Sulfonylurea use and the risk of hospital readmission in patients with type 2 diabetes

A retrospective cohort analysis was conducted using longitudinal data (two-year panels from 1999 to 2010) from the Medical Expenditure Panel Survey (MEPS), a nationally representative and publicly available survey of healthcare utilization and costs in the United States. In the MEPS panels, sampled households participated in five rounds of interviews over the course of two years. Survey responses were then validated by the household’s healthcare providers, such as pharmacists, physicians, and hospitals. All healthcare expenditures for the households and sources of payment, over the two-year period, were also collected by the MEPS. Partly because of supplemental data obtained for households with members suffering from diabetes, the MEPS has been used extensively to study diabetes [13, 14], including drug utilization by those with diabetes [15].

The MEPS database can be used to generate national estimates of healthcare utilization and cost if appropriate weights obtained from stratification and clustering variables are applied. The MEPS includes both general patient weights and patient-with-diabetes-specific weights. When available, the latter were used in this study; when not, the general weights were applied. The consistency over time in the MEPS design allows for pooling data over multiple panels, often necessary in order to obtain reliable national estimates based on high enough (>100) raw cell counts [16].

Patients with a diabetes-related hospital admission during their two-year panel were identified from the inpatient and emergency-department MEPS files by having one of the ICD-9 diagnosis codes in Appendix 1 Table 4 as one of their listed (not necessarily primary) diagnoses [6]. Among this group of patients, only those receiving an oral AHA post-hospitalization, but before any diabetes-related readmission if one occurred, were considered for further study. Patients receiving insulin, any time prior to their second hospital admission, identified by the level-2 therapeutic class code 215 from the prescribed-medicines file in the MEPS, were excluded. Patients who were diagnosed as pregnant (ICD-9 code V22.x) were also excluded.

Patients with a hospital readmission were identified as having a subsequent hospital admission for a diabetes-related condition within one year of the first admission. For patients with no readmission, we examined claims for three rounds following the round of the first admission or until the end of the two-year panel, whichever occurred first.

Depending on their medication use (see Appendix 1 Table 5 for level-2 therapeutic classification codes for drugs) following their first hospital admission, patients were categorized as follows: those on SU monotherapy (SU cohort); those on combination therapy that included an SU (SU+ cohort); those on monotherapy with a non-SU AHA (noSU cohort); and those on combination therapy without an SU (noSU+ cohort).

Although descriptive readmission and cost statistics were obtained for all four patient cohorts, all other analyses were performed using monotherapy patients (SU and noSU groups) in order to preserve a clean comparison between the SU and other drug classes. The outcome variable, time to readmission, was computed as the number of days between the first and second hospitalization.

Covariates were all baseline patient characteristics potentially associated with initial drug assignment, as well as with time to readmission. They included demographic characteristics (age, gender, race, geographic location, and marital status); insurance coverage; period of first hospital admission (to detect any potential trend over time); comorbidities (cardiovascular disease, renal disease, and eye disease); medical care received (number of HbA1c tests and difficulty in getting care); and disease severity (perceived health status, perceived mental health status, and limitations in physical functioning).

A logistic regression model, with assignment to the SU cohort as the dependent variable and the covariates as independent variables, was estimated to obtain propensity scores for SU use. To determine whether the propensity score achieved balance across the covariates, each covariate was regressed, by way of an appropriate dependent-variable specification (logit, multinomial logit, or ordered logit), on the treatment (SU versus noSU) and the patient’s propensity score [17, 18]. Balance was determined based on a p-value for the estimated treatment coefficient of > 0.05. Because balance for all covariates was indeed achieved according to the 0.05 criterion following this procedure, no additional statistical measures were taken to guard against selection bias.

Kaplan-Meier survival curves were generated from the time-to-readmission data for each of the study arms. Then, using a Cox proportional hazard model, time to readmission was regressed on study-arm assignment, a subset of the baseline patient characteristics (for parsimony, not all variables were included), and the propensity scores, following a covariate-adjustment procedure from the propensity-score literature [19]. Patients without a readmission experienced a censoring endpoint, which was either one year from the date of first hospital admission or the end of the MEPS panel, whichever occurred first.

Additional statistical tests were run to ensure correct model specification. All analyses were conducted using SAS version 9.2 and R version 2.14.2. With the exceptions of the Kaplan-Meier curves and the t-test for equality of mean age between cohorts, which could not be accomplished in SAS from weighted survey data, all analyses, including the propensity-score analysis, used population-weighted MEPS data, necessary to obtain unbiased treatment effect estimates that are generalizable to the original survey target population [20]. Because of the public nature of the MEPS database, the research did not require review and approval by the Institutional Review Board at the University of Cincinnati.

Hospital readmissions are a major concern of healthcare providers and policymakers because they indicate deterioration, instead of improvement, of patients’ health. Whereas some readmissions have been associated with patient frailty and inevitable disease progression, many have been found to occur due to substandard care during the initial hospitalization, including poor resolution of the main problem, unstable therapy at discharge, and inadequate post-discharge care [21]. A section of the Patient Protection and Affordable Care Act of 2010, which allowed the Centers for Medicare & Medicaid Services to implement the Hospital Readmissions Reduction Program, has placed a spotlight on the need to reduce readmissions as a way to improve the quality of care and reduce costs across disease conditions, including diabetes [22]. To the extent that more effective medication management can occur pre- and/or post-discharge following an initial hospitalization, it may be possible to reduce the readmission rate for diabetes sufferers.

The number of people with diabetes and comorbid conditions is expected to increase precipitously in the near future. In fact, the $245 billion cost burden of diabetes in 2012 represented a 41 % increase over the $174 billion burden in 2007, in large part driven by a 27 % growth in diabetes prevalence over that five-year period [3]. According to the Centers for Disease Control and Prevention, an estimated 86 million adults are currently considered to have prediabetes, meaning that their blood glucose levels are higher than normal but below the diabetes range [23]. Moreover, one study has projected the incidence of diabetes to be about 15 cases per 1,000 in 2050 compared to 8 cases per 1,000 in 2008 (essentially a doubling, in other words) [24]. Hence, to the extent that hospitalizations can be prevented by more effective pharmacotherapy in the future, there will be substantial potential cost savings available.

The conservative readmission rates (approximately 15 %–25 %) in our study, relative to those mentioned in the introduction, probably result from the omission of insulin users, who are likely to be patients with more comorbidities. Ng and colleagues found that, for a Canadian cohort of patients with type 2 diabetes, those taking insulin had a significantly higher likelihood of a hospitalization (OR = 1.7, 95 % CI: 1.4–2.0), controlling for numerous demographic, socioeconomic, and health-status characteristics [25].

Our study’s results are consistent with research that has compared the antidiabetic drugs with respect to outcomes that could lead to hospitalizations. Bodmer and colleagues found that use of SUs was associated with a 2.8 times higher risk of hypoglycemia than metformin. [26] Evans and colleagues found that patients treated with SUs only were at higher risk of adverse cardiovascular outcomes than those treated with metformin alone [27]. Eurich and colleagues found that, compared with SU monotherapy, metformin, alone or in combination, was associated with a statistically significant lower morbidity and mortality for patients suffering from both type 2 diabetes and heart failure [28]. Finally, Horsdal and colleagues found lower 30-day mortality rates among users of metformin (HR 0.32, 95 % CI: 0.15–0.68) as well as patients without pharmacotherapy (HR 0.58, 95 % CI: 0.36–0.93) compared with users of SUs [29].

In this study, the patients in the SU cohort were on average 8 years older than those taking a non-SU AHA, a fact that helps to explain the difference in readmission cost between the two cohorts. Older patients were both more likely to have additional comorbidities, requiring attention during the hospitalization, as well as more advanced diabetes. Moreover, as seen in Table 2, the SU cohort, before propensity-score adjustment, had a significantly higher percentage of patients with fair or poor perceived health status than those in the other cohort of patients, potentially lengthening and/or complicating their hospital stays. This result is consistent with that of Raebel and colleagues who found that older patients as well as those with elevated serum creatinine were more likely to initiate SU therapy than younger, less severe patients [30]. This result, however, is not consistent with that of Desai and colleagues who found that older patients (70 years old and older) were significantly more likely to receive metformin (not an SU) as initial therapy than those younger than 70 [31].

Unmarried patients had a statistically significant higher risk of hospital readmission than patients who were married. This result is consistent with a number of studies showing that marriage has a protective effect on mortality and hospitalization [32, 33]. The statistically significant effect of eye disease makes sense given that the strongest predictor for development and progression of retinopathy is duration of diabetes [34]. Patients suffering diabetes for longer are more likely to develop eye problems as well as other complications leading to hospitalizations.

There was a significant fall in the number of multiple hospitalizations for patients with diabetes over time. Readmissions were less likely to occur later in the study period than earlier. This result is consistent with what Cunningham and Carrier found in their study of the trend in medical-care costs for nonelderly adults with diabetes [35].

The treatment of diabetes changed substantially over the decade of the 2000s. Specifically, there was a de-emphasis on the use of the SU drug class. From the MEPS data (Table 2), the share of SU patients was higher in the earliest time period than the latest, while the share of noSU patients was higher in the latest period. Consistent with our results, Desai and colleagues found that the proportion of patients initially treated with an SU decreased from 2006 to 2008. Over the same period, there was a significant increase in the use of metformin and DPP-4 inhibitors and a significant decline in the use of TZDs [31]. A longer trend analysis was undertaken by Alexander and colleagues. In their study, SU utilization (monotherapy or combined) decreased from 67 % of treatment visits in 1994 to 34 % in 2007 [36]. A new class of antidiabetic agents for the treatment of type 2 diabetes, the sodium-glucose co-transporter 2 (SGLT2) inhibitors, has become available since the end of our study period. Canagliflozin (Invocana®), dapagliflozin (Forxiga®), and empagliflozin (Jardiance®) were approved by the Food and Drug Administration in April 2013, January 2014, and August 2014, respectively [37‐39]. Additional treatment options may encourage a decline in the use of older medications.

In the 2015 position statement (2015 diabetes guidelines) of the American Diabetes Association, the suggested approach to the management of hyperglycemia in most individuals with type 2 diabetes includes intervention at the time of diagnosis with metformin (the only marketed biguanide in the U.S.) in combination with lifestyle changes [9]. If noninsulin monotherapy at maximal tolerated dose is not effective within 3 to 6 months, a second oral agent or insulin should be added as a means of achieving and maintaining recommended levels of glycemic control [9]. For some older adults, however, metformin may be contraindicated because of renal insufficiency or heart failure. Moreover, again a concern primarily for older adults, TZDs may cause fluid retention, which may lead to heart failure [9]. For these patients, the use of alternative AHAs, including SUs, may be warranted. Because our study period ended in 2010, healthcare providers did not have access to the 2015 guidelines. Earlier position statements (see Nathan and colleagues [40, 41]), however, provided pharmacotherapy guidance for physicians. They also suggested reduced reliance on the SUs.

The MEPS has been widely used in the study of diabetes and has served as the database of choice to examine many aspects of this chronic condition. However, the lack of clinical data on disease severity (there are no Hba1c values, for example) and progression limits our ability to estimate causal relationships between drug use and risk of hospital readmission. To the extent that unobserved variables were associated with the propensity to receive SU treatment, our results may be biased. Additionally, we cannot rule out the possibility that differences in readmission rates may be related to beneficial effects of other AHAs, especially those of metformin, rather than detrimental effects of SUs. Because of the limited longitudinal nature of the MEPS database (as opposed to claims databases), we were unable to identify (1) patients who were initiating antidiabetic pharmacotherapy; (2) patients’ incident hospitalizations, implying that their first hospitalizations may themselves have been readmissions; or (3) patients’ prior healthcare utilization, which could be predictive of future utilization, including hospitalizations. Furthermore, we defined multiple-hospitalization patients as those who had two or more hospital admissions during a year’s time. Whereas previous literature has looked specifically at time between discharge and readmission [5, 6], the MEPS data were not complete enough in many cases to measure time between hospitalizations that precisely.