Introduction
Diabetes is a chronic metabolic disorder and is associated with a wide array of complications. In 2023, 10 percent of Canadians were living with diabetes. Type 2 diabetes mellitus (T2DM) is the more prevalent form, accounting for more than 90 percent of cases (1). Complications from diabetes can have a significant impact on the healthcare system. The estimated direct healthcare system costs associated with diabetes in Canada in 2020 were CAD 3.8 billion (USD 3.2 billion) (2).
Given these substantial economic and health implications, there is an increased interest in interventions to reduce T2DM and the related economic burden, with a particular emphasis on prevention, as preventative interventions have demonstrated effectiveness in delaying or potentially preventing the onset of T2DM (Reference Alouki, Delisle, Bermúdez-Tamayo and Johri3;Reference Roberts, Barry and Craig4). Evidence evaluating the potential health and economic outcomes of T2DM interventions is essential for helping policymakers make informed funding decisions, especially when comparing interventions aiming at preventing versus managing T2DM. Simulation modeling is particularly well suited to this, as T2DM interventions typically incur high upfront costs while the benefits associated with reducing the burden of disease and lowering healthcare costs are realized gradually over years or even decades.
Several T2DM simulation models have been published to evaluate the long-term cost-effectiveness of various T2DM interventions (Reference Hayes, Leal, Gray, Holman and Clarke5–Reference Hoerger, Hilscher and Neuwahl28). However, the existing models have limitations in their ability to evaluate preventive interventions, as they do not simulate either health states prior to the diagnosis of T2DM (Reference Hayes, Leal, Gray, Holman and Clarke5–Reference Sobotka, Cipriano, Coyle, Thiruchelvam and Shah11;Reference Hoerger, Hilscher and Neuwahl28), or the correlated progression of T2DM risk factors over time and comorbidities before and after T2DM onset (Reference Dall, Storm and Semilla12–Reference Vandenberghe27).
While the implications of not simulating health states prior to the diagnosis of T2DM on model suitability for the evaluation of preventative interventions are self-evident, the implications of not simulating the correlated progression of T2DM risk factors merit further explanation. Studies have shown that many different risk factors are important in predicting T2DM-related comorbidities and death (Reference Hayes, Leal, Gray, Holman and Clarke5). Multiple risk factors are also used to identify individuals at a high risk of developing T2DM who can benefit from preventive interventions (29). Interventions affect multiple risk factors simultaneously, and the relationships between risk factors affect the risk of other comorbidities, such as cardiovascular disease in patients before they develop T2DM (Reference Knowler and Fowler30;Reference Breeze, Squires and Chilcott31). Until recently, there has been a paucity of published studies on the correlated progression of T2DM risk factors over time (Reference Palmer32;Reference Leal, Alva and Gregory33). Modelers have had to make simplifying assumptions, such as holding risk factors constant or estimating T2DM progression conditioned on a single risk factor, such as impaired glucose tolerance or BMI (Reference Galani, Schneider and Rutten13;Reference Gillies, Lambert and Abrams14), which can limit the ability of models to accurately predict long-term health outcomes.
We sought to develop a T2DM simulation model, the Institute of Health Economics Diabetes Model (IHE-DM), that addresses these gaps. First, our model simulates progression from normal glucose tolerance (NGT) to pre-T2DM to T2DM, allowing for the evaluation of interventions at any point in the care pathway, including preventative interventions. Second, it incorporates the correlated progression of patient-level risk factors across this full continuum. Finally, our model simulates comorbidities both before and after the onset of T2DM. This comprehensive approach allows us to capture the full benefits of interventions, including their impact on health before the onset of T2DM, and provides a robust tool for assessing the long-term health and economic outcomes of different T2DM interventions applied across the care continuum.
This paper aims to detail the development and validation of the IHE-DM. We also demonstrate its capabilities by investigating the long-term cost-effectiveness of a hypothetical lifestyle intervention-based T2DM prevention program.
Methods
Model overview
The IHE-DM is an individual-level discrete-time microsimulation model that simulates: (i) progression from NGT to T2DM, (ii) the occurrence and timing of T2DM-related comorbidities and mortality, and (iii) time-varying changes in individual-level T2DM risk factors.
All patients move through two model levels during each 1-year model cycle. This continues until either a user-defined time horizon (e.g., 25 years) is reached or the patient dies. The top level of the model determines the patient’s T2DM health state, where they can be in one of two independent T2DM health states determined by their HbA1c: (i) NGT/pre-T2DM, with an HbA1c of less than 6.5 percent, or (ii) T2DM, with an HbA1c of 6.5 percent or higher (34).
The second level of the model predicts the occurrence and timing of T2DM-related comorbidities (congestive heart failure [CHF], ischemic heart disease [IHD], blindness, renal failure, myocardial infarction [MI], stroke, amputation, diabetic ulcer) and death conditioned on the patient’s T2DM health state and their individual risk factors (e.g., age, history of comorbidities, BMI). Time-varying risk factors are also updated based on the patient’s T2DM health state and their individual risk factors in this section of the model. Figure 1 provides an overview of the model structure.
Model structure. AF, atrial fibrillation; BMI, body mass index; CHF, congestive heart failure; eGFR, estimated glomerular filtration rate; HAEM, hemoglobin; HbA1c, glycated hemoglobin; HDL, high-density lipoprotein; HR, heart rate; IHD, ischemic heart disease; LDL, low-density lipoprotein; LVH, left ventricular hypertrophy; MMALB, micro- or macro-albuminuria; MI, myocardial infarction; NGT, normal glucose tolerance; NMB, net monetary benefit; pre-T2DM, pre-type 2 diabetes mellitus; PVD, peripheral vascular disease; QALYs, quality-adjusted life-years; SBP, systolic blood pressure; T2DM, type 2 diabetes mellitus; UKPDS-OM2, United Kingdom Prospective Diabetes Study Outcomes Model Version 2; WBC, white blood cell count.
Simulated population
The IHE-DM starts by generating initial characteristics for each patient by drawing random values from specific distributions corresponding to each characteristic (Section 1.2 of the Supplementary Material), which can be adjusted to simulate different populations. The patient’s initial HbA1c is used to determine their initial T2DM health state.
Patients in the NGT/pre-T2DM state
While in the NGT/pre-T2DM health state, patients can experience three comorbidities: MI (Reference Anderson, Odell, Wilson and Kannel35), CHF, and stroke (Reference D’Agostino, Vasan and Pencina36). The probability of each comorbidity occurring is calculated for each patient in every time cycle based on their individual characteristics (Figure 2). These probabilities are evaluated in a randomized order, allowing comorbidities to occur in a random sequence throughout the year. Once it occurs, CHF is considered a chronic comorbidity that persists through the remainder of the simulation or until death. Acute events include MI and stroke, which can occur once in the NGT/pre-T2DM health state and up to two times if the patient transitions to T2DM.
Independent variables are included in the model equations. Red shading indicates independent variables that are time-invariant (i.e., sex) or time logic (age)/current period, blue shading indicates independent variables that are calculated based on any prior period, purple shading indicates independent variables that are calculated based on the current period and any prior period; Albuminuria, micro- or macro-albuminuria; CHF, congestive heart failure; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HbA1c, glycated hemoglobin; HDL, high-density lipoprotein; IHD, ischemic heart disease; LDL, low-density lipoprotein; LVH, left ventricular hypertrophy; MI, myocardial infarction; PVD, peripheral vascular disease; SBP, systolic blood pressure; T2DM, type 2 diabetes mellitus; WBC, white blood cell count. Mortality Eq (i): patients with no history of comorbidities and no comorbidities occur in the current year; Mortality Eq (ii): patients with no history of comorbidities but who experience one or more comorbidities in the current year; Mortality Eq (iii): patients with a history of comorbidities but experience no new comorbidities in the current year; Mortality Eq (iv): patients with a history of comorbidities who also experience at least one new comorbidity in the current year. *The prediction equation for LVH comes from an analysis by de Simone et al. (1994).
In any patient-year that a comorbidity occurs, the probability of death is based on the 1-year, age- and sex-specific all-cause mortality probability for that comorbidity (37–Reference McAlister, Bakal and Kaul39). In patient-years without any comorbidities, the annual probability of death from all other causes is used, adjusted for mortality from MI, CHF, and stroke, which have already been accounted for (40).
Transitions to T2DM
At the end of each patient-year, the patient’s time-varying risk factors are updated based on their individual characteristics, using the appropriate equation (Figure 2). For patients with pre-T2DM, research indicates that the only significant difference in HbA1c progression between those who develop T2DM and those who do not occurs in the year before T2DM onset, with patients who transition to T2DM experiencing a sudden increase in HbA1c (Reference Heianza, Arase and Fujihara41;Reference Heianza, Arase and Fujihara42).
To incorporate these findings in the IHE-DM, if a patient’s HbA1c is between 6.0 and 6.4 percent (pre-T2DM), we include a 0.129 probability that their HbA1c will increase by 0.54 and they will transition to T2DM. Otherwise, their HbA1c is updated based on their individual characteristics using the HbA1c equation. The updated HbA1c is used at the start of the next model cycle to determine the patient’s T2DM health state (NGT/pre-T2DM or T2DM). A similar method of modeling T2DM state transitions has been used by Dall et al. (Reference Dall, Storm and Semilla12).
Patients in the T2DM state
For patients who transition to T2DM, the model uses the UKPDS-OM2 risk equations to predict comorbidities, mortality, and risk factor progression (Reference Hayes, Leal, Gray, Holman and Clarke5;Reference Leal, Alva and Gregory33). These equations have been externally validated (Reference Pagano, Konings and Di Cuonzo43;Reference Keng, Leal and Mafham44) and are widely used in T2DM simulation models (Reference Dall, Storm and Semilla12;Reference Lindgren, Lindström and Tuomilehto15). They were derived from data collected during the UKPDS trial and post-trial follow-up (Reference Holman, Paul, Bethel, Matthews and Neil45), which included newly diagnosed T2DM patients receiving treatments ranging from conventional glucose control through diet and lifestyle to intensive glucose control with metformin, sulfonylureas, insulin, or combination therapy – reflecting the current standard of care at the time of the study. Therefore, the IHE-DM assumes that all patients who progress to T2DM receive treatment consistent with the standard of care in the UKPDS.
Patients can experience eight comorbidities: CHF, IHD, blindness, renal failure, MI, stroke, amputation, and diabetic ulcer (Reference Hayes, Leal, Gray, Holman and Clarke5). The probability of each comorbidity occurring is calculated per patient-year based on the patient’s individual characteristics (Figure 2). These probabilities are evaluated in a randomized order, allowing comorbidities to occur in a random sequence throughout the year. Once they occur, chronic comorbidities (CHF, IHD, blindness, and renal failure), persist through the remainder of the simulation or until death. Acute events include MI, stroke, and amputation, each of which can occur up to two times.
The probability of mortality is calculated using one of four mutually exclusive equations for patients with: (i) no history of comorbidities and no comorbidities in the current year, (ii) no history of comorbidities but experience one or more comorbidities in the current year, (iii) history of comorbidities but no new comorbidities in the current year, and (iv) history of comorbidities and experience at least one new comorbidity in the current year.
At the end of each patient-year, time-varying risk factors are updated based on the patient’s individual characteristics (Figure 2).
Costs and health-related quality of life
All costs are reported in Canadian dollars, inflated to 2022 prices using the healthcare component of the Canadian Consumer Price Index (46). The costs associated with T2DM and related comorbidities include the cost of inpatient hospital stays, outpatient visits, prescription drugs (both related to diabetes as well as other medications for patients aged 65 and older), emergency room visits, long-term care, and home care (Reference O’Reilly, Hopkins and Blackhouse7). To obtain the comorbidity costs of MI, CHF, and stroke for patients without T2DM, we use condition-specific cost ratios (Reference Choi, Booth and Jung47).
Baseline QALY values by age and sex and utility decrements for T2DM and comorbidities were sourced from literature as indicated in Table 1, which can be adjusted to reflect different populations.
Model equations and parameters data sources
Note: AF, atrial fibrillation; BMI, body mass index; CHF, congestive heart failure; eGFR, estimated glomerular filtration rate; HAEM, hemoglobin; HbA1c, glycated hemoglobin; HDL, high-density lipoprotein; HR, heart rate; IHD, ischemic heart disease; LDL, low-density lipoprotein; LVH, left ventricular hypertrophy; MI, myocardial infarction; MMALB, micro- or macro-albuminuria; NGT, normal glucose tolerance; pre-T2DM, pre-type 2 diabetes mellitus; PVD, peripheral vascular disease; QALYs, quality-adjusted life-years; SBP, systolic blood pressure; T2DM, type 2 diabetes mellitus; UKPDS-OM2, United Kingdom Prospective Diabetes Study Outcomes Model Version 2; WBC, white blood cell count.
Each component of the model and the model parameters are described in greater detail in Section 1.3 of the Supplementary Material. Table 1 summarizes the parameters, equations, and data sources for each major component of the model. Figure 2 shows the individual-level characteristics that are predictors (independent variables) for comorbidities, mortality, and changes in risk factors over time. Probability values for model parameters not calculated using equations (e.g., probability of mortality in the NGT/pre-T2DM health state), along with cost and QALY values, are available in Section 1.3 of the Supplementary Material.
Handling uncertainty
The IHE-DM deals with multiple levels of uncertainty. Initial characteristics and risk factors for each patient are generated by drawing random values from specified distributions to account for sampling uncertainty and patient heterogeneity (Reference Dakin, Leal and Briggs56). To minimize stochastic uncertainty, the model simulates a large number of individuals, which can be adjusted to ensure stable means for the outcomes of interest. After generating the initial population, parameter uncertainty in the clinical, cost, and QALY parameters is addressed through probabilistic sensitivity analysis (PSA), where 1,000 sets of input parameters are randomly sampled from the underlying probability distributions, and the results are simulated for each set of parameters. Further details on the PSA are provided in Section 1.5 of the Supplementary Material.
Model validation
Through multiple iterations of model development, pressure testing, and extreme value analysis, we verified that all model elements were implemented correctly and that the model generated logical results (Section 1.4 of the Supplementary Material). Internal validation assessed how well the model predicted outcomes in the dataset used to derive the UKPDS-OM2 equations. External validation tested the model’s ability to predict T2DM occurrence in an external dataset from the Diabetes Prevention Program Outcome Study (DPPOS) (Reference Knowler and Fowler30;57).
Internal validation
Once patients transition to T2DM, the model uses the UKPDS-OM2 risk equations to predict T2DM-related comorbidities and death (Reference Hayes, Leal, Gray, Holman and Clarke5) and to simulate risk factor progression over time (Reference Leal, Alva and Gregory33). To validate the implementation of these equations in the IHE-DM, we used the published baseline characteristics of UKPDS participants to generate a simulated cohort of 5,100 patients with newly diagnosed T2DM, reflecting the population size of the UKPDS. We then simulated risk factor progression, the development of comorbidities, and mortality for 20 years.
We compared the observed Kaplan–Meier (KM) cumulative failure probability from the UKPDS-OM2 to the simulated values from the model by plotting the simulated versus observed values for each of the comorbidities and mortality at years 5, 10, 15, and 20. If the model predicted the observed data perfectly for each endpoint, the plot points would fall along a 45-degree identity line. A majority of points falling above this line would indicate that the model overestimates the observed values. Conversely, a majority of points falling below the line would indicate that the model underestimates the observed values. We evaluated the accuracy of all predictions by fitting a regression line through the observed and predicted values for all endpoints, and estimating the intercept, slope coefficient, and the coefficient of determination (R2) (Reference Steyerberg, Vickers and Cook58). Perfectly accurate model predictions would have an intercept of zero, a slope coefficient of one, and an R2 of one. We used the same method to assess macrovascular (CHF, IHD, MI, stroke) and microvascular (blindness, renal failure, amputation, diabetic ulcer) endpoints separately. Last, we compared the simulated average risk factor progression for each risk factor with the average observed risk factor progression from the UKPDS over a 20-year follow-up period.
External validation
For external validation we required a longitudinal data source that reported: (i) a population without T2DM at baseline, (ii) baseline risk factors similar to those in our model, particularly HbA1c as this was the measure of blood glucose we used to determine T2DM health states, and (iii) outcomes consistent with our model outcomes (e.g., T2DM incidence). We selected the DPPOS (Reference Knowler and Fowler30;57), a long-term follow-up of the Diabetes Prevention Program trial (DPP), as it met these criteria.
We used the published baseline characteristics of the trial participants enrolled in the DDPOS to generate simulated cohorts of 5,100 patients for both the placebo and intensive lifestyle intervention (ILI) arm of the trial. When parameter values for initial risk factors were unavailable in the DPPOS, we used values from the UKPDS.
We simulated the annual effects of the ILI on BMI and HbA1c changes for the first 10 years following the intervention, as reported in the DPPOS 10-year follow-up (Reference Knowler and Fowler30). For patients with pre-T2DM, we applied a reduction to their probability of experiencing a sudden increase in HbA1c to more closely reflect the decrease in T2DM incidence observed in the ILI arm compared to the placebo arm in the DPPOS (Section 1.3 of the Supplementary Material).
We compared the mean simulated and observed KM cumulative incidence of T2DM annually for 15 years from randomization for each study arm (57). We calculated the average ratio of simulated to observed KM cumulative incidence. Values close to one would indicate good alignment between the model predictions and observed values from the DPPOS. Values greater than one would indicate that the model overpredicted the observed values, while values less than one would indicate underprediction.
Case study: Cost-effectiveness of a potential T2DM prevention program
To illustrate the capabilities of the IHE-DM, we evaluated the 15-year cost-effectiveness of a hypothetical lifestyle intervention-based T2DM prevention program in Alberta, Canada. To align with the size of our validation cohorts, we simulated cohorts of 5,100 patients in each arm (standard care and intervention).
We based the T2DM prevention program on the ILI arm of the DPP trial. We simulated the annual effects of the ILI on BMI and HbA1c as described in the external validation section. The program compliance rate was assumed to be 50 percent (Reference Cannon, Masalovich and Ng59;Reference Ritchie, Baucom and Sauder60). The program cost was assumed to be the same as the direct medical cost of the intensive lifestyle intervention in the DPP (Reference Hernan, Brandle and Zhang61). We inflated the DPP cost to 2022 dollars using the US Consumer Price Index (62) and converted to Canadian dollars using the purchasing power parities (PPP) rate for 2022 (63).
Costs and QALYs were discounted at a 1.5 percent annual rate per the Canada Drug Agency Guidelines (64). We used a willingness-to-pay (WTP) threshold of CAD 30,000 (USD 25,424) (Reference Ochalek, Lomas and Claxton65). Section 1.5 of the Supplementary Material provides a complete description of the analysis.
Results
Internal validation
Results of the internal validation of the T2DM health state are shown in Figure 3, which plots the simulated and observed KM cumulative failure probability for each T2DM-related complication and mortality at years 5, 10, 15, and 20. Figure 3A plots all endpoints, Figure 3B plots microvascular endpoints only, and Figure 3C plots macrovascular endpoints only.
Predicted versus observed Kaplan–Meier cumulative failure probability, internal validation. Simulated mean Kaplan–Meier cumulative failure probability for each endpoint is calculated from 1,000 PSA iterations of 5,100 patients. The dashed line on the graphs represents the 45-degree reference line, and the solid line represents the fitted regression line. Each of the comorbidities and mortality are represented by three endpoints corresponding to years 5, 10, 15, and 20. CHF, congestive heart failure; IHD, ischemic heart disease; MI, myocardial infarction.
The intercepts are not statistically different from zero in all cases (all endpoints: 0.001 [95% confidence interval [CI]: −0.006, 0.008], microvascular endpoints: −0.002 [95% CI: −0.006, 0.003], and macrovascular endpoints: −0.010 [95% CI: −0.025, 0.006]). The slopes of the regression lines for each of the plots are not statistically different from one for both all endpoints and for microvascular endpoints (all endpoints: 1.007 [95% CI: 0.950, 1.065], microvascular endpoints: 1.088 [95% CI: 0.943, 1.233]). The slope for macrovascular endpoints is statistically significantly different from one (1.157 [95% CI: 1.008, 1.306]) but is relatively small in magnitude. The R2 value is greater than 0.95 in all cases.
The mean absolute difference between predicted and observed values was under 1 percent across all comorbidities and mortality. The exact simulated and observed values for each complication and mortality and the mean simulated and observed risk factor progression can be found in Section 2.1 of the Supplementary Material.
External validation
Figure 4 reports the mean simulated and observed KM cumulative incidence of T2DM annually for 15 years from randomization for the placebo arm (Figure 4A) and the ILI arm (Figure 4B). The ratio of simulated to observed KM cumulative incidence of T2DM is less than one in each year for the first 9 years in the ILI arm and in each year for the first 10 years in the placebo arm. At year 10, the simulated to observed KM cumulative incidence ratio is 1.04 and 0.99 for the ILI and placebo arms, respectively. From years 11 through 15, the simulated to observed KM cumulative incidence ratio is greater than one for both arms. At year 15, the simulated to observed KM cumulative incidence ratio is 1.38 and 1.30 for the ILI and placebo arms, respectively. The exact values for each year can be found in Section 2.1 of the Supplementary Material.
Simulated and observed cumulative incidence of T2DM, external validation. Simulated mean cumulative incidence of T2DM is calculated from 1,000 PSA iterations of 5,100 patients per iteration for each arm (placebo and intervention). PSA, probabilistic sensitivity analysis; T2DM, type 2 diabetes mellitus.
Case study: Cost-effectiveness of a potential T2DM prevention program
We estimate that the hypothetical intervention has an incremental net monetary benefit of CAD 2,701 (USD 2,289) (95% Uncertainty Interval [UI]: CAD 1,316 to 4,000 [USD 1,115–3,390]) over the 15-year time horizon, becoming cost-effective in year 10. The intervention increases costs by CAD 541 (USD 458) (95% UI: CAD −123 to 1,230 [USD −145 to 1,042]) and increases QALYs by 0.11 (95% UI: 0.08–0.13). Additionally, the intervention reduces the T2DM event rate by 3,931 (95% UI: 2,946–5,016) cases per 100,000 person-years at risk compared to standard care over the 15-year model time horizon and reduces the event rates for all comorbidities. Detailed cost-effectiveness results and event rates are reported in Section 2.2 of the Supplementary Material alongside a completed Consolidated Health Economic Evaluation Reporting Standards 2022 (CHEERS 2022) checklist (Reference Husereau, Drummond and Augustovski66).
Discussion
We developed a microsimulation model for T2DM to estimate the long-term health impacts, costs, and cost-effectiveness of T2DM interventions. Our model improves on other T2DM simulation models by: (i) simulating progression from NGT to pre-T2DM to T2DM, allowing for the evaluation of interventions at any point in the care pathway, (ii) simulating the correlated progression of multiple individual-level T2DM risk factors (e.g., BMI, HbA1c, blood pressure) over time to improve modelling of long-term health outcomes, and (iii) simulating comorbidities both before and after patients transition to T2DM to capture the full benefits of preventive interventions. Risk equations and parameter values were sourced from the literature. We illustrated the application of the model by evaluating the cost-effectiveness of a potential lifestyle intervention-based T2DM prevention program in Alberta, Canada.
It is difficult for policymakers to compare different interventions when their cost-effectiveness is evaluated using various models that may vary in terms of model states, assumptions, or input data. Multi-use disease models are increasingly recognized as important in addressing this challenge (Reference Scholte, Ramaekers and Danopoulos67;Reference Wang, Pouwels and Ramaekers68). In the case of T2DM, several such models have been developed, but they predominantly focus on patients with newly diagnosed T2DM (Reference Hayes, Leal, Gray, Holman and Clarke5–Reference Sobotka, Cipriano, Coyle, Thiruchelvam and Shah11;Reference Hoerger, Hilscher and Neuwahl28). With the growing emphasis on preventive interventions, there is a need for multi-use T2DM models that can evaluate the potential health and economic outcomes of interventions across the full care continuum, from prevention to management. Our model fills this gap and can help policymakers make informed funding decisions, especially when comparing interventions aiming at preventing versus managing T2DM.
Internal and external validation of the model showed good model performance. In internal validation of the T2DM health state, the predicted KM cumulative failure probability of T2DM-related comorbidities and mortality aligned well with the observed KM cumulative failure probability of comorbidities and mortality in the UKPDS over 20 years. The mean absolute difference between predicted and observed values was under 1 percent across all comorbidities and mortality. The model predicted microvascular comorbidities slightly better than macrovascular, with mean absolute differences of 0.1 and 0.4 percent, respectively. Overall, our model demonstrates strong predictive performance for T2DM-related comorbidities and mortality within the UKPDS cohort.
For external validation, we compared the model predictions of the annual KM cumulative incidence of T2DM with the observed values in the 15-year follow-up study of the DPP trial. The model under-predicted T2DM incidence in the first 8 years, predicted reasonably well from years 8 through 11, and over-predicted from years 12 through 15 for both the placebo and the ILI arm. The over-prediction may stem from extrapolating findings from longitudinal studies. For patients with pre-T2DM, we applied an annual probability of 0.129 for a sudden increase in HbA1c, which results in progression to T2DM, based on longitudinal studies with mean follow-up durations of 5 and 9.5 years. Extrapolating these findings over a 15-year period may have led to the over-prediction of T2DM incidence over longer time periods. Additionally, the DPP diagnosed T2DM using fasting plasma glucose (FPG) and oral glucose tolerance test (OGTT), whereas our model used HbA1c to determine T2DM health states. While HbA1c is very specific, it is less sensitive than FPG and OGTT (Reference Punthakee, Goldenberg and Katz69). Some evidence indicates that HbA1c may identify more people as having T2DM than FPG in Canada; however other studies suggest HbA1c may identify fewer people as having T2DM compared to FPG or OGTT (Reference Punthakee, Goldenberg and Katz69), which could explain our model’s under-prediction of T2DM incidence from years 1 through 8. For these reasons, although the external validation indicates some discrepancies between the model predictions and the external dataset, interpretation of these discrepancies is difficult. External validation against other independent datasets that use HbA1c to diagnose T2DM would provide additional insights and will be an important future step in evaluating the model’s predictive performance. In addition, comparing our model’s performance with other published T2DM models through additional external validation activities, such as the Mount Hood Diabetes Challenge, will be an important area of future research (70).
Our study has limitations. First, we used the best available data in the literature to parameterize the IHE-DM. Nonetheless, the parameters and equations in the model are from multiple sources and are subject to the limitations of the source studies themselves. The characteristics of the populations in the source studies varied, potentially impacting the coherence of the various input parameters. Additionally, the primary source studies for predicting comorbidities, the FHS (1968–1987) and the UKPDS (1977–2002), are relatively dated. As standard care has improved over time (e.g., new medications and treatment devices for T2DM, statin use for high cholesterol), these equations may overestimate the risk for some comorbidities. Future work should endeavor to generate and validate newer values for model parameters and equations, ideally from the same population source for internal consistency.
Second, we used HbA1c to determine patients’ T2DM health state. Three tests can be used for T2DM diagnosis: HbA1c, FPG, and OGTT (34). We chose to use HbA1c because it is the measure of glucose control used in the UKPDS-OM2 equations (Reference Hayes, Leal, Gray, Holman and Clarke5), a primary source in our model. Using a different diagnostic test or a combination of tests may yield different results, either increasing or decreasing the occurrence of T2DM and related comorbidities. Consideration should be given to incorporating other measures of glucose control into future versions of the model.
Third, there can be a significant delay between the onset of T2DM and clinical diagnosis. Canadian data estimates that T2DM onset occurs 4–7 years before diagnosis (34). The IHE-DM does not take this into account. Once a patient’s HbA1c reaches 6.5 percent, they start the next model year with T2DM. From this point forward, it is assumed that the patient receives treatment for T2DM consistent with the treatment received by patients in the UKPDS. This likely results in an underestimate of the occurrence of comorbidities as patients are assumed to receive treatment for T2DM earlier than in practice.
The case study illustrates our model’s capabilities; however, a comprehensive cost-effectiveness analysis was beyond this paper’s scope. Nonetheless, it is important to acknowledge the case study’s limitations. We converted the direct medical cost of the ILI in the DPP to Canadian dollars using PPP. However, adapting US cost data to the Canadian setting is more complex than changing price weights, as US healthcare costs are generally higher than in Canada (64). We view the intervention cost in the case study as an upper bound for the cost of a similar program in Alberta. We estimated intervention compliance based on retention and participation in the US National DPP (Reference Cannon, Masalovich and Ng59;Reference Ritchie, Baucom and Sauder60). However, retention rates can vary depending on implementation strategy (Reference Rosella, Lebenbaum and Fitzpatrick50). Future analysis of a lifestyle intervention-based T2DM prevention program should include sensitivity and scenario analysis to understand the impacts of the intervention cost and compliance on cost-effectiveness.
Conclusions
This study presents the development and validation of a microsimulation model for T2DM. Whereas prominent T2DM models focus on newly diagnosed patients, our model contributes to the literature by simulating progression from normal glucose tolerance and/or pre-T2DM to T2DM. Our model also simulates the correlated progression of multiple T2DM risk factors over time and comorbidities before and after T2DM onset. These adaptations allow us to evaluate preventative interventions and more accurately capture the long-term impacts of interventions, filling an important gap in the evidence base. Our model can be used to inform future funding decisions and define target populations for T2DM interventions across the care continuum.
Supplementary material
The supplementary material for this article can be found at http://doi.org/10.1017/S0266462325100172.
Acknowledgments
The authors thank Arianna Waye for her support and feedback, and Himani Pandey for her contributions to background research and data collection. Information about the model code can be made available upon request to Megan Wiggins.
Funding statement
This work was supported by Alberta Health Services (no grant number).
Competing interests
The authors declare none.
Open access
