Summary

The main conclusions for each of the Key Questions are reported in Table 24. The table includes our assessment of the strength of the body of evidence on each Key Question, broken down by the medication comparisons for which data were reported.

Table 24

Evidence of Comparative Effectiveness of Oral Diabetes Medications.

Discussion

This report addresses the comparative effectiveness and safety of the oral diabetes medications used most frequently in the United States. As expected, short-term or proximal outcomes such as HbA1c levels were studied more frequently in RCTs than were distal outcomes such as long-term complications of diabetes. HbA1c is unequivocally linked to the risk of microvascular disease, making it a good proximal outcome measure; in addition, it may also be linked to macrovascular disease. We found that most oral diabetes medications reduced HbA1c to a similar degree, except for nateglinide and acarbose, which appeared less effective in indirect comparisons. Inzucchi and coworkers have conducted a systematic review evaluating HbA1c among the oral diabetes medications and have drawn similar conclusions. ²⁶⁶ Our study has added to this body of work by including more recent articles, especially on the meglitinides, and by conducting meta-analyses. Also, Inzucchi et al. only included articles in which HbA1c was listed as the major endpoint, while we evaluated articles that reported HbA1c even if it was not a major endpoint of the study. ²⁶⁶

Oral diabetes medications varied in their effects on body weight/BMI, with metformin and acarbose being the only medications associated with weight loss, as compared to other oral diabetes medications that increase weight. Our results are consistent with those of other reviews; ^{17, 266, 267} however, no other systematic review to date has included meta-analyses of the effects on weight of as many different types of oral diabetes medications. One might argue that the weight loss seen with metformin and acarbose was only as a result of the removal of medications that increase weight in the run-in and that, in general, metformin and acarbose are associated with weight neutrality, rather than weight loss. This interpretation appears particularly reasonable because the published placebo-controlled trials have shown no weight change for acarbose and metformin. These findings suggest that any weight loss becomes significant in direct comparisons between medications only when the two medications have contrasting effects on weight.

Generally, the reported weight gain was small to moderate, even in the longer-duration RCTs such as the UKPDS (<5 kg). However, even small amounts of weight gain (5–10% of body weight) may be associated with increased morbidity because they worsen insulin resistance and lipid levels, decrease patient adherence, and can cause CHF exacerbations, ^{268, 269 – 271} while weight loss improves control of diabetes. ^{272 – 274} Also, different types of weight gain (central versus peripheral) may have different effects on morbidity, with central adiposity being considered to be more predictive of cardiovascular outcomes. ^{275, 276} Few studies have evaluated whether weight gain is related to increases in visceral adipose tissue, subcutaneous fat, or plasma volume; however, a few recent studies have suggested that different medications affect weight through different mechanisms. Sulfonylureas are thought to increase weight by increasing insulin release, which increases glucose uptake by cells and leads to increased storage of glucose as fat. ²⁷⁷ In two clinical studies, ^{277, 278} pioglitazone has been associated with an increase in total body water and subcutaneous fat and a decrease in visceral fat, most likely reflecting its effect on peroxisome proliferators-activated receptor (PPAR) gamma. A larger number of studies of more oral diabetes medications are needed to allow us to establish conclusively the existence of such differences and their potential impact on long-term outcomes.

Blood pressure control is extremely important in patients with diabetes. ^{279 – 285} The UKPDS showed that for every decrease in blood pressure of 10 mmHg, there is a 15% decrease in diabetes-related deaths, with no threshold effect. ²⁸³ Oral diabetes medications generally had minimal effects on blood pressure. Two systematic reviews have evaluated thiazolidinediones' effects on blood pressure when compared with placebo and reported a slight but significant reduction in blood pressure of 2 to 4 mmHg; ²⁸⁶ however, no one has compared the blood pressure effects of thiazolidinediones with other oral diabetes medications. Thiazolidinediones are thought to affect blood pressure through a variety of potential mechanisms, including intracellular increases in calcium, increased production of angiotensin II type I receptor in vascular smooth muscle cells, and inhibition of the production of various vascular cell types, to name only a few. ²⁸⁷ A suggestion that blood pressure is decreased in the thiazolidinedione group when compared with second generation sulfonylureas and acarbose has been made and requires further exploration. Given the small non-significant between-group differences of 3–5 mmHg, the clinical relevance of these differences is questionable, especially since these RCTs were of short duration.

Effects on lipid levels have been found to vary across medication type, but most effects were small to moderate. For instance, pooled analyses showed between-group differences of about 5 to 10 mg/dL in LDL and 10 to 30 mg/dL in TGs, and summary data for HDL showed differences of about 3 mg/dL. Buse et al. published a systematic review on lipid levels in 2002 that reported similar results. ²⁵ Our review updates theirs; we were also able to add more detail on specific differences, since we presented meta-analyses for LDL and TGs, whereas Buse and colleagues derived their results by totaling the numbers of studies with statistically significant differences.

In our evaluation of lipids, we noted that a single medication can have favorable effects on one lipid outcome and unfavorable effects on another lipid outcome. For instance, most oral diabetes medications decreased TGs, yet only thiazolidinediones increased HDL and LDL. Elevated LDL, elevated TGs, and low HDL are associated with cardiovascular morbidity in epidemiologic studies. ²⁸⁸ In addition, treatment of lipids (especially LDL) has also been associated with improved cardiovascular outcome. ²⁸⁸ While lipid-lowering treatments exist, patients do not reach adequate treatment goals for lipids for a multitude of reasons. ^{289, 290} Therefore, decisions regarding medications that may adversely affect lipids carry a higher relative importance. Few trials in our review evaluated the effects on lipid sub-fractions, an analysis that is thought by some to confer important additional information on cardiovascular risk. This topic may be an area for future research once more data have been acquired with regard to the effects of lipid sub-fractionations on cardiovascular outcomes.

Several caveats deserve mention regarding the proximal outcomes. First, a study's results may have been influenced by the baseline levels obtained. For instance, lower baseline levels for particular outcomes would generally be associated with smaller between-group differences, and vice versa. We were limited in our assessment of heterogeneity based on baseline levels. Because of concerns regarding ecologic fallacy, we did not use individual level characteristics in our metaregression. However, although baseline levels might contribute to between-study variance, they would not influence the direction of the point estimates, and so would not be expected to markedly influence the results.

Second, many studies failed to report the significance of between-group differences and their measures of dispersion, thereby hindering efforts to estimate effect size across trials. We used a conservative estimate of 0.5 for the correlation between baseline and final values when calculating variance. As a result, studies with calculated variances had less weight than studies that reported variance in the publication, a difference that may have influenced our results.

Third, many trials were industry-sponsored, raising the possibility of publication bias. ²⁹¹ While obvious publication bias was generally not observed, these analyses had only limited power to detect publication bias because of the small numbers of studies available for many of the drug comparisons.

Fourth, the indirect comparison results must be viewed with caution, since indirect comparisons tend to overestimate effects. ¹⁷⁵ In addition, we were unable to fully assess the heterogeneity in the placebo groups. Furthermore, we indirectly compared pooled estimates of acarbose versus placebo from another systematic review with our own pooled estimates, creating further heterogeneity. However, these trial results were similar to the placebo-controlled trial results in our review. Finally, few of these RCTs, with the exception of the UKPDS, were long-term studies, making it difficult to assess potential attenuation or exacerbation of effects over time. For instance, in UKPDS, HbA1c initially was reduced in the first few years, then began to rise in both the metformin and sulfonylurea groups. ^{15, 16}

Very few published studies have compared oral diabetes medications in terms of clinically important distal outcomes, such as cardiovascular events, mortality, and microvascular outcomes. We found that metformin, sulfonylureas, and pioglitazone were the only medications associated with long-term reductions in vascular risk. Inzucchi and colleagues reported similar findings in 2003. ²⁶⁶ However, we were able to add pioglitazone to their list, since the PROactive study had been published since their review. The UKPDS has provided the bulk of the relevant data on distal outcomes, along with several cohort studies.

Several caveats deserve mention here: First, long-term events such as renal failure or death and amputations from peripheral vascular disease were generally rare, limiting statistical power to detect differences between medications for these outcomes. Second, study end-points and populations varied greatly. For instance, some studies evaluated primary prevention of cardiovascular disease, while others evaluated secondary prevention. Cardiovascular outcomes ranged from electrocardiogram abnormalities to in-stent re-stenosis rates to nonfatal and fatal myocardial infarction or stroke. Nephropathy outcomes were reported as urinary albumin to creatinine ratios, microalbuminuria, proteinuria, change in glomerular filtration rate, and renal failure. Third, all-cause mortality was difficult to assess across studies, since many smaller shorter-duration trials failed to report mortality, even when no deaths occurred. Fourth, methods to assess and classify outcomes varied across studies: Some used vital statistics, others used claims data, and others used medical record review. Fifth, cohort studies often failed to adjust for potentially important confounders such as the duration of diabetes, HbA1c level, or blood pressure level. Confounding by indication was a large concern, since many patients who have more serious disease or have a longer duration of diabetes are put on sulfonylureas or combinations of sulfonylurea and metformin. These limitations made it difficult to draw firm conclusions about differing effects of the oral diabetes medications on distal outcomes.

Finally, we evaluated the comparative safety of the oral diabetes medications. Our conclusions were consistent with other systematic reviews that have analyzed specific adverse events associated with a single oral diabetes medication. ^{22, 292, 293} However, no previous systematic review has systematically assessed all serious adverse events of all oral diabetes medications in one report.

Minor and major hypoglycemia was more common among sulfonylureas (especially glyburide) and combinations including sulfonylureas than for other oral diabetes medications except repaglinide, which resembled the sulfonylureas. However, repaglinide may be associated with less serious hypoglycemia than the sulfonylureas, as indicated in one randomized crossover study in the elderly, ¹⁹⁹ or with less overall hypoglycemia than glyburide (0 vs 6 events) in patients who skip meals, as seen in one RCT. This last trial was not included in our review because it lasted less than 3 months. ²⁹⁴ Few studies stratified minor or major hypoglycemia according to glycemic control, although these data would be important for subjects with serious hypoglycemia. Despite this omission, little heterogeneity was found in many of these meta-analyses.

Lactic acidosis is the adverse effect most commonly mentioned as a specific concern for patients taking metformin. Because of this concern, metformin is contraindicated in patients with impaired renal function and congestive heart failure. However, neither our review nor the systematic review by Salpeter et al. ⁴⁹ produced consistent evidence of an elevated risk of lactic acidosis in patients taking metformin, when compared with other oral diabetes medications. The concern with regard to lactic acidosis mainly represents a response to about 300 case reports that we did not evaluate in our review. The problem with using case report data is that it is difficult to determine cause and effect, and the effects reported may reflect underlying disease rather than medication effects. In fact, most of the reported cases have occurred in patients with severe acute conditions, such as myocardial infarction or acute renal failure, that could have caused the lactic acidosis. ^{295, 296} We did not have enough information on subjects who were taking metformin and had chronic conditions such as chronic renal insufficiency, chronic liver disease, congestive heart failure, or severe pulmonary disease; therefore, we were unable to determine the safety of taking metformin in patients with co-morbidities that predispose subjects to lactic acidosis.

In our review of published studies, thiazolidinediones were consistently associated with an increase in the number of episodes of self-reported edema; unpublished FDA data corroborate this finding. An advisory letter was issued by GlaxoSmithKline in December 2005, reporting that macular edema had been reported post-marketing by subjects taking rosiglitazone or thiazolidinediones; ²⁹⁷ most of these patients also had peripheral edema. We did not, however, find any reports of macular edema in our review, since most of the reported events came from case reports. This potential adverse effect will need further investigation in observational studies. Thiazolidinediones were also associated with anemia, with an average drop in hematocrit of 1–3% that is likely not clinically relevant unless the individual already has moderate to severe anemia. Thiazolidinediones are thought to cause edema and anemia by increasing the plasma volume. ²⁷⁷ The FDA data indicated that hematocrit decreased by more than 10% in fewer than 1 in 100 subjects. However, although the anemia develops infrequently, it seems reasonable to check for anemia after starting thiazolidinediones.

CHF is an adverse event that is mentioned on the product label for the thiazolidinediones; the label cites two different studies: One compared pioglitazone with glyburide, and the other was a study of rosiglitazone plus insulin, showing increased CHF in the thiazolidinedione arms. ^{298, 299} Rosiglitazone is also currently contraindicated in patients with New York Heart Association (NYHA) class III or IV CHF. ²⁹⁹ In our review, thiazolidinediones also conferred a greater risk of congestive heart failure than did metformin or second generation sulfonylureas. We found several RCTs that reported on CHF as an adverse event, including PROactive, a one-year, large randomized, double-blind trial. ¹⁸⁵ However, many studies did not report on this adverse event, even to state that there were no events. Although a few observational studies have evaluated this outcome, they were limited by their ability to address key confounders such as HbA1c control, duration of diabetes, blood pressure level, adherence to medications, and medication dosing.

Summaries of data with regard to withdrawals due to unspecified adverse events were included in many studies, but these data have limited usefulness because the authors did not identify specific reasons for the withdrawals. Eighteen percent of trials did not report on withdrawals or losses to follow-up, and it would be important to know whether the withdrawals that were reported were due to specific adverse events. If trials reported this information consistently, comparative data would be more meaningful for this outcome.

Several caveats deserve mention with regard to adverse events. First, while almost all studies reported the incidence of hypoglycemia, they were inconsistent in terms of reporting on other adverse events. Second, the definitions of adverse events varied across studies and were often aggregated. For instance, GI events could be defined as nausea, vomiting, abdominal pain, flatulence, or a mixture of the above, making it difficult to compare across studies. Third, as expected, the incidence of adverse events was generally higher in cohort studies, which were of longer duration, than in the short-duration RCTs. The estimates from the RCTs (with the exception of the longer-duration UKPDS) are therefore likely to be lower than what one would expect in diabetic subjects over time. Also, because of potential publication bias, cohort studies may be more likely to get published if a difference between medications is shown. However, the cohort studies have generally shown between-group differences that were similar to those in the RCTs. Fourth, few RCTs evaluated certain outcomes, such as elevated liver transaminases, CHF, anemia, cancer, and allergic reactions; therefore, we relied on a small number of cohort studies for many of these outcomes. The available cohort studies, however, were limited by their ability to adjust for key confounders such as HbA1c levels, blood pressure, lab data, duration of diabetes, adherence to medications, and doses of medications. Finally, most studies occurred in patients without contraindications, such as renal or hepatic insufficiency; therefore, we cannot generalize these findings to those specific co-morbidities.

Several general limitations to this systematic review should also be kept in mind. First, we did find that the study populations were fairly similar to the general population of adults with type 2 diabetes. However, many studies excluded individuals with complications of diabetes or other co-morbidities, as well as less adherent subjects, thereby limiting the generalizability of the findings.

Second, we excluded articles comparing oral diabetes medications with insulin, since insulin was not part of this review. Therefore, we may have missed some relevant information regarding the oral diabetes medications of interest. However, these data would have been indirect comparison data and less strong than the head-to-head data we present. We also did not study the oral diabetes medications in combination with injectable diabetes medications such as insulin, amylin, or exenatide, another potential limitation to the generalizability. In the UKPDS, 22–40% of subjects had insulin added to their regimen after 6 years of follow-up; ⁹³ such addition of insulin could lead to new adverse events or exacerbation of existing adverse events. Also, a systematic review already exists that evaluated insulin in combination with oral medications, as well as insulin monotherapy. ³⁰⁰

Third, we only included articles that had more than 20 participants in each arm or a total of 40 participants or more in the study. We felt that very small studies were unlikely to influence our conclusions, given the large number of studies that were available for inclusion in this systematic review. Fourth, we only evaluated specific safety issues for which there was an a priori hypothesis of potential harm. However, we did abstract data on all serious adverse events as well as all the well-known side effects. We also evaluated FDA data, paying particular attention to safety concerns. Therefore, we do not feel we have missed any clinically relevant concerns.

In conclusion, oral diabetes medications had similar effects on glycemic control and slightly different effects on other proximal outcomes. We were unable to draw firm conclusions about the differences among oral diabetes medications in terms of their effects on distal outcomes. Safety differences did exist among the oral diabetes medications and deserve further investigation.

Future Research

This review of existing evidence identified a number of issues requiring further research. These remaining issues are grouped by key question below.