA^sub 1c^ values based on a Poisson regression formula used to calculate the first episode of hypoglycaemia and quadratic trend regression for the risk of any event, including recurrent events, based on the DCCT (Appendix 7)89.

The second method uses age-specific probability tables for major hypoglycaemic events requiring third party medical intervention. These probabilities are based on published data from the DCCT89, but may be edited and updated to take into account newer forms of insulin that have lower risks of hypoglycaemia (e.g. detemir or glargine insulin).

For patients with type 2 diabetes, probabilities are medication- dependent, with different probabilities for major hypoglycaemic events in patients receiving metformin, sulphonylurea, insulin or acarbose, or other treatments as published in the UKPDS 33 paper77.

The probability of death following hypoglycaemic events requiring intervention and events resulting in coma and/or seizure are taken from the DCCT for type 1 diabetes70. For type 2 diabetes, probabilities of major hypoglycaemic events requiring third party medical intervention are based on published data from Poland and Israel90,91. The probability of death following a hypoglycaemic event is not influenced by any other parameters in the CORE Diabetes Model.

A risk adjustment to the probability of hypoglycaemic events is made for ACEI treatment and is based on the publications by Morris et al.92 and Herings et al.93 Adjustments are applied to patients with type 1 and type 2 diabetes receiving ACEI leading to an increased risk of major hypoglycaemic events due to increased insulin sensitivity seen with ACEI use.

Ketoacidosis

The ketoacidosis sub-model is applied only in simulations of type 1 diabetes and is a simple two state model: Alive and Death (due to ketoacidosis). Patients in the Alive state can experience a ketoacidosis incident, which may result in death (switch to Death state) or survival (remain in the Alive state). The probability of ketoacidosis events for patients with type 1 diabetes can be adjusted separately for conventional and intensive glycaemic control and can be indexed by duration of diabetes. The default settings in the CORE Diabetes Model are based on DCCT data71, but can be adjusted to account for multiple daily injection versus continuous subcutaneous insulin infusion or other interventions with different risks of ketoacidosis. The probability of death following ketoacidosis is based on published data from Australia and corresponds to a case fatality rate of 2.7%!)4 but may be as high as 5%95.

Lactic Acidosis

The lactic acidosis sub-model is only applied to simulations of type 2 diabetes and, like the ketoacidosis sub-model, it is a simple two state model: Alive and Death (due to lactic acidosis). Only patients receiving metformin are considered to be at risk of lactic acidosis events. The probability of a lactic acidosis event for these patients is based on a literature review of cases of lactic acidosis between 1972 and 1982, which reported 42 cases of lactic acidosis in patients receiving metformin96. Although lactic acidosis may be a problem in type 2 patients taking biguanides, this complication is otherwise very rare and the default setting in the model is zero (resulting in the model only being applied to type 2 simulations97. Following a lactic acidosis event, the probability of death is set to 0.43 (43% risk of death)36.

Non-specific Mortality

The non-specific mortality sub-model consists of two states: Alive and Dead. The probabilities of non-specific death - that is, mortality from causes not covered in the other sub-models - are based on country-specific mortality statistics. The default setting utilises US mortality statistics from 1999 (published in 2001)98. Probabilities are stored in clinical tables indexed by age and ethnic group. Appendices 8 and 9 show the non-specific mortality probability values for male and female patients (respectively) in the US setting. These values are applied equally to patients with type 1 and type 2 diabetes. Diabetes interventions are assumed to have no impact on non-specific mortality.

Cycle Length

In the CORE Diabetes Model, the progress of the cohort is viewed and evaluated at discrete, pre-defined time intervals. The time interval for the calculation cycle is one year in most sub-models, with the exceptions of the foot ulcer sub-model (cycle length 1 month) and the hypoglycaemia sub-model (cycle length 3 months). Between cycles, patients in the simulation can move from one state to the next in the model, or stay in the same state according to transition probabilities based on published sources. The terminal condition, or point at which the simulation stops, is flexible and can be set by the user, with time horizons between 1 and 90 years possible.

Treatment Patterns

The treatment module in the CORE Diabetes Model covers information on treatments, treatment effect and the change of each physiological parameter in the simulation as a consequence of treatment. After the first year, the model determines the progression of each physiological parameter over time based on equations taken from published data or as fixed incremental adjustments every year (in the absence of any relevant published data). The pattern of progression associated with each treatment in the CORE Diabetes Model can be altered by the user. In simulations of type 1 diabetes, as well as defining new treatment strategies, users can specify whether patients will receive intensive or conventional insulin therapy, which influences the incidence of microvascular and macro-vascular complications, and defines the rate of major hypoglycaemic events for any given treatment. For instance, users are able to define the long-term effects of treatment on HbA^sub 1c^, the percentage of patients reaching target HbA^sub 1c^, SBP, total cholesterol, LDL, HDL, triglycerides, BMI and the incidence of hypoglycaemic episodes. For type 2 diabetes, each treatment is associated with an annual probability of treatment failure and/or side-effects and users can specify the risks of major hypoglycaemic events and lactic acidosis according to treatment pattern. Users can set the long-term effects of treatment on HbA^sub 1c^, SBP, total cholesterol, LDL, HDL, triglycerides, BMI, based on the current HbA^sub 1c^ level of post-prandial glucose levels associated with the treatment. After treatment failure or side effects, the user can define a long-term treatment pathway in a treatment tree that reflects local common practice. The type 2 diabetes treatment pathways in the model are fully flexible and adaptable to take into account differences in local treatment patterns. For example, a patient receiving inhaled insulin will have a set probability (user-defined) of experiencing treatment failure or side effects. The treatment tree design allows the simulated patient to change treatments, for example to injectable insulin, or continue with the same treatment and carry on in the simulation. The treatment trees in the CORE Diabetes Model allow up to three changes of therapy following treatment failure or side effects during a simulation.

Changes in HbA^sub 1c^ over time are taken into account in simulations of type 2 diabetes, as it has been shown that HbA^sub 1c^ levels gradually increase as beta-cell function diminishes over time77,87. This usually requires additional or more intensive therapy to maintain glycaemic control99. HbA^sub 1c^ levels are tracked from baseline levels (defined in the cohort definition section) over the course of the simulation. For simulations of type 1 diabetes, it is assumed that patients' insulin production is zero at the start of the simulation and HbA^sub 1c^ levels \are controlled by solely by exogenous insulin therapy.

Application of Costs and Quality of Life Utilities

The economics module of the CORE Diabetes Model allows users to define the data that are used to perform economic analysis, including direct costs (management costs, costs of ongoing disease complications and costs of acute events), data for the determination of indirect costs (based on the human capital approach), discount rates, and health state and event utilities/disutilities for all possible disease states and events captured in the model. The model is fully adaptable to allow the user to define costs based on the audience for the analysis, for example societal costs or third party healthcare payer costs could be used for any given simulation.

Total costs and quality-adjusted life expectancy are calculated as functions of the states of diabetic complications reached during a given year of simulation, plus any acute events that may occur during that year. Acute event costs and disutilities are accounted as they occur. The state costs and utilities are accounted annually and are cumulative. An example cost and utility data input set is provided in Appendices 10 and 11.

To provide default settings for the model, health state utilities were derived wherever possible from the UKPDS100, with data gaps filled using data from the Australian Institute of Health and Welfare Burden of Illness in Australia report101, and Tengs et al.102 (Appendix 11). Utilities are applied equally to type 1 and type 2 diabetes, as very few data are available on the former. The values for utilities/disutilities can be updated as improved data become available.

Output Databases

The output databases in the model store the results of the simulations to be sourced for presentation via the user interface. The model is designed to provide a set of standard default results in the output of every simulation. These include graphical representation of survival curves; total cumulative direct, indirect and combined costs (direct + indirect); total annual direct, indirect and combined costs (direct + indirect); and a breakdown of costs by complication and treatment costs. Numerical details are provided of the following (where Δ refers to the difference in outcome, setting 1 minus setting 2): mean, median and standard deviation of life expectancy; mean, median and standard deviation of quality-adjusted life expectancy; Δ life expectancy; Δ quality-adjusted life expectancy; Δ costs; Δ costs/Δ life expectancy (incremental cost-effectiveness ratio based on life expectancy); Δ costs/Δ quality-adjusted life expectancy (incremental cost-effectiveness ratio based on quality-adjusted life expectancy); and breakdown costs (for total costs related to each complication). In addition, users can also view cumulative events, cumulative incidence, annual costs or cumulative costs for the complications of diabetes. Closed cohort simulations also include the option of performing an acceptability curve and/or a net health benefit analysis (NHB) with upper and lower values for willingness to pay (WTP) defined by the user.

Sensitivity Analyses

The robustness of results from the CORE Diabetes Model can be tested by altering key variables in sensitivity analyses. Every cohort, economic, treatment, treatment cost, patient management and clinical variable in the model can be altered by the user by simply selecting upper and lower limits for the variable and the number of intervals to be tested in between. This allows users to estimate uncertainty in results from the model and evaluate the influence of key variables in any given simulation.

Discussion

During the creation of the CORE Diabetes Model, extensive literature searches were performed to identify appropriate data sources for the model. Studies were selected based on the criteria that they provided the appropriate data, and were the most recent and largest studies available. Preference was then given to epidemiological studies which collected 'real life' data over a long period of follow-up rather than comparative, controlled clinical trials. As a result, the model was created using data primarily from the UKPDS, DCCT, Framingham and, to a lesser extent, WESDR. This could be viewed as a potential criticism of the model, as it could be argued that these data sources may not be suitable for every population or setting simulated, but it should be noted that the CORE Diabetes Model has been designed such that any of the probabilities can be altered without any re-programming to suit user preference or incorporate newly-published data. A limitation of all health economic models is that, because they are based on data from clinical studies, they do not accurately reflect the real-life situation where factors such as non-compliance and varying standards of care may have an influence. This is a shortcoming of the CORE Diabetes Model, but it is difficult to avoid given that the majority of data used to create any disease model must come from clinical studies of the disease. When designing the model we preferentially included data from epidemiological studies, which could be predicted to reflect the real-life situation more accurately than clinical studies, and therefore minimise this problem.

As with any new health economics model, the external validity of the CORE Diabetes Model is a critical issue. A recent review of decision-analytic models of Helicobacter pylori eradication provided evidence that, due to flawed key assumptions, the models provided spurious information on the cost-effectiveness of treatments103. The external validity of the CORE Diabetes Model has been addressed in a separate publication104, which describes a total of 66 second- (internal) and third- (external) order validation analyses performed across a range of complications and outcomes simulated by the CORE Diabetes Model. Correlation analysis produced an R^sup 2^ value of 0.9224 (perfect fit = 1). Second-order validation analyses (model predictions versus observed outcomes reported in studies used to construct the model) gave an R^sup 2^ value of 0.9574, and the value for third-order analyses (model predictions versus observed outcomes reported in studies not used to construct the model) was 0.9023.

Conclusions

Health economic models of the modern era must be able to support decision making and education, as well as aid the communication of results to a wide-ranging audience. To fulfill this role, models need to be accepted by relevant target groups, especially by medical experts, providers of care, and healthcare payers. Therefore models must be transparent, with users having access to full descriptions of how they work and the data upon which they are based. In addition, the complex nature of advanced models of chronic illness means that clear descriptions are required to assess the credibility of the assumptions used to create the model.

The CORE Diabetes Model utilises standard decision analysis techniques. The combination of Monte Carlo simulation and Markov modelling has been implemented to capture the long-term and progressive nature of diabetes and the interactivity of its complications. It has been designed to be customised to meet audience-specific needs, and to present results in a format that is easy to interpret. Economic and clinical data in the model can be edited by the user, thus ensuring adaptability by allowing the inclusion of new data as they become available; creation of country- , health maintenance organisation- or provider-specific versions of the model; and investigation of new hypotheses. Having the model based on a central server with Internet access makes it possible to refine, update and improve the model while maintaining rigid version control, overcoming a typical problem with multiple releases of stand-alone models. This approach also ensures that all users have the most up-to-date version and data. It is hoped that the flexibility of the model structure, with specific regard to treatment pathways, will ensure that any intervention or policy implementation can be accurately simulated and that the model will continue to evolve and keep pace with new data and treatment concepts.

We have described the CORE Diabetes Model in detail to provide users with an overview and understanding of how the model works, and to explain the assumptions and data sources upon which it was built. The CORE Diabetes Model allows extrapolation of results obtained from short-term trials to long-term outcomes. Diabetes management strategies can be compared in different patient populations in a variety of realistic clinical settings, allowing investigations geared towards improving the quality of care for diabetes patients. It is hoped that this advanced and expansive mathematical model for type 1 and type 2 diabetes will become a valuable tool for decision makers and developers of diabetes interventions alike.

Acknowledgements

The CORE Diabetes Model was developed independently by CORE - Center for Outcomes Research, with no external funding.

References

1. Pharmaceutical Benefits Advisory Committee. Guidelines for the pharmaceutical industry on preparation of submissions to the PBAC. 2002.

2. A proposal for methodological guidelines for economic evaluations of pharmaceuticals. Gent: Belgian Society for Pharmaco- epidemiology BESPE, 1995.

3. Canadian Coordinating Office for Health Technology Assessment. Guidelines for economic evaluations of pharmaceuticals: Canada. 1997. Ottawa, CCOHTA.

4. Hannover Consensus Group. German recommendations for healthcare economic evaluation studies. Revised version of the Hannover consensus. Med Klin 2000; 95(1):52-5.

5. Capri S, Ceci A, Terranova L. Guidelines for economic evaluations in Italy: recommendations from the Italian Group of Pharmacoeconomic Studies. Drug Information Journal 2001; 35(1):189- 201.

6. National Institute f\or Clinical Excellence. Guide to the technology appraisal process. 2001.

7. Sullivan SD, Lyles A, Luce B et al. AMCP guidelines for submission of clinical and economic evaluation data to support formulary listing in the US health plans and pharmacy benefits management organizations. J Managed Care Pharm 2001; 7:272-82.

8. The Regence Group. The Reference Group Formulary Submission Guidelines 2002. A format for submission of clinical and economic evaluation data in support of reimburement and formulary consideration to the Regence Group. The Regence Group 2002.

9. Weinstein MC, O'Brien B, Hornberger J, Jackson J, Johannesson M, McCabe C et al. Principles of good practice for decision analytic modeling in health-care evaluation: report of the ISPOR Task Force on Good Research Practices - Modeling Studies. Value Health 2003; 6(1):9-17.

10. Rubin RJ, Altman WM, Mendelson DN. Health care expenditures for people with diabetes mellitus, 1992. J Clin Endocrinol Metab 1994; 78:809A-F.

11. Brandle M, Zhou H, Smith BR, Marriott D, Burke R, Tabaei BP, Brown MB, Herman WH. The direct medical cost of type 2 diabetes. Diabetes Care 2003; 26(8):2300-4.

12. Harris MI, Flegal KM, Cowie CC, Eberhardt MS, Goldstein DE, Little RR et al. Prevalence of diabetes, impaired fasting glucose, and impaired glucose tolerance in U.S. adults. The Third National Health and Nutrition Examination Survey, 1988-1994. Diabetes Care 1998; 21(4):518-24.

13. Takiya L, Chawla S. Therapeutic options for the management of type 2 diabetes mellitus. Am J Manag Care 2002; 8(11):1009-23.

14. Amos AF, McCarty DJ, Zimmet P. The rising global burden of diabetes and its complications: estimates and projections to the year 2010. Diabet Med 1997; 14 Suppl 5:51-85.

15. Zimmet P, Alberti KG, Shaw J. Global and societal implications of the diabetes epidemic. Nature 2001; 414(6865):782- 7.

16. Uusitupa M, Siitonen O, Aro A, Pyorala K. Prevalence of coronary heart disease, left ventricular failure and hypertension in middle-aged, newly diagnosed type 2 (non-insulin-dependent) diabetic subjects. Diabetologia 1985; 28(1):22-7.

17. Harris MI. Undiagnosed NIDDM: clinical and public health issues. Diabetes Care 1993; 16(4):642-52.

18. Partanen J, Niskanen L, Lehtinen J, Mervaala E, Siitonen O, Uusitupa M. Natural history of peripheral neuropathy in patients with non-insulin- dependent diabetes mellitus. N Engl J Med 1995; 333(2):89-94.

19. Atkinson MA, Maclaren NK. The pathogenesis of insulin- dependent diabetes mellitus. N Engl J Med 1994; 331(21):1428-36.

20. Kretowski A, Kowalska I, Peczynska J, Urban M, Green A, Kinalska I. The large increase in incidence of Type I diabetes mellitus in Poland. Diabetologia 2001; 44 Suppl 3:B48-B50.

21. Podar T, Solntsev A, Karvonen M, Padaiga Z, Brigis G, Urbonaite B et al. Increasing incidence of childhood-onset type I diabetes in 3 Baltic countries and Finland 1983-1998. Diabetologia 2001; 44 Suppl 3:B17-20.

22. Pundziute-Lycka A, Dahlquist G, Nystrom L, Arnqvist H, Bjork E, Blohme G et al. The incidence of Type I diabetes has not increased but shifted to a younger age at diagnosis in the 0-34 years group in Sweden 1983-1998. Diabetologia 2002; 45(6):783-91.

23. Weets I, De Leeuw IH, Du Caju My Rooman R, Keymeulen B, Mathieu C et al. The incidence of type 1 diabetes in the age group 0- 39 years has not increased in Antwerp (Belgium) between 1989 and 2000: evidence for earlier disease manifestation. Diabetes Care 2002; 25(5):840-6.

24. Keller RJ, Eisenbarth GS, Jackson RA. Insulin prophylaxis in individuals at high risk of type I diabetes. Lancet 1993; 341 (8850):927-8.

25. Ziegler AG, Bachmann W, Rabl W Prophylactic insulin treatment in relatives at high risk for type 1 diabetes. Diabetes Metab Rev 1993; 9(4):289-93.

26. Fuchtenbusch M, Rabl W, Grassl B, Bachmann Vf, Standl E, Ziegler AG. Delay of type I diabetes in high risk, first degree relatives by parenteral antigen administration: the Schwabing Insulin Prophylaxis Pilot Trial. Diabetologia 1998; 41(5):536-41.

27. Eastman RC, Javitt JC, Herman WH, Dasbach EJ, Zbrozek AS, Dong F et al. Model of complications of NIDDM. I. Model construction and assumptions. Diabetes Care 1997; 20(5):725-34.

28. Brown JB, Russell A, Chan W, Pedula K, Aickin M. The global diabetes model: user friendly version 3.0. Diabetes Res Clin Pract 2000; 50 Suppl 3:S15-46.

29. Palmer AJ, Brandt A, Gozzoli V, Weiss C, Stock H, Wenzel H. Outline of a diabetes disease management model: principles and applications. Diabetes Res Clin Pract 2000; 50 Suppl 3:S47-S56.

30. Eddy DM, Schlessinger L. Archimedes: a trial-validated model of diabetes. Diabetes Care 2003; 26(11):3093-101.

31. National Diabetes Data Group, ed. Diabetes in America. Bethesda, MD: National Institutes of Health, National Institute of Diabetes and Digestive and Kidney Diseases, 1995; Chapters 8, 9, 12- 20:165-220, 259-456.

32. King H, Aubert RE, Herman WH. Global burden of diabetes, 1995- 2025: prevalence, numerical estimates, and projections. Diabetes Care 1998; 21(9):1414-31.

33. Gold MR, Siegel JE, Russell LB, Weinstein MC. Cost- effectiveness in health and medicine. 1 edn. New York: Oxford University Press, 1996.

34. Sonnenberg FA, Beck JR. Markov models in medical decision making: a practical guide. Med

Source: Current Medical Research and Opinion

More News in this Category