Analyses of Results From Cardiovascular Safety Trials With DPP-4 Inhibitors: Cardiovascular Outcomes, Predefined Safety Outcomes, and Pooled Analysis and Meta-analysis

The intensification of diabetes treatment appears to improve CV outcomes only in patients with recent-onset diabetes and after a prolonged latency (11,12). The FDA guidance requires the recruitment of patients at high CV risk, often with prolonged diabetes duration, advanced age, and renal failure (1), who are the least likely to benefit from any intervention regarding CV risk factors. In addition, most of those patients already receive multiple treatments for CV disease prevention. In the Steno-2 trial, the use of statins at the end of the intervention in the intensive versus conventional treatment arms was 0 vs. 3% at baseline and 85 vs. 22% at the end of the intervention stage of the trial, respectively (13). In SAVOR-TIMI 53, EXAMINE, and TECOS, statins were used at baseline by 78.3, 90.4, and 79.9% of the trial population, respectively (3–5). The residual CV risk left, in spite of statin treatment, was smaller, and therefore any CV superiority was more difficult to prove.

The FDA requirements for new ADD development programs were to exclude an unacceptable increased relative CV risk with an upper limit of 95% CI of <1.3 (1). For this reason, trials are designed to show noninferiority between treatment arms with respect to CV events while maintaining similar glycemic control (14): investigators are free to modify glucose-lowering medications other than the experimental drug in order to maintain good glycemic control in both treatment groups, and in fact, between-group differences in HbA_1c are small (3–5). Conversely, in superiority trials the study is designed to highlight the benefits of drug treatment: for example, in lipid trials additional lipid-lowering therapies could not be modified during the study, creating a sizable difference in the surrogate marker (e.g., LDL cholesterol) that would hopefully lead to improved CV outcome.

The FDA required the CV safety trials to extend for a minimum of 2 years. A longer trial duration increases the risk for missing data (6), possibly affecting the magnitude and even the direction of results (15). The short duration of CV trials is also driven by market issues, such as patent duration, the extravagant cost of large clinical trials (the three currently available trials with DPP-4 inhibitors alone, taken together, accumulated >85,000 patient-years of follow-up), and competition between pharmaceutical companies.

In summary, these trials should be called “CV safety trials” and not “CV outcome trials” owing to the trial goals and design, the patient population chosen, the minimization of glycemic differences between treatment arms, and the relatively short duration. The recent finding of a significant benefit of empagliflozin in a similarly designed CV safety trial (6) was completely unexpected; despite this unpredictable result, the design of those trials remains inadequate for exploring treatment benefits.

The FDA guidance specifies that CV end points should be prospectively and blindly adjudicated by an independent committee in all trials of clinical development of new drugs. The CV end points that must be adjudicated are CV mortality, myocardial infarction (MI), and stroke, combined in the composite end point, major adverse CV events (MACEs). Other CV end points specifically mentioned in the FDA guidance that can also be adjudicated are hospitalization for acute coronary syndrome (ACS), urgent revascularization procedures, “and possibly other endpoints” (1). Some of the companies have chosen to add hospitalization for unstable angina, urgent coronary revascularization, and/or hospitalization for heart failure (hHF) to the definition of MACE (Table 1), although these end points are less “clear-cut” and more difficult to define; this increases overall event rates, thus reducing required sample size, but it limits the reliability of results, augmenting background noise and biasing results in favor of “no difference” between treatment arms (16).

In the FDA 2008 guidance (1), patients with “higher CV risk” to be included in CV safety analysis were defined as “patients with relatively advanced disease, elderly patients, and patients with some degree of renal impairment.” Pharmaceutical companies and researchers followed these instructions yet chose different inclusion and exclusion criteria in different CV safety trials (Table 1). Patients in EXAMINE and ELIXA (4) had experienced recent ACS, whereas those in TECOS had previous CV disease (5). Conversely, in the SAVOR-TIMI 53 trial 21.7% of enrolled patients had no prior CV events but showed at least two risk factors (17). The selection of a higher-risk population increases CV event rate, thus reducing the sample size and/or the duration of the trial. For example, if the expected annual event rate is 3% instead of 2%, with a 5-year follow-up, the sample needed to prove CV safety according to the FDA requirement is 5,400 patients instead of 8,000 (16). However, the downside of using such a specific high–CV risk population is the external validity of trial results. Alogliptin was shown to be safe in patients with ACS during the last 90 days (4), but the inference of those results to patients with stable CV disease, or with high CV risk without prior events is of uncertain validity.

Unexpectedly, in the SAVOR-TIMI 53 trial, saxagliptin was associated with an increased rate of hHF (3.5 vs. 2.8%; hazard ratio [HR] 1.27 [95% CI 1.07–1.51]; P = 0.007) (3,18). This phenomenon occurred during the first year, with no significant difference thereafter. In EXAMINE, in which hHF was a component of a prespecified exploratory extended MACE end point, there was no statistically significant increase in the risk of first event of hHF with alogliptin versus placebo (3.1 vs. 2.9%; 1.07 [0.79–1.46]) (19). However, these results are still not sufficient to exclude a small increase in risk (20). In TECOS, in which hHF was a secondary outcome, the rate of hHF in the sitagliptin arm (228 [3.1%]) was not statistically or numerically different from that in the placebo arm (229 [3.1%]) (HR 1.00 [95% CI 0.83–1.2] P = 0.98) (5). The results of saxagliptin on hHF in the SAVOR-TIMI 53 trial could be a play of chance, a specific effect of that drug, or an effect of the class of drugs on specific (and still unidentified) subpopulations of patients. In fact, the TECOS and SAVOR-TIMI 53 populations, although similar in age (66 vs. 65.1 years) and diabetes duration (9.4 vs. 10.3 years), differ in other features: for example, inclusion of patients without prior CV disease (0 vs. 21.6%), HbA_1c (7.2 vs. 8.0%), ethnicity (whites 67.9 vs. 75.3%), renal impairment (estimated glomerular filtration rate <50 mL/min [9.3 vs. 15.6%]), and/or use of insulin at baseline (23.2 vs. 41.4%).

None of the major CV safety trials with DPP-4 inhibitors were specifically designed to assess the effect of these drugs on heart failure, which was a secondary end point or part of a composite end point. This is a typical case in which a meta-analysis can add relevant information (see below).

Common adverse events (AEs) can be easily identified in phase II/III trials; conversely, rare AEs can be missed by those studies due to insufficient sample size and/or duration of observation (21). In addition, most early trials with diabetes drugs include only relatively healthy populations, who may be at lower risk of AEs (22). Large-scale trials designed for CV end points may yield important results for other safety issues.

Targeting intensive glycemic control significantly increases the risk of severe hypoglycemia (23), which may increase CV disease risk (24).

However, there is a consistent heterogeneity in the incidence of severe hypoglycemia among trials, which could be attributable to various factors such as frailty, age, duration of disease, HbA_1c targets, and therapeutic strategies. Severe hypoglycemia tends to be positively associated with baseline HbA_1c; the higher the starting HbA_1c value of the patient, the higher the event rate of severe hypoglycemia.

Notably, incretin-based therapy should be void of hypoglycemic risk owing to the ability of these drugs to stimulate insulin secretion with a glucose-dependent effect. However, in SAVOR-TIMI 53, episodes of severe hypoglycemia were significantly more frequent in the saxagliptin arm (2.1 vs. 1.7%, P = 0.047) (3), whereas any hypoglycemic events were more frequent only in patients with a baseline HbA_1c <7%. Almost all episodes were observed in patients treated with sulfonylureas (SUs), especially in those with HbA_1c <7%. In EXAMINE (4) and TECOS (5), no increase in rates of hypoglycemia was reported with alogliptin and sitagliptin, respectively.

The observed adherence to oral hypoglycemic drugs in patients with type 2 diabetes varies from 48.5 to 72.5% (25). Drug-induced weight gain is a well-known predictor of nonadherence (25). In addition, the increase in body weight induced by some glucose-lowering drugs may have an adverse effect on quality of life and on some CV risk factors (26). Weight gain induced by treatment intensification is influenced by the amount of HbA_1c reduction, background therapy, and treatment strategy. When insulin is added to background metformin and SUs, weight gain is almost 3 kg; when TZDs are added the observed weight gain is >4 kg (27). In a meta-analysis of 40 randomized controlled trials, the use of secretagogues (SUs plus meglitinides) added to background metformin was associated with a weight gain of 1.9 kg (28).

Conversely, the use of DPP-4 inhibitors is not associated with weight gain: in SAVOR-TIMI 53, EXAMINE, and TECOS, no significant weight changes were observed. The main reasons for this may be related to lesser degrees of serious hypoglycemic events, to a delay in gastric emptying, and/or to a sense of satiety secondary to a direct hypothalamic effect (29).

Type 2 diabetes may be associated with increased risk of malignancies (30); in addition, it is possible that some glucose-lowering drugs increase the risk for all cancers or specific forms of cancer. Such a concern has been raised also for GLP-1 receptor agonists and DPP-4 inhibitors. GLP-1 receptor activation directly promotes cell proliferation and enhances cell survival in several tissues (31). It has been observed that GLP-1–based therapies can induce histological changes suggestive of pancreatic damage in rodents (32). An analysis of AEs reported to the FDA suggested that GLP-1 receptor agonists and DPP-4 inhibitors could be associated with pancreatitis and pancreatic cancer (33), but that result could have been heavily affected by underreporting and reporting bias. In fact, cohort studies failed to detect any signal of risk either for pancreatitis or for pancreatic cancer (34).

Reassuring findings are also available from SAVOR-TIMI 53, EXAMINE, and TECOS, which did not report any significant increase in risk of either pancreatitis or pancreatic cancer, although a modest trend toward an increased incidence of pancreatitis was reported in TECOS (3–5).

Finally, it has been observed in preclinical studies that the incidence of thyroid C-cell tumors is increased in rodents treated with GLP-1 analogs (35). Notably, in mouse and rat thyroid gland C cells, GLP-1 receptor densities are much higher than in humans and primates (36). The available CV safety trials confirm the safety of DPP-4 inhibitors in this respect; in fact, no case of thyroid medullary carcinoma was reported in SAVOR-TIMI 53, EXAMINE, or TECOS (3–5).

Meta-analyses are typically performed when there is a multiplicity of available studies that either provide inconsistent results or lack adequate statistical power because of insufficient sample size. CV trials performed with molecules of the same class, which can be assumed to have a similar profile of action on CV risk, can be combined in a meta-analysis, thus increasing their power to detect even small differences between treatment groups. However, it is unlikely that such a meta-analysis would provide any relevant additional information on the effect of DPP-4 inhibitors on major CV events, in which the sample size was calculated to have an adequate power to explore the principal end point (i.e., MACE).

On the other hand, a meta-analysis of CV trials could be useful for the exploration of secondary end points or of individual components of the principal end points, for which each trial is underpowered. A typical example is hHF, which was significantly increased with saxagliptin in the SAVOR-TIMI 53 trial (3). In the CV trial with alogliptin, the difference between treatment groups for hHF was not statistically significant, but this could have been the consequence of a smaller overall number of events, with a lower statistical power (19); on the other hand, no such effect was detected with sitagliptin in TECOS (5). A meta-analysis of all available trials with DPP-4 inhibitors (including those with a non-CV end point, but not TECOS, which was unavailable at the time) confirmed an increased risk for heart failure (37). However, when trials with a non-CV end point (i.e., phase II–III studies) were considered separately, no signal of risk was detected, with no significant difference across molecules (37). This result is consistent with the hypothesis that the increased risk observed with saxagliptin in SAVOR-TIMI 53 may have been related to the characteristics of the patients enrolled. In the case of heart failure with DPP-4 inhibitors, the number of available large-scale trials is insufficient to draw definitive conclusions through meta-analysis. Further insight may be provided by a pooled analysis of patient-level data from the CV safety trials (SAVOR-TIMI 53, EXAMINE, and TECOS), exploring the effect of DPP-4 inhibitor treatment on different subgroups of patients categorized for their baseline characteristics, with the molecule used as a categorical covariate. However, such an analysis would pose many technical and organizational challenges, considering differences in protocols across studies.

Information on CV events provided by meta-analyses and pooled analyses is particularly relevant when no CV outcome study is yet available. Under current regulations, most new drugs are marketed on the basis of a clinical program composed of studies with metabolic end points, with a large-scale, noninferiority CV study performed only after marketing, to confirm the safety of the drug. While long-term studies are ongoing, combined analyses of shorter-term trials can provide some insights on possible CV actions of new treatments. In the case of DPP-4 inhibitors, pooled analyses of CV events in randomized phase II–III studies have been published for all molecules (38–42); all these analyses yielded risk estimates <1, although differences from comparator groups were not statistically significant (Table 2). Data from different analyses are not comparable, because the definition of events and the trial entry criteria are heterogeneous. In addition, in some of the analyses CV events had not been adjudicated (38).

Pooled analyses, which combine patient-level data, can be performed only by companies that hold the property of data collected during clinical development of new drugs. Independent researchers, without access to patient-level data, can only pool trial-level results from studies in a post hoc meta-analysis, which allows the combination of trials with different drugs from the same class, estimating an overall class effect. Since some of the trials submitted to regulatory authorities remain unpublished, it is important that all disclosed trials, including those not published in medical journals, are included in the meta-analysis. Notably, in such analysis only the crude number of events is available, whereas time to event remains unknown; this produces an imprecision in risk estimates, which is negligible in short-term studies but which can be relevant if the duration of treatment is very long. The lack of adjudication of events and the short duration of treatment are also major limitations for post hoc meta-analyses. Furthermore, it is possible that some trials with unfavorable results for experimental drugs remain undisclosed or that CV outcomes are only partially reported, thus producing a systematic bias. All these methodological issues suggest caution in the interpretation of results from post hoc meta-analyses. However, it is notable that all post hoc analyses (8–10,43) reported a significant reduction in the incidence of major CV events, with risk estimates ranging from 0.36 to 0.71 (Table 3); the only exception is a meta-analysis that included two CV studies (SAVOR-TIMI 53 and EXAMINE) along with earlier trials, with the former largely driving the overall result (44). One of the largest of those meta-analyses (8) also reported a significant reduction of all-cause mortality—the only end point unaffected by event adjudication.

The difference in results between pooled analyses and meta-analyses of early studies, on one side, and CV trials, on the other, is quite substantial. The methodological limitation of pooled analyses and meta-analyses summarized above, although relevant, does not seem sufficient for inducing such a large distortion. It has been suggested that differences in results could be determined by the duration of trials, with DPP-4 inhibitors producing a benefit in the short-term, which is lost with longer-term treatment (45). This view is not confirmed by CV trials, in which no evidence of a temporary benefit is present in the first few months after randomization; in addition, a significant reduction of major CV events is observed in non–CV outcome trials with DPP-4 inhibitors with duration of treatment >52 weeks (46). Another possible explanation for the observed difference is the diversity in the characteristics of enrolled patients. Subjects recruited in CV studies, who are selected for high CV risk, are typically older with longer duration of diabetes, greater impairment of renal function, higher comorbidity, and higher number of concomitant treatments than those participating in earlier trials with metabolic end points. It is possible that DPP-4 inhibitors produce a CV benefit in relatively low-risk patients, whereas they are neutral in subjects at very high risk or with prior CV events.

The enrollment of subjects at high CV risk is a practical strategy for yielding a high number of events with a limited sample size and duration of observation. It is also reasonable to verify safety in those patients who are at greatest risk of adverse outcomes. On the other hand, patients enrolled in CV trials with DPP-4 inhibitors are scarcely/hardly representative of those actually receiving a prescription of those drugs in clinical practice.

The results of the first CV trials with DPP-4 inhibitors (3–5) did not meet the expectations of many clinicians, who hoped for a demonstration of a CV benefit, as suggested by meta-analyses of previous, shorter-term studies (8–10,43). However, CV trials provided exactly the result that they had been designed for: noninferiority versus placebo. CV trials in diabetes simply demonstrate that drugs do not catastrophically increase CV risk. The unexpected finding of a significantly beneficial effect of the active drug, as with empagliflozin (6), given the design of those studies, is almost a case of serendipity.

The explanation of differences in CV outcomes between earlier studies and CV safety trials is complex. Our opinion is that methodological issues are not sufficient to account for such difference and that the more favorable result of an earlier trial is the effect of the enrollment of younger, healthier patients with fewer complications, with a lower proportion of subjects receiving other medications for the reduction of CV risk. However, available data are not sufficient to confirm such a hypothesis, which could be proven only through a specifically designed CV outcome trial in patients with diabetes at low risk—which is very unlikely to be ever performed.

CV trials provide a huge amount of safety data. Large-scale trials of appropriate duration, although designed for major CV events, yield relevant information on other safety end points (for example, with DPP-4 inhibitors, pancreatitis). Exposure of several thousand patient-years to an experimental drug also warrants the detection of relatively infrequent AEs. However, in this latter case, unless results are consistent across trials with molecules of the same class, the interpretation of results can be overwhelmingly difficult. This is what happened for the unexpected (and problematic) finding of increased hHF with saxagliptin in SAVOR-TIMI 53.

Although the results of CV safety studies could underestimate the actual benefit of DPP-4 inhibitors on CV outcomes, a recent study with a similar design did show a significant reduction of CV morbidity and mortality with the SGLT2 inhibitor empagliflozin (6). Although the characteristics of patients enrolled in this latter trial (5) are not exactly the same as in trials with DPP-4 inhibitors (3–5), the fact that results with empagliflozin in patients with prior CV events were much improved compared with than those observed with saxagliptin, alogliptin, or sitagliptin should be clearly recognized.

Inevitably, information on safety is excellent for newer agents but largely missing for older drugs. This creates an information disequilibrium, for which safety issues have a greater probability of being raised for newer drugs, while they may remain unnoticed, or underestimated, for older agents. Another problem with the current approach is that populations enrolled in CV trials can be quite different from those that, on average, receive a prescription for a drug in the real-world. This can lead to an overestimation of safety issues with newer agents.

Finally, the most important drawback of CV safety trials is their excruciatingly high cost, which has an impact on the price, and therefore the widespread availability, of newer agents. Furthermore, pharmaceutical companies, being forced to invest huge budgets on placebo-controlled CV trials, tend to shrink their development programs on the side of phase III active-comparator controlled trials, which are probably more useful for clinical decision making and which could be more cost-efficient.

The decision of the FDA to require safety data from large-scale, specifically designed trials had the merit of focusing the attention of the scientific community on the relevance of CV safety of glucose-lowering drugs. At the same time, the unintended negative consequences of this decision may outweigh the benefits. As an alternative, CV safety could be assessed through pooled analyses of patient-level data from wider programs of active comparator-controlled phase III trials of appropriate duration, provided that a relevant fraction of patients at high CV risk is included and that MACEs are appropriately adjudicated. This would probably require a safety criterion less stringent than the present 1.3 for upper confidence limit (1); however, additional real-world data could be obtained through specific programs for postmarketing surveillance, thus avoiding specific CV safety trials.

Funding and Duality of Interest. E.M. has received speaking and/or consulting honoraria from AstraZeneca, Boehringer Ingelheim, Eli Lilly, Jannsen, Merck, Novartis, Novo Nordisk, Sanofi, and Takeda and research grants from AstraZeneca, Boehringer Ingelheim, Eli Lilly, Novartis, Novo Nordisk, Takeda, and Sanofi. O.M. has received speaking fees from AstraZeneca and Bristol-Myers Squibb, Novo Nordisk, Eli Lilly, Sanofi, Novartis, Merck Sharp & Dohme, and Kyowa Hakko Kirin; consulting fees from Novo Nordisk, Eli Lilly, Sanofi, Novartis, AstraZeneca, Boehringer Ingelheim, and Janssen; and research grants from AstraZeneca, Bristol-Myers Squibb, and Novo Nordisk. A.A. has received research grants from the Strategic Project DYCENDI from the University of Padova STPD11ALFE and has also received speaking honoraria from Novartis, Eli Lilly, Novo Nordisk, Sanofi, Boehringer Ingelheim, AstraZeneca, Mediolanum, Servier, Janssen, and Merck Sharp & Dohme; financial support for attending symposia from Novartis, Eli Lilly, Novo Nordisk, Sanofi, Boehringer Ingelheim, and AstraZeneca; financial support for educational programs from Novartis; consulting fees from Novartis, Eli Lilly, Novo Nordisk, Sanofi, Boehringer Ingelheim, AstraZeneca, Mediolanum, Servier, Janssen, and Merck Sharp & Dohme; and research grants from Boehringer Ingelheim and AstraZeneca. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. E.M., O.M., and A.A. contributed to collection of relevant references, discussion of their clinical implications, writing of the manuscript, and revision of the text.