Targeting of the Diabetes Prevention Program leads to substantial benefits when capacity is constrained

Abstract

Objective:

Approximately 84 million people in the US have pre-diabetes, but only a fraction of them receive proven effective therapies to prevent type 2 diabetes. We estimated the value of prioritizing individuals at highest risk of progression to diabetes for treatment, compared to non-targeted treatment of individuals meeting inclusion criteria for the Diabetes Prevention Program (DPP).

Methods:

Using microsimulation to project outcomes in the DPP trial population, we compared two interventions to usual care: (1) lifestyle modification and (2) metformin administration. For each intervention, we compared targeted and non-targeted strategies, assuming either limited or unlimited program capacity. We modeled the individualized risk of developing diabetes and projected diabetic outcomes to yield lifetime costs and quality-adjusted life expectancy, from which we estimated net monetary benefits (NMB) for both lifestyle and metformin versus usual care.

Results:

Compared to usual care, lifestyle modification conferred positive benefits and reduced lifetime costs for all eligible individuals. Metformin’s NMB was negative for the lowest population risk quintile. By avoiding use when costs outweighed benefits, targeted administration of metformin conferred a benefit of $500 per person. If only 20% of the population could receive treatment, when prioritizing individuals based on diabetes risk, rather than treating a 20% random sample, the difference in NMB ranged from $14,000 to $20,000 per person.

Conclusions:

Targeting active diabetes prevention to patients at highest risk could improve health outcomes and reduce costs compared to providing the same intervention to a similar number of patients with pre-diabetes without targeted selection.

Introduction

A number of studies have found that intensive lifestyle intervention, focused on maintaining healthy diet and exercise, such as the Diabetes Prevention Program (DPP), or administration of metformin are effective at delaying or preventing diabetes in individuals with pre-diabetes.^1,2 Although these interventions have been found to be cost-effective on average,^1,2 treating all eligible individuals the same way may not be efficient or feasible. The DPP lifestyle intervention is intensive and requires a major commitment from the patient: 6 months to a year of classes, counselling sessions, and continued support in changing lifestyle habits. Despite its proven efficacy, health systems are still working on implementing and ramping up the DPP lifestyle intervention, partnering with community-based organizations, serving tens of thousands of people, but still short of capacity to serve the millions that are eligible.^3–6

Substantial differences in the extent to which individuals can potentially benefit from DPP also casts doubt on the optimality of its universal application. Estimated 3-year risk of developing diabetes in the DPP-eligible population ranges from 1% to more than 90%. The distribution for this risk is skewed, so while its mean is 21%, risk modeling suggests that it is less than 15% for half of the population included in the DPP trial,^7,8 and maybe lower still in the general ‘prediabetic’ population not undergoing the stringent selection in the trial. Because individual risks vary considerably, intervention benefits for individuals also vary. Moreover, while on average both metformin and lifestyle intervention improved outcomes compared to usual care, many DPP trial participants at low risk received no benefit from metformin and only a modest absolute benefit from lifestyle intervention.^7,9 Targeting only those individuals at highest risk might accrue a considerable proportion of the potential population benefits, while substantially reducing costs and treatment burden. Importantly, clinical risk prediction models based on combining demographic and clinical characteristics, such as age and sex, and their lab values (e.g., hemoglobin A1c, fasting glucose, triglyceride, and others) can identify higher risk individuals.^7,10 Several economic studies of diabetes prevention have previously applied a risk-stratified approach, but were limited by using average intervention effectiveness rates across all risk groups in the treated populations.^11–14 To our knowledge, no previous economic study has fully incorporated the heterogeneity of treatment effect of DPP by diabetes risk.^15–17

This study aimed to estimate the value of using individualized risk information, based on multivariable modeling and individualized analysis of intervention effectiveness, to target use of either (1) intensive lifestyle intervention or (2) metformin administration for diabetes prevention versus “usual care” (i.e., no additional intervention).

Research Design and Methods

Value comparisons

To estimate the value of targeting, we used the “net monetary benefit” (NMB) metric, or the difference between the monetized value of health and the incurred cost of health care. The monetized value of health is the product of the quality adjusted life expectancy and the value of each quality-adjusted life year (QALY).^18,19 In our base case we used the willingness to pay (WTP) value of $50,000/QALY, per recommendations of the Second Panel on Cost-Effectiveness in Health and Medicine.^19,20

Extrapolating the DPP clinical trial cohort to the approximately 84 million members of the US population satisfying the DPP inclusion criteria, we compared the NMB of either active intervention (lifestyle or metformin) to usual care: (1) treating all population members the same way (either treating all members or treating no members) based on the average NMB of treatment for the entire cohort; and (2) targeting for treatment only those population members whose individual NMB is expected to be positive.

Our second analysis assumed that resource constraints permitted active prevention (either lifestyle or metformin) for only 20% of the US population meeting the DPP criteria (approximately 17 million individuals). We compared administering active prevention to (1) 17 million individuals selected at random from this population and (2) the 17 million individuals projected to gain the most from active prevention.

Cost and health benefit estimation

We constructed a decision-analytic microsimulation model to project individual care costs and health outcomes for each of the two active prevention strategies and for usual care.

For each individual, we used a risk model to predict probabilities of diabetes onset at different ages when receiving the lifestyle intervention, metformin, or usual care. We simulated individuals moving through the model’s health states, starting from the initial state of prediabetes. Over their lifetime, they might progress to diabetes and develop diabetes-related complications, such as heart disease, stroke, renal impairment, blindness, or neuropathy (Figure 1). An individual accrues associated costs and QALYs based on each specific health state and complication they experience over time (Table 1). By averaging over multiple replications, we obtained expected total lifetime costs and QALYs for each person under each prevention strategy.

Figure 1: Overview of Simulation Model.

Note: Further methodological details on the simulation models in Appendix 1

Table 1:

Study cohort description and simulation model assumptions

DPP trial cohort at baseline (n=3081)	Mean or number	SD or %
Age, years	50.6	9.0
Female sex	2053	66.6%
Race/ethnicity
White	1768	57.4%
Black	644	20.9%
Hispanic	508	16.5%
Family history of diabetes	2127	69.0%
Current smoker	216	7%
Diagnosis of hypertension	835	27.1%
BMI	33.5	5.8
Hemoglobin A1c, %	5.9	0.5
Triglycerides, mg/dL	162.9	93.5
Fasting plasma glucose, mg/dL	107.2	7.7
History of high blood glucose	614	19.9%
Waist to hip ratio	0.9	0.1
Waist circumference, cm	105.0	14.6
Simulation Model Input Assumptions
Costs of DPP interventions (Herman et al.2)	Initial 4 years	Maintenance annual cost
Lifestyle intervention	$4,000	$140
Metformin	$1,416	$160
Costs of diabetes treatment and complications*	One-time event	Ongoing
Baseline annual cost with no complications		$2,315
Nonproliferative retinopathy	$103	$103
Macular edema or proliferative retinopathy	$1,101	$103
Blindness	$2,951	$2,951
Microalbuminuria	$437	$437
Proteinuria	$748	$748
End-stage renal disease with hemodialysis	$99,046	$99,046
End-stage renal disease with renal transplant	$138,071	$44,331
Clinical neuropathy	$511	$511
Amputation	$42,929	$1,500
Angina	$8,282	$2,139
Myocardial infarction	$41,744	$2,307
Percutaneous transluminal coronary angioplasty	$8,282	$2,139
Coronary artery bypass graft	$60,685	$2,307
Myocardial infarction with coronary artery bypass graft	$60,685	$2,307
Congestive heart failure	$34,635	$7,620
Ischemic stroke	$55,278	$18,448
Acute metabolic complication: Hypoglycemia requiring hospitalization	$16,991
Health utilities
Impaired glucose regulation	0.77	Palmer et al.36
Diabetes (intercept, adjusted by disutilities for complications and comorbidities, not shown)	0.69	Zhou et al.25

^*Full details on the Michigan Model for Diabetes structure and input assumptions are available in the model manual (accessed at http://diabetesresearch.med.umich.edu/peripherals/DiseaseModel/MDRTC%20Diabetes%20Model/UserManual_MichiganModel_for_Diabetes_ver2.pdf)

Note: all costs shown in 2014 US dollars

The analysis applied a healthcare sector perspective with costs and health benefits discounted 3% annually per recommendations of the second Panel on Cost-Effectiveness in Health and Medicine.²⁰ We measured costs in 2014 US Dollars.

Study population and data sources

We based the simulated population on the 3-year randomized DPP trial cohort of 3081 individuals who had impaired glucose regulation and were eligible for diabetes prevention programs, but had not progressed to type 2 diabetes (Table 1).²¹ Starting with the baseline characteristics of the simulated individuals from the DPP patient-level data sets, we projected time-varying measures, e.g., BMI, based on trajectories observed in the trial. We used published estimates for costs and health-related quality of life for the pre-diabetes and diabetes health states and associated complications, mortality rates from the age- and sex-specific US life tables for years 1981–2013²², and age- and sex-specific medical costs from the Medical Expenditures Panel Survey (MEPS) dataset for 2015 (Table 1).²³

To estimate lifetime costs and QALYs for individuals who develop diabetes, we used the Michigan Diabetes Research Center model^24,25 which assumes standard of care diabetes treatment, with its associated costs, outcomes and adherence rates, and incorporates costs and health-related quality of life impact of adverse outcomes associated with diabetes, including microvascular and macrovascular complications and mortality. The Michigan Diabetes model version 2.1, released in 2015, was the most recent update at the time of this study²⁶ (Appendix 1).

We performed statistical analysis in SAS and R and implemented the decision model in Excel.

Prediction of progression to diabetes

To predict the individual risk of developing diabetes and time to progression in adults with impaired glucose regulation, given baseline characteristics and a pre-specified prevention strategy, we used a Cox proportional hazards model. Starting with a previously published short-term prediction model⁷ derived from individuals enrolled in the DPP trial and their 3-year outcomes, we extended the model using long-term patient level follow-up data from the Diabetes Prevention Program Outcomes Study (DPPOS), a 7-year open-label follow-up of 2,766 participants from the DPP trial.²⁷ The original model included 7 independent variables (height, waist to hip ratio, waist circumference, fasting plasma glucose, history of high blood glucose, and serum levels of HbA1c and triglycerides) selected using significance level p≤0.1 in multivariable models to predict the 3-year risk of diabetes, and also incorporated the effects of the diabetes prevention intervention and the interaction between linear predictor of 3-year diabetes risk and metformin therapy to account for variation in the treatment effect by risk (higher relative benefit for higher risk patients). The lifestyle intervention had a consistent relative effect across all risk quantiles. The final model had a good discrimination with a c-statistic of 0.73. Its performance was externally validated against the Framingham diabetes model,¹⁰ described in detail previously.⁷ (Appendix 2) Unlike prior analyses, our model incorporates the patient-level heterogeneity of treatment effect, using diabetes risk-based hazard ratios to estimate treatment effect at each level of diabetes risk, rather than population-wide average hazard ratios. This approach affects the results due to statistically significant interaction between diabetes risk (using the multivariable linear predictor) and the treatment effect of metformin, whereby higher risk patients have a greater risk reduction.⁷

Published literature suggests that treatment efficacy for diabetes prevention varies over time with substantial uncertainty.^16,28 To estimate efficacy beyond year 4, we developed two alternative scenarios – a conservative, or short-term treatment effect scenario, which assumes that beyond the first 48 months interventions do not provide any reduction in hazard of diabetes onset; and a second, more optimistic, long-term treatment effect scenario assuming a constant hazard reduction for diabetes onset over time.

Sensitivity analysis

For each treatment effect duration scenario we tested the impact of different values of a QALY, different levels of discounting, and the utility values assigned to the pre-diabetes and diabetes health states (Appendix 3).

Additional sensitivity analyses tested alternative assumptions for mortality rates, increasing general population estimates for age- and sex-specific mortality to account for potentially higher mortality risk in the population with pre-diabetes. Besides examining a constant increase across the entire cohort, we also tested a differential adjustment with mortality rates positively associated with risk of diabetes. We applied mortality rates similar to general population for individuals at lowest diabetes risk and added a penalty equivalent to being 5 years older for individuals with pre-diabetes in the highest diabetes risk quintile with a reduced intermediate penalties for individuals in the intermediate risk quintiles, linearly extrapolated from 5 to 1 year.²⁹ This adjustment has the net effect of attenuating the contrast between the pre-diabetic state and the diabetic state, particularly for those at the higher risk end of the pre-diabetic spectrum and minimizes the relative benefit of diabetes prevention from the pre-diabetic state.

Finally, medical literature indicates that some individuals experience significant “treatment burden” of metformin—i.e. the impact of taking a daily medication on health-related quality of life and adverse events from metformin preventive treatment.^30,31 We examined how this treatment burden could affect the overall net benefits of metformin by applying an annual absolute reduction of health utility value, or disutility, ranging from 0.002–0.05 based on published estimates.^30,31 In the absence of patient-level data linking patient preferences with other characteristics and outcomes, our sensitivity analysis applied the disutility uniformly across the modeled population.

Results

Treating the entire eligible population

When compared to usual care, metformin delayed the onset of diabetes by a mean of 1.3 (SD 0.96) years in the short-term treatment effect scenario and 2.0 (SD 1.3) years with the long-term effect scenario, while lifestyle delayed it by a mean of 3.1 (SD 1.0) and 7.3 (SD 1.2) years, respectively. These averages, however, obscured substantial heterogeneity. Individuals in the lowest risk quintile did not benefit from metformin, while mid- to high-risk individuals benefitted considerably. For individuals in the lowest risk quintile, metformin did not delay diabetes onset versus usual care, but in the highest quintile delayed it by 2.7 (SD 0.4) years in the short treatment effect scenario.

Compared with usual care metformin was not cost-effective in the lowest diabetes risk quintile, but was cost-saving for higher risk patients (Figure 2).

Figure 2: Population net monetary benefit (NMB) for each treatment by quintile of 3 year diabetes risk with alternative assumptions of treatment effect duration.

Note: The short treatment (tx) effect scenario assumes diabetes onset hazard reduction in the first four years only, and extended treatment (tx) effect scenario assumes diabetes onset hazard reduction extends over the entire lifetime.

Net monetary benefit (NMB) calculated compared to usual care, based on $50,000 willingness to pay threshold

In contrast, the lifestyle intervention delayed diabetes onset by 1.7 (SD 0.3) years in the lowest risk quintile, and 4.1 (SD 0.8) years in the highest quintile (Appendix 2 figure 3). Lifestyle intervention was cost-saving across all risk quintiles.

NMB provides additional insights for comparing the value of cost-saving strategies. Risk-stratified NMB largely reflected the heterogeneity of diabetes risk and intervention treatment effect seen across risk quintiles. In both the short-term and long-term treatment effect scenarios, lifestyle intervention had positive increasing value across 3-year diabetes risk quintiles. In contrast, metformin had a negative value in the lowest risk quintile, with progressively increasing positive values in higher risk quintiles (Figure 2).

Targeting treatment strategies

For the DPP lifestyle intervention, there was no difference between strategies of treating everyone and treating only those with a positive expected NMB, because it had a positive expected NMB for all individuals in the cohort. When treated with metformin, 75–80% of the population had positive expected NMB, depending on assumptions of treatment effect duration. Targeting treatment to only these individuals yielded a NMB gain of $523–880 per person, or 7% of NMB of treatment without targeting.

In the case of total DPP program capacity of 17 million people, or 20% of the eligible population, targeting treatment to individuals in the highest diabetes risk quintile, which had the highest predicted NMB, increased average NMB per person substantially. For lifestyle the difference in NMB ranged from $14,000 to $18,000 per person on average, depending on assumptions of treatment effect duration. For metformin the difference in NMB ranged from $16,000 to $20,000 per person (Figure 3).

Figure 3:

Average net monetary benefit (NMB) per person, a comparison of targeting treatment to those with highest expected benefit versus randomly sampling from the entire eligible population for a subset of 17 million (20%) of the eligible population of adults with impaired glucose regulation.

Sensitivity analysis

In sensitivity analyses the patterns of NMB across diabetes risk quintiles remained stable, with metformin having negative values for the lowest diabetes risk quintile, and becoming positive and increasing with greater risk of diabetes (Appendix 3). For the lifestyle intervention, all risk quintiles had positive NMB across the tested variable ranges. For most tested scenarios, the value of targeting remained stable in the range of 5–10% NMB gain. When assuming long-term treatment effect, active interventions provided greater NMB, a higher proportion of patients with positive NMB, and greater absolute NMB increase from targeting, compared to scenarios assuming short-term treatment effect.

Treatment burden from daily metformin administration had a strong effect in decreasing NMB of metformin, as well as the proportion of patients with positive NMB, increasing the value of targeting treatment (107–145%). Increasing the discount rate tended to reduce all net benefits, with greatest impact on lower risk individuals, when assuming a long-term treatment effect, suggesting that more of the QALY benefits occurred in the more distant future.

Discussion

Our model found that targeting active diabetes prevention to those patients at highest risk could improve health outcomes and reduce costs compared to providing the same intervention to a similar number of patients with pre-diabetes without targeted selection. While on average both metformin and the DPP lifestyle intervention appeared to be cost-saving (compared to usual care), the value of these interventions varied considerably over the DPP-eligible population. These analyses emphasize the potential health and economic benefits of more routinely examining the distribution of effects across risk groups in clinical trials^32,33 for which methodological guidance has recently been proposed.³³

In a scenario in which a healthcare system can treat the entire eligible population of adults with impaired glucose regulation, our results suggest it is worthwhile to offer the lifestyle intervention to all. However, if programs can afford to treat only a fraction of eligible individuals, societal value could be increased by prioritizing lifestyle intervention based on diabetes risk, because greatest value is achieved in individuals in higher diabetes risk quintiles.

In contrast to the lifestyle intervention, preventive use of metformin may not be worth the cost for individuals in the lowest diabetes risk quintile, because these individuals did not appear to benefit from the therapy, particularly when treatment burden and costs were factored in. For patients unable or unwilling to participate in the lifestyle intervention, our analysis suggests that prioritization of metformin prevention to higher risk individuals – i.e., those in whom the net benefit is positive – would confer higher value to society.

The National DPP, an ambitious program to provide a complex intervention to a large population, has faced many challenges while scaling up from the research program to real world settings. The types of lifestyle programs shown to be effective are resource-intensive, involving multiple one-on-one sessions with case managers to educate patients about diet, exercise, and behavior modification, as well as ongoing monthly individual and group sessions to support a healthy lifestyle. Even with current nationwide efforts to expand participation and retention in the National DPP program, many eligible individuals are not participating, and substantial additional investment is needed to maintain participant engagement and retention. Prioritizing efforts and resources towards engaging highest risk participants should confer higher program value.

Conclusions of this study should reasonably generalize to the US population eligible for DPP programs. The eligibility criteria for the National DPP program are, however, less restrictive, than the enrollment criteria for the DPP trial. The criteria for fasting plasma glucose 100–125 mg/dL and 2 hour plasma glucose 140–199 mg/dL remain the same, but the National DPP program criteria now include individuals with hemoglobin A1c levels of 5.7%–6.4%, as well as all individuals with prior gestational diabetes. The program also allows participation starting at age 18, rather than 25 in the trial, and for persons who are slightly more overweight (body mass index 25, rather than 24, or higher). Based on these less restrictive criteria, we expect the National DPP eligible population contains more individuals at low risk of diabetes than the trial population that underlies our analysis. If so, prioritization should confer even greater value than our current model estimates.

One of the strengths of the modeling approach is the ability to project outcome scenarios beyond the observed data in clinical trials and observational studies. Rather than using the often cited average clinical trial results from the literature, we had access to detailed patient-level data and were able to not only model risk using patient-level data, but also project costs and outcomes for each individual in the clinical trial based on their specific detailed baseline characteristics in the microsimulation. Our models for the risk of diabetes onset over time accounted for the diminishing impact of baseline covariates and examined several alternative scenarios for the effects of DPP, both a limited short-term treatment effect, and a persistent long-term effect, These scenarios were consistent with the uncertainty in the durability of DPP efficacy found in a recent meta-analysis, which also showed that the effects of medication therapy were less durable than lifestyle interventions.²⁸

Prior analyses of DPP data have been generally consistent with our findings. One analysis estimated the median delay in diabetes onset to be 11 years with lifestyle intervention and 3.4 years with metformin.² Although these estimates appear more positive than our extended treatment effect scenario, they assumed a greater treatment effect (lower hazard ratios) for 0–48 months, with the highest treatment effect occurring initially during months 0–24. Unlike prior analyses, our study incorporates the patient-level heterogeneity of treatment effect, using diabetes risk-based hazard ratios, rather than population-wide average hazard ratios. Because of the significant interaction between diabetes risk and intervention effectiveness, incorporating heterogeneity into the analysis substantially affects the results.

A limitation of our analysis is that we estimated the effects of lifestyle intervention on health outcomes and costs that are directly related to prevention or delay of diabetes onset and complications. Since better diet and exercise habits may have other far reaching effects on overall health and complication risks, our analysis may underestimate the value of successful lifestyle interventions. Although we use the best and most granular data available for our analysis, there may be additional important factors involved in patient outcomes, especially over a long-term time horizon, that out simulation models do not account for. Future research should explore the effects of variation in the length of lifestyle intervention programs, effects of socioeconomic factors, race, and other factors that may potentially affect the effectiveness of DPP.

Consistent with other economic analyses of heterogeneity of treatment effect,^34,35 we found that NMB or cost-effectiveness varies by diabetes risk strata, based on the heterogeneity in diabetes risk and the patterns of treatment effect for DPP interventions across the population of adults with impaired glucose regulation. Individual level information may be helpful in prioritizing or targeting treatment to those who benefit most and avoid the costs of treatment in individuals who derive little to no benefit. Our analyses showed that while lifestyle intervention provides health benefits at reasonable societal cost to all adults in the cohort, the greatest expected gains occur in higher risk individuals, and—where capacity is limited—these individuals should be targeted. Additionally, according to our analyses metformin administration should be targeted to higher risk individuals, since patients in the lowest risk quartile do not appear to benefit.