The cost-effectiveness of interventions targeting lifestyle change for the prevention of diabetes in a Swedish primary care and community based prevention program

Diabetes poses a huge burden on society both in human suffering and health expenditure terms. Diabetes is a common and costly disease that is expected to even continue to grow in prevalence and health expenditures over the coming decades. The International Diabetes Federation (IDF) estimated that approximately 52 million adults (20–79 years of age) had diabetes in the IDF EUR Region with a prevalence of 7.9% in 2014 [1]. In Sweden, the estimated number of cases was 426,800 among adults in 2014, amounting to a prevalence of 6.1% [1]. The number of diabetes-related deaths among Swedish adults was estimated at 2930 while the cost per person with diabetes in Sweden was 6310 USD [1]. It is estimated that incidence and prevalence of diabetes and consequently its costs in Sweden and Europe will increase in the future making prevention initiatives essential. An epidemiological model based on German data confirmed that unless enormous efforts are conducted into prevention programs, the number of persons with diabetes are expected to increase in the next two decades [2].

The three main types of diabetes are gestational diabetes, type 1 diabetes and type 2 diabetes (T2D); we focus on T2D in this analysis. T2D is the most common diabetes type and is characterized by insulin resistance and relative insulin deficiency. T2D develops over a long time period and is often undetected over years. During this time, people almost always first develop any of the pre-diabetic states, i.e., impaired fasting glucose (IFG), impaired glucose tolerance (IGT) or a combination of both (IFG and IGT). IFG is largely associated with an impaired insulin secretion and impaired suppression of hepatic glucose output, while IGT is mainly associated with muscle insulin resistance and defective insulin secretion [3].

Markov models are used to estimate future costs and effects of treatment options. They extrapolate current knowledge to be able to make decisions that have impacts on the future. Markov models may help to optimize the selection of individuals eligible for a focused intervention, such as specific lifestyle interventions [4].

Several trials have shown that prevention of T2D among individuals at high risk through lifestyle change is possible, effective and cost-effective, especially targeting diet and exercise to reduce weight [5‐11]. Lifestyle modification is considered the first choice of intervention for T2D prevention as it has a good cost- and treatment-effectiveness [12]. However, the long-term cost-effectiveness of a lifestyle intervention program in a Swedish setting is still not clear. On the other hand, policymakers need to know the cost-effectiveness of interventions to prevent T2D before implementing them at population level [13].

Various studies have estimated the effectiveness and cost-effectiveness of T2D prevention initiatives starting in the increased risk state IGT [14‐18]. Some studies focused on the effect of screening for T2D [18, 19]. Other studies have investigated the cost-effectiveness of healthy individuals (NGT) but only IGT as pre-diabetic state [20, 21]. However, no study to our knowledge has estimated the cost-effectiveness of T2D prevention strategies from a population-based perspective, including healthy individuals and also considering IFG and IGT as two distinct pre-diabetic states. Previous research has, nonetheless, shown that IFG and IGT are different regarding risk of developing a worsened health state and quality of life [22‐25]. Norris and colleagues consequently underlined that further research is needed to define the duration of the pre-diabetes phase and to identify measurable risk factors for progression to T2D and its complications [26].

The objective of this study was to estimate the cost-effectiveness of a T2D prevention initiative targeting weight reduction, increased physical activity and healthier diet in persons in pre-diabetic states by comparing a hypothetical intervention comparable to the Finish Diabetes Prevention Study (DPS) versus no intervention in a Swedish setting.

The model consisted of six different, mutually exclusive states (Fig. 1). The states described the possible development of T2D through a simulation of hypothetical persons using average values of input parameters in combination with assumed distributions. The Markov model is characterized by a population with mean values for risk factors that are adjusted by the intervention effect.

A hypothetical person in a healthy state, NGT, could develop IFG, IGT or IFG and IGT, the pre-diabetic states. From any of the pre-diabetic states, the person could develop T2D. In return, a person with T2D could move to a pre-diabetic state or they could move from a pre-diabetic state to NGT. Direct changes between the pre-diabetic states were not possible in the model. Even though such changes are plausible, they have not been estimated in this model for simplicity reasons and because the available number of individuals for estimating the risk equation was not sufficient to build robust estimates stratified by risk factors.

The transition to move back from T2D to any pre-diabetic state was kept in the model, even though a person who once entered the T2D state is not comparable with the one who never had T2D. On the other hand, this change is possible and lifestyle interventions could have an impact on this transition. In addition, some people who had once been diagnosed with T2D might attain healthy values and will not have the same costs and complications as an average person with T2D.

At any time, the person could die and consequently move to the absorbing state death.

The length of one cycle was 1 year. A lifetime horizon was applied. For cost estimation, the societal perspective was applied.

First, analyses were done deterministically using all input parameters as fixed values. Deterministic modeling always returns the same results using the same input parameters.

Second, analyses were conducted probabilistically based on Monte Carlo simulation using defined statistical distributions of input parameters to simulate results. Probabilistic modeling was used to investigate the impact of statistical scattering of the input parameters on the model results, which elicits the uncertainty of the results. One thousand replicates were performed based on underlying probability distributions for input parameters and the mean of the costs and QALYs was calculated. Percentile confidence intervals were estimated for cost and QALYs. To estimate the percentile confidence intervals, the 2.5th and 97.5th ordered values were used as lower and upper limits of the confidence interval, respectively.

Six hypothetical scenarios were run, for men and women at younger (30), middle (50) and older (70) age, to describe the cost-effectiveness of the intervention for both sexes at different age groups. Cost and QALY differences between no intervention and intervention, and incremental cost-effectiveness ratios (ICERs, cost per QALY gained) were estimated and visualized in the cost-effectiveness quadrants (cost-effectiveness planes, CE planes) as scatterplots of the Monte Carlo simulation results. The cost-effectiveness acceptability curves (CEA curves) displayed the probability that the data are consistent with a true cost-effectiveness ratio falling below a specified threshold value. CE planes and CEA curves were estimated for all six sex-age scenarios to enhance readability of model results. Cost-effectiveness planes and cost-effectiveness acceptability curves assist in reading the distribution of the probabilistic outcomes and the probability of whether the intervention is cost-effective assuming different threshold values.

Running the model deterministically (stable input parameters) with different scenarios of age and sex showed the effect that the intervention was more costly than no intervention but also gained higher QALYs (Table 6). The higher QALYs for the intervention group showed that the intervention was effective compared to no intervention. The ICER was highest for men at older age (men, age 70: 9215 € per QALY gained) and lowest for women at younger age (women, age 30: 3833 € per QALY gained). The ICER was more favorable for women compared to men and for younger age compared with older age (Table 6). All ICERs were cost-effective assuming a cost-effectiveness threshold of 50,000 € per QALY gained.

Running the model probabilistically 1000 times while assuming specified statistical distributions elicits the uncertainty of the results. The mean cost difference was higher than the median cost difference for all scenarios, while the difference in effect between mean and median values was very small (Table 7). The difference between mean and median in cost was due to the right skewed distribution of the data. The percentile interval for the cost difference was large, ranging from cost-saving results to higher cost for the intervention (Table 7). This large range derived from the large assumed standard deviation of the cost, i.e., standard deviation being equal to the mean value of the costs. The ICERs derived by probabilistic analysis were comparable to the deterministic results (Tables 6, 7).

The results of this analysis showed that the described hypothetical intervention was very cost-effective for men and women at all ages. The cost-effectiveness planes showed the distribution of the cost and outcomes and the cost-effectiveness acceptability curves underlined that the probability that the intervention was cost-effective was high assuming a threshold value of 50,000 € per QALY gained. Assuming that the described hypothetical intervention and its effect was similar to those of the DPS, this analysis showed what would happen if a DPS-like intervention were implemented in Västerbotten.

Several studies have estimated the cost-effectiveness of lifestyle intervention. However, none focused on the costs and effects of NGT, IFG, IGT, and IFG and IGT on the development of T2D. In addition, it is difficult to compare the results of different cost-effectiveness studies with each other as methods, health care systems, costs included, perspectives, lifestyle interventions and other factors differ [13]. Nonetheless, in the following, the results of this study were compared with other modeling studies focusing on T2D prevention through lifestyle change.

Similar lifestyle interventions for preventing T2D were also cost-effective [14, 15, 19, 20, 42], cost-saving [16‐18, 21, 43, 44] or had mixed results [45].

A Markov Monte Carlo simulation on the cost-effectiveness of screening for T2D found that early detection and prevention of T2D by either lifestyle intervention or metformin may be cost-effective, with 562.54 € per QALY for lifestyle intervention [19]. Forster and colleagues analyzed the cost-effectiveness of two diet and exercise interventions, showing that those interventions have the potential to be cost-effective [42]. Smith and colleagues estimated that the DPP-like intervention cost less than $20,000 per QALY gained in approximately 78% and less than $50,000 per QALY gained in approximately 86% of model iterations using cost-effectiveness acceptability curves [14]. An estimation of the cost-effectiveness of the DPP showed that the intervention could provide a financial return on investment for private payers and long-term benefits for Medicare with an estimated cost of $1288 per QALY gained at age 50 [15]. A modeling study based on an example from Germany estimated that lifestyle interventions for primary prevention of T2D have a high potential to be cost-effective [20].

In a modeling study examining the effect of lifestyle intervention for preventing T2D in an Australian setting, the program was dominant over no treatment (cost- and life-saving) [21]. Häussler and Breyer estimated that a German lifestyle intervention to reduce obesity among obese people without increased glucose tolerance can pay for itself from the point of view of a health insurer [43]. The evaluation of the European Community-based project “10,000 Steps Ghent” showed that the intervention, which focused on increased physical activity, was dominant compared to no intervention [44]. A cost-effectiveness evaluation of the Diabetes Prevention Program Outcomes Study estimated that lifestyle intervention among people with pre-diabetes produces long-term social benefits which exceed intervention costs [16]. A simulation model based on Swedish data estimated that lifestyle intervention programs save a mean total of 1853 € per patient [17]. Liu and colleagues estimated that programs to prevent T2D using lifestyle interventions were cost-saving in a Chinese setting [18].

This study has several strengths. Both the health utility weight and the risk equations that define the transition probabilities were estimated from the same source, the Västerbotten Intervention Program (VIP) and for NGT, IFG, IGT, and T2D. As other studies only included IGT instead of IFG and IGT, and very few studies considered NGT, this analysis provides essential information on the prevention of T2D and its pre-states. Several sensitivity analyses including probabilistic sensitivity analysis enable the reader to judge the soundness of the model and its input parameters. In addition, CEA curves and CE planes assist the reader in judging the stability of the results and the change with assumed threshold value.

On the other hand, some limitations are to be mentioned. As the focus of this study was the prevention of T2D, this model did not differentiate between T2D with and without complications. However, neither complications nor prevention of complications were in the focus of this analysis. Hypothetical persons who just entered the T2D state would be underestimated in HUW and overestimated in cost terms. Over time, the complications increase and the cost increases, while the HUWs decrease. Hypothetical persons who were in T2D for a long time would be underestimated in cost estimates and overestimated in HUWs in this model. However, as a lifetime time horizon was used here, these differences disappear or are marginal since sometimes over- and sometimes underestimations would balance using the average value. In addition, due to the lack of information on cost of IFG compared to IGT, the cost within the pre-diabetic states was assumed to be the same. Further, data for transition probabilities and health utility weights for T2D with or without complications were not available. However, it has to be kept in mind that the cost of T2D is strongly associated with complications, while the combination of micro- and macrovascular complications incurs the highest cost [45, 46]. Further, as the population of inhabitants in Västerbotten is fairly small (approximately 260,000), the robustness of the mortality values could be questioned. However, cross-checking the mortality in Västerbotten with the mortality in entire Sweden showed no significant difference. As input parameters for the model were taken from different sources, this could limit the internal validity of the model. However, it was impossible to extract all necessary information for the model from the same source. Nevertheless, most data were taken from the Västerbotten population or the Diabetes Prevention Study in Finland. In addition, the internal validity of the model could be reduced as different studies were used to describe the effect of the intervention. Further, IGT was associated with lower HUWs compared to IFG and IGT (Table 4), which seems counter-intuitive. Having IGT in the described risk profile (Table 3) seems to reveal worse HUWs for IGT compared to IFG and IGT. Nonetheless, the HUWs of all pre-diabetic states ranged between NGT and T2D and the differences were too small to have huge impacts on the final results.

All models are based on a certain number of assumptions, as not all data are available. Some assumptions in this model were based on data of the VIP and some assumptions on results from other studies. The main assumptions were: the cost of all pre-diabetic states was the same and was 46% of those of T2D; the development of pre-diabetic states after T2D was possible; the effect of the intervention only lasted for 7 years, with a linear decrease over time; the effect of lifestyle change was according to the Diabetes Prevention Study; there was no increased risk of dying due to pre-diabetic states; the chosen statistical distributions for the probabilistic estimations were used; and the standard deviation equaled the mean cost. The lack of data is indeed one of the reasons modeling is needed. It is recognized that modeling studies are valuable in the decision-making process where real-life data or intervention effectiveness is insufficient or absent [21]. Therefore, informed and well-guided assumptions need to be made for decision-making.

This analysis has shown that model structure is an important issue of modeling. The possible transition between T2D and any pre-diabetic state could be argued. Markov models ignore information on what has happened in the cycles before and, therefore, do not differentiate in any pre-diabetic state between those who moved from T2D to a pre-diabetic state and those in the pre-diabetic state who never had T2D. On the other hand, this transition is technically and physiologically possible and, due to the limitation that this model does not differentiate between less severe T2D and T2D with complications, it allows separation of those people with low-profile T2D from those in more severe states. It was decided to keep this transition, as it is plausible and underlines the dynamics in the model, and to have its influence checked in a sensitivity analysis. The sensitivity analysis showed that this assumption does have a large impact on the result.

Clinical trials have largely targeted IGT, while also other trials have shown that prevention in individuals with IFG can be effective [47]. The effectiveness in our model, i.e., influence on weight, physical activity and diet, was also taken from clinical trials, which only included IGT as the pre-diabetic state. However, the results of the effectiveness in our model were translated to effects based on risk equations for transition probabilities [22] and health utility weights [23]. Therefore, the influence of IFG, IGT, and IFG and IGT were modeled separately based on distinct functions.

Possible transitions between the pre-diabetic states were not modeled, as the population used in the model was too small to lead to robust estimates. On the other hand, such transitions would not have a huge impact on the outcome of the model as both the effect (HUWs) and the costs were very similar between pre-diabetic states.

Even though QALYs gained were low, other benefits that derive from the intervention should not be omitted. Prevention of pre-diabetic conditions and T2D as well as the adoption of a healthy lifestyle also have positive impacts on other diseases such as cardiovascular diseases and some cancers. Therefore, the positive effect in terms of prevention of cost due to cardiovascular diseases and cancer of this model is, therefore, likely to be underestimated.

The intervention was assumed to only have effects on BMI, nutrition and physical activity, and only last for 7 years with a linear decrease. For some people, lifestyle intervention programs could have a longer lasting effect than just 7 years. The Chinese Da Qing study, for example, showed an effectiveness of lifestyle intervention even after 20 years [48]. However, sensitivity analyses have shown that doubling the duration of time of the intervention showed a small impact on the results, which can be explained by the highest difference in effect between intervention and no intervention being in the 1st year and thereafter decreasing substantially within a few years.

This analysis aimed to provide indicators for long-term allocation of healthcare resources in Sweden and other European countries. In line with other studies, the prevention of T2D through lifestyle change can be very cost-effective. Targeting weight reduction through increase of physical activity and change in diet would have a strong implication on future T2D risk. The return on investment in T2D prevention initiatives in a group setting is high. This analysis shows that implementing a DPS-like intervention in Västerbotten would have a high potential to be cost-effective.

However, one has to be aware that targeting the T2D epidemic only at individual level may likely not suffice. Interventions and actions at policy and environmental level need to supplement individual interventions for a sustainable reduction in T2D incidence [49]. The rising incidence is driven by a wide mix of factors at different levels, such as individual susceptibility, food supply, transportation use, climate change or the economic situation of the country [50]. Targeting individual interventions is one important step towards slowing the progression of T2D but should not be the only initiative.