Background
A pandemic of type 2 diabetes (T2DM) is affecting diverse populations worldwide [
1]. While genetic factors are important precursors [
2,
3], the rapid emergence of T2DM since the middle of the last century parallels a rise in obesity rates associated with unprecedented lifestyle changes affecting caloric intake and physical activity [
1,
3]. In addition, growing numbers of people with diabetes at each end of the life span attest to other contributing factors. Thus, T2DM during childhood and early adulthood is often related to gestational diabetes, which heightens diabetes risk for both affected mothers [
4] and their offspring [
5,
6]. Among aging adults, increasing numbers of baby boomers entering their seventh decade, and improved survival of people with diabetes, are also driving an increase in T2DM prevalence [
1]. Finally, elevated T2DM rates affecting Indigenous peoples [
1,
7] also highlight the importance of social determinants of health [
8,
9] and, in some groups, sex differences [
7] in the genesis of this disorder.
Chronic kidney disease (CKD) is a serious complication of T2DM and can lead to end stage renal disease (ESRD) [
10]. While ESRD affects <1% of prevalent diabetic non-Aboriginal Canadians [
11], diabetes-related ESRD (DM-ESRD) – ESRD among people with diabetes caused by diabetic and non-diabetic factors – is the leading cause of ESRD in Canada, accounting for more than 35% of incident ESRD cases [
12]. Importantly, diabetic Aboriginal people in Canada are at higher risk for developing DM-ESRD than their non-Aboriginal peers [
11,
13,
14]. In addition, because Aboriginal people in Canada typically develop diabetes at a younger age in part because of the inter-generational impact of diabetic pregnancies experienced more frequently by Aboriginal women [
5,
6], they are more likely to live long enough to transition through earlier stages of CKD and to develop DM-ESRD [
15,
16].
DM-ESRD is a devastating disorder for affected people and their families, and renal replacement therapy (RRT) with peritoneal dialysis (PD), hemodialysis (HD) and renal transplantation consumes a disproportionate share of health care resources [
17]. Furthermore, while the incidence of DM-ESRD in Canada has stabilized somewhat since the early 1990s, the prevalence of both T2DM and DM-ESRD continues to rise [
7,
13]. Forecasting DM-ESRD numbers, while taking into account an evolving T2DM epidemic and population demographics, would allow prediction of financial, human resource and facility requirements. Moreover, simulating clinical scenarios could provide insight into how individuals and groups progress through CKD stages and health care processes, and how this influences both the health and cost burden of DM-ESRD. Accordingly, we sought to examine the potential of dynamic computer modeling [
18] in better understanding the epidemiology of DM-ESRD. Our specific objectives were: 1) to develop an agent-based model that can project case numbers and treatment costs of First Nations and non-First Nations people with DM-ESRD in Saskatchewan from 1980 to 2025; and 2) to investigate the potential long term impact of simulated clinical interventions on the DM-ESRD epidemic.
Discussion
This study describes the first agent-based model (ABM) designed to project case numbers, costs and survival of people with diabetes-related ESRD (DM-ESRD) in Canada. Using the province of Saskatchewan as its setting, this ABM closely reproduced historical trends for the incidence, prevalence and costs of DM-ESRD from 1980 to 2011, and is able to project this information into future decades. The model simulated events and activities for a population with diabetes, including year of diabetes diagnosis, progression to ESRD, type of RRT, and death. Furthermore, it considered individual patient characteristics, stages of CKD, RRT modalities, kidney transplant assessment and waiting list processes, and costs. Using this information, the model can aid in resource planning for managing the fast-growing DM-ESRD population in the province. Furthermore, the model can be used by policy makers to simulate “what if” scenarios that may provide insights into the dynamics of a diabetic person’s progression through kidney disease stages and health care processes that are otherwise not possible to achieve. Although there are two published models from Canada [
34,
35] and two from the United States [
36,
37] that projected various elements of case numbers, costs and complications of diabetes into the future, three of the four [
35‐
37] used Markov models. These consider only the latest features of populations of interest and do not trace individual trajectories, or allow for calibration to or validation against longitudinal individual-level data. Importantly, none of the four models addressed ethnicity-based disparities in diabetes and its complications experienced by Indigenous peoples.
Our ABM projected a 2.5 times increase in DM-ESRD prevalence from 2012 to 2025, with First Nations people consistently accounting for approximately 30% of cases. The latter is about twice the current proportion of First Nations people within Saskatchewan’s total population. By 2025, the model projects that there will be almost 1300 prevalent cases with DM-ESRD requiring RRT in the province, with total absolute costs of almost 90 million dollars per year. Barring changes in clinical practice, close to 1000 of these individuals will be receiving hemodialysis, the most expensive form of RRT. Most of the remaining people with DM-ESRD will be on peritoneal dialysis, with fewer than 100 renal transplant recipients (see below). These projections have sobering implications not only for future RRT resource needs but also for the disproportionate demands of RRT on the provincial health care system budget [
17].
While the model was particularly close to reproducing historical trends for DM-ESRD from 1980 to 2005, it displayed a modest divergence from historical data by gradually overestimating prevalent case counts from 2005 to 2011. This was primarily due to the model’s overestimation of both HD and PD patients. This overestimation of dialysis patients was partially offset by the model’s underestimation of renal transplant recipients. Thus, it appears that the model is not carrying out sufficient numbers of renal transplants among those on dialysis. We continue to examine possible factors in the model that could explain this anomaly. One possible reason is that we did not include incident diabetes cases occurring under age 20, a cohort that experiences higher transplant rates because of its young age and healthier status. This problem will be corrected in updated iterations of the model.
The ability to simulate “what if” clinical scenarios using this ABM is a rapid, powerful and economic means of projecting scenarios of interest and testing strategies to mitigate current trends. We have provided two hypothetical scenarios to illustrate the potential of this feature. While clinically implausible, these scenarios can aid in understanding the implications of more modest and plausible scenarios. In the first, we imagined that diabetes mellitus was suddenly and completely preventable. Thus, no new diabetes cases entered the model after 2005. Despite this dramatic intervention, prevalent cases of DM-ESRD (and costs) continued to rise for more than 10 years and did not begin to slowly decrease until 2019. This illustrates the tremendous inertia within the system, and the resource demands associated with caring for existing people with diabetes. It also highlights the potential value of the model in evaluating the likely impact of promising interventions.
In the second scenario, all diabetic people with newly diagnosed DM-ESRD immediately received a renal transplant, and graft losses were rapidly replaced. By 2025, this resulted in an almost 70% increase in prevalent cases of DM-ESRD but with a 30% reduction in annual costs. The increase in prevalence is due to the substantially lower mortality experienced by renal transplant recipients compared to people on dialysis (although it should be noted that an unselected group of DM-ESRD patients receiving kidney transplants would have higher mortality rates than those currently selected for transplantation and would likely also experience a decrease in graft survival). In contrast, the lower costs reflect the fact that caring for renal transplant recipients is significantly cheaper than dialysis treatment, especially following the first year of a successful transplant. Once again, while this scenario is completely unrealistic and overly simplified, it demonstrates the potential of the ABM for helping health care systems in considering different options for the allocation of health care resources.
Strengths of the model include its ability to project DM-ESRD incidence and prevalence by age, gender and ethnicity, the first model in Canada to consider all three parameters simultaneously. Second, the foundational data for the model is of high quality [
24‐
26], and itemised cost information of RRT is taken from published research conducted in the neighboring province of Alberta [
28]. Third, model output was extensively validated against historical data from multiple sources including over 25 years of longitudinal data [
26]. Fourth, the current model has demonstrated its capacity for straightforward integration with other Anylogic models and an ability to incorporate a System Dynamics Model of drivers for diabetes [
23]. Moreover, Anylogic software offers an animated presentation layer that allows stakeholders and other users to understand structures of the system, and provides tools to help policy makers simulate different policies and interventions in a timely manner. Finally, our Saskatchewan DM-ESRD model is equipped with a database which can record demographic information, treatment history, transplant assessment and waitlist, and cost data for every patient in the model population.
In addition to excluding people with childhood-onset diabetes, our model also has a number of limitations associated with model structures and input parameters. First, all DM incident patients in the model either developed ESRD or died before developing ESRD. However, since some patients would have instead left the province to live elsewhere, the model may have overestimated DM-ESRD incident patients. Second, the model selected transplant candidates based solely on age, whereas in reality patient selection is based on additional factors. Without considering these in the model, patients being transplanted might have died sooner or later than in reality. Third, input on cost data was taken from a Calgary study and there might be significant cost differences for managing DM-ESRD patients between Alberta and Saskatchewan. Fourth, the Cox Proportional Hazards model that was used for dialysis patients was conducted on patients from 1999 to 2008, and we adjusted the values for different time periods by adding a calendar covariate. This might have biased mortality rates for some cohorts of dialysis patients in the model. Fifth, we used a scaling factor for extrapolating the Saskatoon Diabetes Population to the Saskatchewan Diabetes Population starting in 2006. However, the proportion of First Nations people in the city of Saskatoon city might be different than the proportion of First Nations people in the province. Also, because the scaling ratio for 2006 was used for the years 2006 to 2025, it likely does not reflect population changes occurring over the later time period. Finally, many other rates used for projection are also based on years in which the values are known, and it is likely that those rates will change in the future.
Conclusions
Over the past decade, system science methods have been increasingly applied to study problems in the public health domain to provide insights not evident using traditional approaches [
18]. Dynamic models such as that presented here can capture the system wide impacts of complexity in a system that not only includes direct elements of the disease being studied but also related features such as demographics, risk factors, economic considerations, facilities and equipment, human resources, policy, budgets, and transportation. Dynamic models can be a strong learning and communication tool through visualization of diverse components in the system, and by allowing comparisons between the simulated behavior resulting from hypothesized relationships and empirical evidence. Furthermore, models can inform researchers as to which missing data could contribute the most value to decision making and understanding of system evolution. Finally, “what if” simulations using the model can help decision makers to evaluate policies and interventions that might be difficult to carry out in the real world because of ethical concerns, and time and resource constraints. Furthermore, such simulations can be used to identify the leverage points in the system and identify cost effective strategies.
In this paper, we have shown how the strengths of system science methodology can be applied to a serious public health issue in Canada by developing a dynamic agent based model of DM-ESRD. By projecting rates and costs of DM-ESRD into future decades while considering a vast array of individual characteristics, and simulating “what if” scenarios, we have shown the immense potential of this approach within a provincial health care system. While this particular project is confined to Saskatchewan, its elements and structure are adaptable and transferrable to other jurisdictions.