Second-Line Agents for Glycemic Control for Type 2 Diabetes: Are Newer Agents Better?

We considered four treatment-intensification regimens: metformin, sulfonylurea, and insulin (T1); metformin, DPP-4 inhibitor, and insulin (T2); metformin, GLP-1 agonist, and insulin (T3); and metformin and insulin (T4). In each regimen, patients started metformin monotherapy when HbA_1c reached the prespecified glycemic control goal. In T1–T3, treatment was sequentially intensified by addition of a second-line agent other than insulin, and if or when HbA_1c again exceeded the glycemic control goal, insulin was initiated (in place of the second-line agent) as the third-line agent in combination with metformin. In T4, treatment was intensified by directly adding insulin once HbA_1c exceeded the glycemic control goal. For all regimens, there were no further treatment changes once insulin was initiated, as it was assumed to maintain glycemic control.

The Markov model is based on the 10 discrete HbA_1c states presented in Supplementary Tables 1 and 2. Each state is defined by the conditional mean HbA_1c in a given interval for a patient newly diagnosed with type 2 diabetes. The mean HbA_1c value for each state increases linearly with respect to age according to a linear trend factor. This common assumption, based on other published glycemic control models (17,18), reflects the expected rise in HbA_1c with age and anticipated deterioration of glycemic control. At the beginning of each 3-month period, treatment is initiated/intensified if HbA_1c exceeds the glycemic control goal. Treatment results in a proportional decrease in HbA_1c according to a medication effect estimated from observational data (Table 1). If no diabetes complications or death occurs, patients undergo continued HbA_1c state transition based on the 3-month transition probability matrices provided in Supplementary Tables 1 and 2. Each treatment regimen was evaluated using the Markov model by backward induction (29). All analyses were conducted using MATLAB R2012b (MathWorks, Inc., Natick, MA).

We considered four outcome measures related to primary prevention: expected LYs, expected QALYs, mean time to insulin dependence, and expected medication cost per QALY for maintaining glycemic control from diagnosis to occurrence of first diabetes-related complication or death. For each period in which no diabetes complications or death occurred, LYs were increased by 3 months, QALYs were adjusted based on the disutility of medications, and the medication cost was calculated based on the sum of the costs of using medications for 3 months discounted at a 3% annual discount rate (30).

A retrospective administrative claims data set that included medical claims, pharmacy claims, laboratory data, and eligibility information from a large, national U.S. health plan was used to estimate 3-month HbA_1c state transition probabilities (Supplementary Data), to estimate the medication effect on reducing HbA_1c (Supplementary Data), and to calibrate and validate our model (Supplementary Data). The individuals covered by this health plan are geographically diverse across the U.S. with greatest representation in the south and midwest U.S. census regions. The plan provides fully insured coverage for professional (e.g., physician), facility (e.g., hospital), and outpatient prescription medication services. Medical (professional, facility) claims include ICD-9, Clinical Modification (ICD-9-CM) diagnosis codes, ICD-9 procedure codes, Current Procedural Terminology, version 4 procedure codes, Healthcare Common Procedure Coding System procedure codes, site of service codes, provider specialty codes, and health plan and patient costs. Outpatient pharmacy claims provide National Drug Codes for dispensed medications, quantity dispensed, drug strength, days’ supply, provider specialty code, and health plan and patient costs. Laboratory results linked to the administrative claims data are available for a subset of these patients. All study data were accessed using techniques that are in compliance with the Health Insurance Portability and Accountability Act of 1996, and no identifiable protected health information was extracted during the course of the study. Because this study involved analysis of preexisting, de-identified data, it was exempt from institutional review board approval.

The population meeting criteria for our study (37,501 individuals) were age of at least 40 years, diagnosis with type 2 diabetes between 1995 and 2010, prescription for their first noninsulin glucose-lowering medication at least 6 months after enrollment, and at least 5 years of continuous enrollment with at least two HbA_1c records and complete pharmacy claim data. Type 2 diabetes was defined using the Healthcare Effectiveness Data and Information Set criteria (31). Healthcare Effectiveness Data and Information Set requirements for pharmacy data include at least one antihyperglycemia medication prescription and, for claim encounter data, the presence of at least one diabetes-specific ICD-9 diagnosis codes 250.XX (exclude 250.X1 and 250.X3), 357.2X, 362.0X, or 366.41 with two annual face-to-face outpatient encounters with different dates of service or one face-to-face in an acute inpatient or emergency department encounter.

Model parameters, including base-case values and ranges for sensitivity analysis, are shown in Table 1. We assumed a diagnosis age of 55.2 years for women and 53.6 years for men based on the median age at time of diagnosis of diabetes in the U.S. as of 2011 (19). The initial HbA_1c state distributions for men and women are shown in Supplementary Tables 1 and 2. Treatment regimens were assumed to be fixed for patients living beyond 100 years, and future life expectancy at age 100 years was assumed to be 2.24 years for women and 2.05 years for men based on a 2008 U.S. life table (32).

The probabilities of diabetes complications were determined by a patient’s age, sex, ethnicity (Afro-Caribbean or not), smoking status, BMI, HbA_1c, systolic blood pressure, total cholesterol, and HDL cholesterol; history of peripheral vascular disease, atrial fibrillation, ischemic heart disease, and congestive heart failure; and blindness at diagnosis using the UKPDS outcomes model (33). Probability of death from other cause was estimated based on the Centers for Disease Control and Prevention 2007 mortality tables (34).

The cost of medications other than insulin was based on the federal median price for generic agents and the average wholesale price for brand name agents provided by the Agency for Healthcare Research and Quality Evidence Practice Centers (16). The cost of insulin therapy, including the cost related to self-monitoring of blood glucose, insulin, and insulin-related supplies, was taken from Yeaw et al. (23). All costs were inflation adjusted to 2013 dollars using the consumer price index method (35). For medications other than insulin, the base-case cost was the mean price of all brand name and generic (if available) medicines, and the cost in the range represents the least and the most expensive medicines. The base-case cost for insulin was the mean cost of all insulin regimens including basal insulin regimens, premixed insulin regimens, and basal-bolus insulin regimens. The cost in the range represents the average cost for basal insulin therapy (the least expensive insulin therapy) and the average cost for basal-bolus insulin therapy (the most expensive insulin therapy), respectively.

Medication effect (other than for insulin) was estimated based on HbA_1c changes seen with use of these agents by patients included in the data set and is presented as the relative reduction in HbA_1c observed during each 3-month treatment interval.

To calibrate and validate the model, we used all available HbA_1c pairs at least for 3.5 months to ensure at least one 3-month transition, as long as the patient was not on insulin during that time period. This provided a total of 97,667 pairs of HbA_1c test results. The linear trend factor was varied from 0 to 0.25 to estimate the trend factor that minimized the mean of the sum of the squared errors (SSE) between the observed HbA_1c state distribution (determined by the second HbA_1c value in each pair) and the model-generated HbA_1c state distributions. The optimal trend factor was 0.1075 for men (mean SSE of 0.0022) and 0.105 for women (mean SSE of 0.0015). Additional details of the model calibration and validation can be found in Supplementary Data.

The Markov model–based results showed that the expected LYs and QALYs from diagnosis to first event produced by the four treatment regimens were similar (Table 2). The maximum difference among regimens in the expected LYs to first event, specifically, the difference between T4 and T1, was 0.03 years (12.73 days) for women and 0.03 years (11.06 days) for men. Similarly, the maximum difference among regimens in the expected QALYs to the first event, specifically, the difference between T4 and T1, was 0.04 QALYs (16.12 quality-adjusted days) for women and 0.04 QALYs (14.20 quality-adjusted days) for men. The observed differences in expected LYs and QALYs among regimens were primarily the result of different expected durations of sustained glycemic control with the three second-line agents (in combination with metformin). The mean time elapsed between failure of metformin monotherapy and the need for insulin initiation was 1.05 years (381.99 days) for women and 1.0 year (364.65 days) for men using T1, 0.62 years (224.50 days) for women and 0.53 years (194.84 days) for men using T2, and 0.68 years (247.96 days) for women and 0.62 years (225.46 days) for men using T3.

Significant differences were observed in the expected medication cost per QALY incurred by the four treatment regimens. Compared with using sulfonylurea as a second-line agent, which was the least expensive treatment regimen, use of DPP-4 inhibitor (T2) was associated with a mean per-person additional medication cost of 141 USD per QALY for women and 160 USD per QALY for men. Use of GLP-1 agonist (T3) incurred a mean additional medication cost of 191 USD per QALY for women and 216 USD per QALY for men compared with T1, and use of insulin as a second-line agent (T4) incurred a mean additional medication cost of 150 USD per QALY for women and 170 USD per QALY for men compared with T1.

For any fixed glycemic control goal ranging between 6.5% (48 mmol/mol) and 8.0% (64 mmol/mol), use of sulfonylurea as the second-line agent incurred the lowest expected medication cost per QALY, and GLP-1 agonist use incurred the highest expected medical cost per QALY, among both men and women (Fig. 1). Targeting a treatment goal of 6.5% (48 mmol/mol) vs. 7% (53 mmol/mol) incurred significantly higher expected medication cost per QALY and a small reduction in the expected QALYs for all treatment regimens (Fig. 1). All treatment regimens resulted in increased expected QALYs and increased medication cost per QALY when targeting a treatment goal of 7% (53 mmol/mol) compared with 8% (64 mmol/mol) (Fig. 1).

The expected medication cost per QALY of each of the four treatment regimens varied significantly (Fig. 2) as a result of differential costs incurred by generic (metformin, sulfonylurea) compared with brand name (DPP-4, GLP-1) medications and basal insulin compared with basal plus bolus insulin regimens. T3 exhibited the largest variation in the expected medication cost per QALY (503 USD per QALY difference for women and 453 USD per QALY difference for men), while T2 was associated with the smallest variation in the expected medication cost per QALY (291 USD for women and 261 USD for men).

When the effects of sulfonylurea, DPP-4 inhibitor, and GLP-1 agonist on HbA_1c were simultaneously set to be the lower bound or upper bound of the randomized control trial (RCT) results on the efficacy of medications (Table 1 ), the four treatment regimens still resulted in similar expected LYs and QALYs from diagnosis to first event. The treatment regimen with sulfonylurea as the second-line agent resulted in the lowest cost per QALY (2,537 USD per QALY for women and 2,612 USD per QALY for men at lower bound and 2,388 USD per QALY for women and 2,454 USD per QALY for men at upper bound), while the treatment regimen with GLP-1 agonist as the second-line agent still produced the highest cost per QALY (2,809 USD per QALY for women and 2,911 USD per QALY for men at lower bound and 2,867 USD per QALY for women and 2,971 USD per QALY for men at upper bound).

The conclusions drawn from this study are based on a model and therefore may not be a perfect representation of what would be observed in practice. Direct comparison of four different diabetes treatment regimens by the Markov model developed and validated in this study demonstrated that all four treatment regimens resulted in similar expected benefits in LYs and QALYs irrespective of glycemic control goal. However, for all glycemic control goals ranging between the currently recommended targets of HbA_1c 6.5% (48 mmol/mol) and 8% (64 mmol/mol), the use of sulfonylurea as the second-line agent incurred the lowest expected medication cost per QALY. These findings hold for both observed effects of medications from real-world data and randomized control trial results. The differences in cost per patient among the four treatment regimens were substantial and thus of potential importance to patients as well as health care providers and payers. In addition, the treatment regimen with a sulfonylurea as the second-line agent resulted in the longest time of insulin independence compared with all other regimens—an important factor to be considered by patients who wish to delay insulin initiation as long as possible. Conversely, the more expensive treatment options that use a DPP-4 inhibitor or a GLP-1 agonist as the second-line agent were associated with slightly less expected benefit in terms of both LYs and QALYs, and a shorter time of insulin independence, compared with the use of sulfonylurea. Use of insulin as the second-line agent resulted in the shortest time to insulin dependence, and was also significantly more expensive than using sulfonylurea with no added benefit in terms of LYs or QALYs.

To date, there has been no comprehensive side-by-side evaluation of the clinical benefits, effects on quality of life, and costs incurred by different diabetes treatment regimens for glycemic control. Our model fills this gap by integrating real-world knowledge of treatment costs, benefits, and harm, thereby allowing clinicians, payers, and patients to directly compare treatment regimens to select the one that is best suited for each individual patient given his/her specific glycemic control goal, cost sensitivity, and preference. Given that >25 million patients have been diagnosed with type 2 diabetes in the U.S., the potential policy implications of these differences uncovered by our model are also significant.

The Glycemia Reduction Approaches in Diabetes: A Comparative Effectiveness Study (GRADE), which is in the recruitment phase now, seeks to compare the same treatment regimens using a prospective clinical trial design; however, our model is significantly different from that of GRADE in that our results compare QALYs and costs for newly diagnosed patients and because our treatment efficacy is based on data that captures long-term adherence effects that are typically much smaller in clinical trials.

Several models have been developed to predict the natural history of diabetes-related complications progression and to gauge their sequelae on patient quality of life (17,18,36–38); however, none of these models were based on real-world data describing the rate of and variations in HbA_1c progression caused by both biological changes and patient behavior with and without different treatment modalities. Moreover, none of the previous published models explicitly compared and contrasted different treatment regimens with regard to their practical efficiency, cost, and clinical benefit based on real-world inputs rather than clinical trial data or select observational study population groups. To our knowledge, this study is the first to develop and validate a glycemic control model that takes into consideration the known adverse effects of treatment, such as hypoglycemia, current medication cost, and various suggested glycemic control goals.

Our model can serve as an adjunctive decision aid to facilitate treatment selection for people newly diagnosed with type 2 diabetes in a way that trades off health and economic implications for patients. It can also be used by health care providers and payers to determine whether a particular treatment option is consistent with the goal of high-value care, e.g., providing a clinically justified benefit given the incurred cost. While no clinical study has yet definitively established the clinical benefit of using incretins in place of sulfonylureas as second-line agents and there is increasing concern regarding sulfonylurea use owing to its association with severe hypoglycemia (10), our model, which considers the side effect of severe hypoglycemia, suggests that for a glycemic control goal of 6.5% (48 mmol/mol) or 7% (53 mmol/mol), sulfonylureas provide higher value than incretin. Indeed, use of incretins as second-line agents (treatment regimens T2 and T3) resulted in significantly higher cost but slightly less clinical benefit as measured by LYs and QALYs to first incident diabetes-related complication or death. However, ultimate value will depend on patient preference.

Our study has several limitations. First, the results presented in this article are based on a Markov model rather than a clinical trial, and no model can provide a perfect representation of reality. Specifically, our model assumes that HbA_1c varies among discrete states and at discrete 3-month time intervals rather than continuously; furthermore, transitions among states are assumed to depend only on the most recent HbA_1c state. For addressing these limitations, the assumptions were carefully validated based on real patient data. Treatment regimens were designed as sequential one-by-one additions of different classes of antihyperglycemic medications, while in clinical practice patients may start two or more drugs at the same time. We also assumed that insulin would replace the previously used second-line drug, as recommended by most clinical practice guidelines, but it is possible for patients to continue using two or more noninsulin agents in conjunction with insulin. We assumed that insulin will ultimately result in achievement of the glycemic goal; this is an idealized assumption that is based on the physiology of insulin action, and there is likely to be substantial variation among patients in whether they achieve and maintain their glycemic goal over time. Finally, the model is based on data that represents a privately insured population. Therefore, it is possible that these results may not be generalizable to the Medicare and Medicaid populations.

Several features that were not incorporated into the current model are due to insufficient evidence in literature such as the potential variability in how medications influence HbA_1c trajectory, the potential variability in the duration of observing the effect of medications, and the potential indirect pleiotropic effects of these medications not mediated by their glucose-lowering properties. Medication disutility values were based on limited empirical data because definitive evidence is not yet available. Our analyses were focused on primary prevention of the most common micro- and macrovascular complications of diabetes, and patients included were treatment naïve and newly diagnosed with diabetes. To the extent possible, we have used previously published data on the utility decrements for complications and treatments; however, utility estimates are limited in that they represent an average measure and do not reflect individual patients' well-being. To address this, we performed sensitivity analysis on the utility estimates. Finally, not all known adverse medication effects were included in the model. We did not consider severe nausea and other gastrointestinal side effects of metformin or DPP-4 inhibitors (16), since these symptoms and availability of alternatives would likely cause the medication to be discontinued. We did not consider pancreatitis risk from the new agents due to the uncertainty of this evidence (39,40). Ultimately, however, our proposed model is sufficiently versatile to allow for easy integration of newly acquired clinical knowledge and its continued refinement.

Two key factors that were not explicitly incorporated into the model are medication adherence and lifestyle modifications, both of which are known to improve glycemic control, particularly in early stages of diabetes. However, this is alleviated by our use of real-world observational data for patients who adhere to their treatments and lifestyle recommendations with the frequency expected from any general population among which such therapies are to be deployed. This affords our model an aspect of generalizability and validity that makes it attractive and relevant to patients, health care providers, and payers.

Funding. This work was funded in part by Agency for Healthcare Research and Quality under grant 1R21-HS-017628 (to N.D.S.). This material is also based in part on work supported by the National Science Foundation under grant no. CMMI-0969885 (to B.T.D.).

The funders had no role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. Y.Z. designed the study, developed the model, performed the data analysis, interpreted the results, wrote the manuscript, and reviewed and edited the manuscript. R.G.M. contributed to the results interpretation and discussion, wrote the manuscript, and reviewed and edited the manuscript. J.E.M. provided critical recommendations regarding study design and methodology, contributed to the results interpretation and discussion, and reviewed and edited the manuscript. S.A.S. provided critical recommendations regarding study design, contributed to the results interpretation and discussion, and reviewed and edited the manuscript. N.D.S. provided funding for the acquisition of data, provided critical recommendations regarding study design, contributed to the results interpretation and discussion, and reviewed and edited the manuscript. B.T.D. provided funding for the study, provided critical recommendations regarding study design and methodology, contributed to the results interpretation and discussion, and reviewed and edited the manuscript. Y.Z. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.