Introduction

In 2011, there were 366 million people with diabetes worldwide, and this is expected to rise to 552 million by 2030, rendering previous estimates very conservative [1]. Excluding accidents, diabetes is the fifth cause of death for women and the fourth for men in United States [2]. Although there is general agreement on the first-line use of metformin in most patients with type 2 diabetes, the ideal drug sequence after metformin failure is an area of increasing uncertainty [3]. Sulphonylureas are the most commonly added oral antidiabetic drugs in this scenario; the dipeptidyl peptidase-4 (DPP-4) inhibitors may offer non-inferior, glucose-lowering efficacy with reduced risk of hypoglycemia and weight gain [4]. Moreover, DPP-4 inhibitors may protect pancreatic beta-cells from enhanced apoptosis, as demonstrated in animal models of diabetes [5], and improve markers of beta-cell function in type 2 diabetic patients [6].

According to popular therapeutic algorithms [7, 8], adjustment of diabetes therapy should be based on the HbA1c level and a change in therapy is recommended when HbA1c is above the target. The expected reduction in HbA1c after failure of the first drug (metformin) ranges between 0.62 and 1.0 % in patients treated with various adjunctive drugs (excluding insulin), while after failure of two drugs (mainly metformin and sulfonylureas), the HbA1c decrease ranges between 0.6 and 1.15 % [9, 10]. The greater the baseline HbA1c level of the patient, the greater the HbA1c reduction following therapy [11]. Tailoring therapeutic strategies to subsets of patients based on their likelihood to respond to particular drugs has been advocated in order to personalize care that maximizes efficacy of available therapies [12].

We were unable to identify any reviews that systematically evaluated clinical or biochemical baseline characteristics that predicted the HbA1c response to DDP-4 inhibitors. In this article, we performed a meta-regression analysis of randomized controlled trials (RCTs) that assessed the effect of different DPP-4 inhibitors on HbA1c level in people with type 2 diabetes. We also hypothesized that durability of glycemic control may be a surrogate marker to test the hypothesis that DPP-4 inhibitors influence beta-cell loss.

Materials and methods

Search strategy

We completed a systematic review of RCTs that met predetermined methodological criteria. We performed an electronic search for RTCs evaluating DPP-4 inhibitors in patients with not adequately controlled type 2 diabetes. We searched Medline, ISI Web of Knowledge, and Scopus from inception to 31 March 2013. The search had no population and language restriction. Our core search consisted of the terms: “type 2 diabetes”, “hemoglobin A1c”, “HbA1c”, “A1C”, “incretins”, “gliptins”, “DPP-4 inhibitors”, “vildagliptin”, “sitagliptin”, “saxagliptin”, “linagliptin”, “alogliptin”, and “randomized clinical trial”. We searched for additional trials in the prescribing information documents of approved medications, at relevant websites (http://www.clinicaltrialresults.org and http://www.clinicaltrials.gov), and in personal reference lists of recovered article. Study selection was conducted by two independent reviewers (D.G. and K.E.). The relevance of studies was assessed with a hierarchical approach on the basis of title, abstract, and the full manuscript. After the initial screening of titles and abstracts, the studies included by both reviewers were compared; disagreement was resolved by consensus. The inter-review agreement was calculated with κ statistics. We followed the PRISMA (preferred reporting items for systematic reviews and meta-analyses) checklist for reporting systematic reviews and meta-analyses [13].

Selection criteria

We included any study arm of RCTs. We specified that RCTs were included if they (1) were published in a peer-reviewed journal, (2) reported on patients aged 18 and older with type 2 diabetes, (3) reported the effect of DPP-4 inhibitors on the HbA1c level in subjects who were either drug naïve, or on background therapy with metformin or other oral agents, (4) included at least 30 subjects in every arm of the study, and (5) lasted at least 12 weeks. Trials were excluded if (1) they reported data on subjects who did not have type 2 diabetes, (2) the intervention included the initiation of two agents at the same time, (3) the doses of DPP-4 inhibitors were different from those approved in the clinical practice (sitagliptin, 100 mg once daily; vildagliptin, 50 mg twice daily; saxagliptin, 5 mg once daily; alogliptin, 25 mg once daily; linagliptin, 5 mg once daily), and (4) they were extension of previous RCTs. We also excluded, on “a posteriori” basis, studies in which a DPP-4 inhibitor was added to insulin, as evidence was very scanty, and local regulatory aspects vary across the world.

Data abstraction

Two investigators (D.G. and K.E.) independently collected the relevant reports using a standardized form, and disagreements were resolved by consensus. The following data were extracted from each retrieved article: name of the first author, year of publication, design of the study, sample size, duration of follow-up, drug, dose, and schedule used, and baseline characteristics. The principal outcome was the effect of DPP-4 inhibitors on HbA1c at the end of the study. Methodological quality of RCTs was not assessed owing to the non-comparative nature of the analysis.

Statistical analysis

The decrease of HbA1c from baseline at the end of treatment was the primary outcome. Heterogeneity between studies was assessed by using Q statistic and I 2 [14]. P value of Q statistic less than 0.10 was considered significant. If overall heterogeneity was significant, a random-effect model was used, otherwise a fixed-effect model was used. Meta-regression was applied to estimate the amount of heterogeneity explained. Meta-regression is a regression model that relates the treatment effect to study-level covariates, while assuming additivity of within-study and between-studies components of variance. Covariates considered included baseline HbA1c, baseline fasting glucose, mean age, duration of treatment with a DPP-4 inhibitor, type of DPP-4 inhibitor, previous diabetes treatment which the DPP-4 inhibitor was added to, and statistical evaluation of the trial (per protocol or intention-to-treat). Categorical variables were included in the model by means of dummy variables. The moderators were chosen using the criteria of the availability and within moderators with known effect on the absolute HbA1c reduction from baseline. Restricted maximum likelihood estimators were used to estimate model parameters [15]. Permutation test (using 1,000 re-allocations) was used for assessing the true statistical significance of an observed meta-regression finding [16]. How much meta-regression model explain heterogeneity of the effect among studies was quantified by the percentage reduction of between studies variability. We also grouped all trials in four length strata: 12–18, 24–30, 52, and 104 weeks, and calculated the weighted HbA1c decrease from baseline, and that derived by the meta-regression after adjustment for baseline HbA1c level. Data were analyzed using Stata, version 11.0 (Stata Corp., College Station, TX, USA). All statistical tests were two-sided, and P values < 0.05 were regarded significant.

Results

We identified 747 citations (525 from Medline, 987 from Scopus, and 999 from ISI), of which we reviewed 100 and selected 78 studies (supplementary references 1–78), with 79 arms and 20,503 patients analyzed for efficacy on HbA1c levels (Fig. 1). In particular, 725 citations were identified from the three databases after exclusion of overlapping, and the remaining 22 from the reference lists of the recovered articles. Agreement between observers on which studies to include was good: the scores for agreement between the two reviewers were 83 % after screening titles and abstracts, and approached 100 % after screening full-text articles. All studies were RCTs (Table 1). Most trials were multinational and sponsored by industry. The trials were published between 2005 and 2013, with 23 RCTs (29.5 %) published between 2010 and 2013 (31 March). All studies were of parallel group; most studies were of double-blind design and only four were of open label design. The duration of the studies ranged from 12 to 104 weeks: 23 arms were of 12–18 weeks, 44 arms of 24–30 weeks, 8 arms of 52 weeks, and 4 arms of 104 weeks. Seventeen studies enrolled drug naive patients; background diabetes treatment included one or more oral antidiabetic drugs (OADs) in 52 studies, and in 9 studies the subjects discontinued OADs prior to randomization. Mean age (weighted by sample size) of participants was 56 years, range 50–71.6 years.

Fig. 1
figure 1

Articles identified and screened for eligibility

Full size image
Table 1 Studies included in the analysis
Full size table

There were 24 arms with vildagliptin and 8,283 patients, 24 arms with sitagliptin and 5,129 patients, 12 arms with saxagliptin and 2,312 patients, 8 arms with linagliptin and 2,986 patients, and 11 arms with alogliptin and 1,793 patients. For all 79 arms, the mean baseline HbA1c value (weighted by sample size) was 8.03 % (64 mmol/mol), range 7.2–9.3 % (55.2–78.1 mmol/mol); the decrease of HbA1c from baseline was −0.74 % (95 % CI −0.80 to −0.67 %), with considerable heterogeneity (I 2 = 97 %, Q test P value < 0.0001) (Table 2). Adjustment for baseline HbA1c value produced similar results: HbA1c decrease was −0.71 % (95 % CI −0.76 to −0.66 %), considering a baseline HbA1c of 8 %. Also shown in Table 2 are the decrements of HbA1c level during treatment with the five different DPP-4 inhibitors: although the primary efficacy response, in terms of absolute HbA1c reduction from baseline, is not similar among the different DPP-4 inhibitors, the lack of head-to-head trials did not allow a direct comparison for significance.

Table 2 Effect of DPP-4 inhibitors on HbA1c reduction from baseline (Δ)
Full size table

Figure 2 shows the decrement of HbA1c according to different study duration. In absolute terms, the greatest HbA1c decrease was seen at 52 weeks (−0.88 %, 95 % CI −1.1 to −0.66 %), with high heterogeneity (I 2 = 98 %, P < 0.0001); the smallest HbA1c decrease was recorded at 104 weeks (−0.31 %, −0.5 to −0.13 %, I 2 = 97 %, P < 0.0001). This pattern of HbA1c decrement did not change when data were adjusted for the different baseline HbA1c values [set at 8 % (64.3 mmol/mol)]: −0.87 %, −1.01 to −0.73 % at 52 weeks, and −0.39 %, −0.60 to −0.19 %, at 104 weeks.

Fig. 2
figure 2

Effect of DPP-4 inhibitors on HbA1c reduction from baseline (random-effect model). The 79 arms from RCTs were divided according to study duration, with most studies lasting 24–30 weeks. The black columns indicate the mean weighted HbA1c decrease (with 95 % CI) from baseline for each time-stratum; the white columns indicate the same decrease as derived by the meta-regression, after adjustment for baseline HbA1c level

Full size image

The initial univariate meta-regression analysis tested the influence of baseline HbA1c on the HbA1c response to treatment and explained 22 % of variance between studies. Additional multivariate meta-regression analyses tested the influence of single covariate added one by one to the baseline HbA1c value: fasting glucose and type of DPP-4 inhibitor explained an additional 19 and 12 % of variance, respectively, while mean age of patients, length of treatment with a DPP-4 inhibitor, the use of previous antidiabetic drugs, and statistical evaluation of RCT resulted in no further increment of explained variance. Multivariate model that included all three significant factors in the univariate analysis explained 61 % of variance between studies. Figure 3a shows that DPP-4 inhibitors have a greater efficacy on HbA1c in patients with higher baseline HbA1c (slope = −0.263, P < 0.001): for every increase of 1 % baseline HbA1c there is a 0.26 % further reduction in the HbA1c response to treatment. On the contrary, DPP-4 inhibitors have a smaller efficacy in patients with higher fasting glucose (Fig. 3b) (slope = 0.010, P < 0.001): for a baseline HbA1c level of 8 %, the estimated reduction of HbA1c during treatment with DPP-4 inhibitors is −0.74 % (95 % CI −2.22–0.73 %) at a fasting glucose level of 150 mg/dl, and −0.58 % (−1.54–0.37 %) at a fasting glucose level of 190 mg/dl. Age was without effect (Fig. 3c), with a slope of −0.002 (P = 0.835).

Fig. 3
figure 3

Meta-regression of the HbA1c to DPP-4 inhibitors in relation to average baseline characteristics, including HbA1c (a), fasting glucose (b) and age (c). Each dot represents the decrease of HbA1c from baseline; the size of the dots is proportional to the inverse of variance

Full size image

Discussion

The efficacy of DPP-4 inhibitors on HbA1c levels in type 2 diabetes is in the same range as that of other antidiabetic drugs [9]; drugs of this class are often prescribed in combination with the first-line agent metformin or other oral antidiabetic drugs in order to achieve tight glycemic control. Meta-regression analysis with trial-level covariates can be used to explore the correlation between HbA1c reduction and baseline features of patients enrolled. The present analysis of 79 arms from 78 RCTs evaluated the efficacy of DPP-4 inhibitors on absolute changes in HbA1c level in patients with type 2 diabetes. In all arms, including 20,503 type 2 diabetic patients, we found an adjusted decrease of HbA1c from baseline of −0.71 %, which is in line with previous assessments performed on a smaller number of arms and patients [11]. It seems unlikely that future studies with current or future components of this class of drugs (DPP-4 inhibitors) may change these estimates, which also need to face the proportion of patients reaching the HbA1c target. The analysis of data for adults with self-reported diabetes from the National Health and Nutrition Examination Survey and the Behavioral Risk Factor Surveillance System over the 1999–2010 period [17] shows that 52 % of survey participants achieved the HbA1c target <7 % (from 2007 to 2010).

Moving away from “one-size-fits-all” medicine, personalized medicine has the potential of tailoring therapies to subsets of patients based on their likelihood to respond to therapy. The challenge is how to identify these patients, and deliver personalized care that maximizes benefit [12, 18]. The distance from the target, i.e., the difference between the current HbA1c value of the patient and the individualized HbA1c target, may be useful as a predictor of therapeutic success [3]. Intuitively, the greater the distance from the target, the lower the probability to reach it. We have identified by meta-regression three factors that explain 61 % of variance between RCTs, with a change to identify a greater reduction of HbA1c to DPP-4 inhibitors. These predictors include, in decreasing order of importance, baseline HbA1c, baseline fasting glucose, and the type of DPP-4 inhibitor.

We have previously shown that the greater the baseline HbA1c level of the patient, the greater the HbA1c reduction following therapy with insulin or noninsulin drugs [11]. However, a separate analysis for each drug class was until now not available. The present analysis indicates a greater reduction of HbA1c (0.26 % more) in response to DPP-4 inhibitors for each 1 % point of baseline HbA1c >7 %. The mean baseline HbA1c level of participants in most RCTs included in the present analysis ranged from 7.5 to 8.5 %: the mean HbA1c response to treatment may therefore range from −0.57 % (at a baseline HbA1c of 7.5 %) to −0.83 % (at a baseline HbA1c of 8.5 %), with a mean adjusted response of −0.71 % (at a baseline level adjusted to 8 %). Although this difference may appear small, it is quite sufficient to achieve the individualized HbA1c targets if they were set at 7 or 7.7 %, respectively.

The second determinant is baseline fasting glucose, as it can modulate the effect of baseline HbA1c. We have found that the greater the baseline fasting glucose level of the patient, the smaller the HbA1c reduction following therapy with DPP-4 inhibitors. Considering that the mean baseline HbA1c level of all 78 RCTs is 8.03 %, we can predict that the −0.71 % HbA1c decrease for a baseline HbA1c level of 8 %, can move from −0.83 % at a fasting glycemia of 130 mg/dl, to −0.75 % at a fasting glycemia of 150 mg/dl, to −0.58 % at a fasting glycemia of 190 mg/dl. So, at the same baseline HbA1c level of 8 %, there may be a benefit of 0.25 % HbA1c decrease at lower fasting glucose level (130 mg/dl). Intuitively, a greater HbA1c decrease at lower fasting glucose levels suggests an effect more pronounced for DDP-4 inhibitors on nonfasting (postprandial) glycemia. Although this hypothesis may be reasonable, as supported by the physiological postprandial increase the levels of endogenous glucagon-like peptide-1 and gastro-intestinal polypeptide, not all RCTs reported the assessment of nonfasting glycemia, and the methods used for measurement of postprandial glucose vary widely across trials, preventing any reliable meta-regression. From a practical point of view, relying on two easily available factors (baseline HbA1c and fasting glucose) allows predicting a more (or less) HbA1c reduction in response to DPP-4 inhibitors in type 2 diabetes.

The third determinant of HbA1c decrease relates to the different type of DPP 4-inhibitor which, however, is a less important predictor of HbA1c response. It must be recall that the evidence of superiority, or non-inferiority for a drug versus another can only come for head-to-head trials, which at the moment are very few: only one head-to-head RCT demonstrated the non-inferiority of saxagliptin versus sitagliptin [19].

Declining beta-cell function is the predominant reason for deterioration in glucose tolerance. Our analysis shows that the efficacy of different DPP-4 inhibitors to lower HbA1c declines after 2 years on treatment. Only four studies lasted 104 weeks (supplemental references 17, 21, 42, 65) and compared a DPP-4 inhibitor (vildagliptin, sitagliptin or linagliptin) versus a sulfonylurea (gliclazide, glipzide, glimepiride), showing a non-inferior reduction of HbA1c. We excluded two extension studies comparing vildagliptin [20] or sitagliptin [21] versus metformin, with divergent results (one favoring metformin, one with similar efficacy): data from extended trials are more likely to be biased, as those patients who had loss of glycemic control were not enrolled in the extension part of the RCT. The results of two ample, recently published RCTs [22, 23] throw some light on the importance of long-term studies and durability of glycemic control. One RCT [22] assigned 16,492 patients with type 2 diabetes to receive saxagliptin (5 mg/day) or placebo, and followed them for a median of 2.1 years (maximum follow-up time was 2.9 years): HbA1c levels were significantly lower in the saxagliptin group than in the placebo group at 1 year and at 2 years (difference 0.3 %), although at the end of the treatment period the difference was smaller (0.2 %). In the other RCT [23], a total of 5,380 patients with type 2 diabetes underwent randomization to receive alogliptin or placebo, and were followed for up to 40 months (median, 18 months): by the end of the study period, the mean change from baseline was −0.33 % in the alogliptin group, although the maximal HbA1c decrease occurred at 3 months. So, the results of these largest RCT are concordant with our analysis indicating that at 2 years HbA1c decrease with saxagliptin or alogliptin may be around 0.3–0.4 %.

The present analysis does have limitations. The main limitation relates to meta-regression by itself, as it may be exposed to the risk of ecological fallacy, i.e., the impulse to apply group level characteristics onto individuals within that group [24]. Moreover, a positive association was found between baseline HbA1c and baseline fasting glucose, both included in the final multivariate meta-regression model suggesting a possible problem of collinearity. Analyses on pooled, patient-level data are useful to find the true role of both these parameters on the HbA1c reduction; however, these data are seldom available, as most trials are sponsored by the industry. On the other, the large number of arms (79 arms) and patients (20,503), the multicenter nature of most RCTs, implying a wide ethnic representative, the 21.6 year range of mean age (50–71.6 years), the possibility to explore a wide range of baseline HbA1c (from 7.2 to 9.3 %), and the 61 % explained variance between RCTs may attenuate the limitation. Another limitation (more apparent than real) of the analysis is that we used the absolute HbA1c reduction, without taking into account the placebo effect. However, 46 out of the 78 RCTs (59 %) included in the analysis used a placebo arm, while the remaining used a comparator drug. Moreover, in the real word of the diabetic patients the placebo effect lies within the prescription of the drug by the physician. Moreover, the comparison with placebo is often requested by regulatory agencies of drugs. So, the absolute HbA1c decrease during treatment may be a better measure of what may happen in the real life. Finally, gray literature sources were not searched, which may introduce the potential of some publication bias.

In conclusion, our analysis of 20,503 type 2 diabetic patients using five different DPP-4 inhibitors (vildagliptin, sitagliptin, saxagliptin, linagliptin, and alogliptin) indicates that the mean adjusted HbA1c decrease from baseline is −0.71 %, with high heterogeneity. However, this effect can be modulated by baseline HbA1c and fasting glucose level, and also by the type of DPP-4 inhibitors. These three factors explain 61 % of the variance between studies: a greater HbA1c response is seen in patients with higher baseline HbA1c and lower fasting glucose level.