Association between use of sodium-glucose cotransporter 2 inhibitors, glucagon-like peptide 1 agonists, and dipeptidyl peptidase 4 inhibitors with kidney outcomes in patients with type 2 diabetes: A systematic review and network meta-analysis

We comprehensively searched PubMed, Embase, and the Cochrane Central Register of Controlled Trials (CENTRAL) from database inception to September 28, 2020. Details of the search strategy are presented in Appendix 1 in S1 Text. An additional manual search of the reference lists of included studies, published meta-analyses, and ClinicalTrials.gov was carried out to identify additional trials.

We selected the trials based on the following inclusion criteria: (1) RCTs (either single-blind, double-blind, triple-blind, quadruple-blind, or open-label) in patients with type 2 diabetes mellitus and were published in the English language; (2) Compared SGLT-2 inhibitors, GLP-1 agonists, and DPP-4 inhibitors at market-approved doses with each other or with a control group (defined as placebo or no treatment) (Appendix 12 in S1 Text); (3) Trial durations ≥12 weeks; and (4) Trials provided data on any of the predefined primary and secondary outcomes.

Exclusion criteria included a history of diabetic ketoacidosis, type 1 diabetes mellitus (T1DM), pancreas or beta cell transplantation, diabetes caused by pancreatitis, or pancreatectomy.

After removing duplicate records, the author (S.Y.) was responsible for reading the title and abstract of identified studies to assess eligibility. Two authors (W.H. and Y.C.M.) independently extracted data using a standardised data extraction form (Appendix 13 in S1 Text). They assessed the full texts of there maining results based on the prespecified inclusion criteria to determine the final included studies. In cases of disagreement, a third author (L.Z.) was consulted to reach a consensus.

Two reviewers (W.H. and L.Z.) independently extracted data using a predefined data extraction form, based on relevant templates in the Cochrane Handbook for Systematic Reviews [12]. Disagreements were resolved by discussion or consensus with a third reviewer (S.Y.). Data extracted included study characteristics (first author, publication year, study drug and control treatments, and duration of follow-up) and patient characteristics (background treatments, mean age, the proportion of men, duration of T2DM, baseline HbA1c, body mass index, mean eGFR, pre-existing chronic kidney disease (%), pre-existing cardiovascular disease (%) and background ACEi/ARB treatment (%)). If multiple reports from the same population were retrieved, only the most complete and/or most recently reported data would be used. If kidney events were not reported in the manuscripts, data was extracted from the “serious adverse events” section of the Clintrials.gov website. Only data from the randomized controlled periods was used for trials with open-label extension periods. Studies with multiple treatment groups were incorporated into the network as multi-arm studies where appropriate, allowing for multiple pairwise comparisons from a single study to be incorporated into the network without duplication of data. For studies with insufficient information, the reviewers contacted the primary authors, when possible, to acquire and verify the data. We contacted 17 authors of relevant articles, of whom five replied, but they could not provide any additional study data.

The quality of eligible studies was assessed by two authors (Y.C.M. and L.Z.) independently using the Cochrane Collaboration risk of bias tool across five domains (sequence generation, allocation concealment, blinding, detection bias, and attrition bias) [12]. Any disagreements were resolved through consensus. Each domain was judged as low risk, high risk, or unclear risk. Since it was helpful to evaluate the overall quality of evidence from network meta-analysis, the contribution plot was used to analyze the contribution of each direct comparison to the assessment of each network meta-analysis summary effect. Furthermore, a comparison-adjusted funnel plot was used to identify small study effects. Furthermore, the Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework was used to assess the quality of evidence contributing to each network estimate, which characterizes the quality of a body of evidence for the outcomes based on study limitations, imprecision, inconsistency, indirectness, and publication bias [13].

The primary outcome was composite renal events that included the development of new-onset macroalbuminuria, a decline in estimated glomerular filtration rate (or increase in creatinine), renal impairment, renal failure, and progression of end-stage kidney disease, renal replacement therapy, or renal death [14]. The secondary outcome was acute kidney injury (AKI) events. AKI was defined by each trial or, according to the Medical Dictionary for Regulatory Activities preferred terms for serious adverse events.

The network meta-analysis uses a Bayesian approach and allows comparisons among all studies’ treatment arms, including direct and indirect comparisons simultaneously. A direct meta-analysis was carried out using the DerSimonian and Laird random-effects models to estimate pooled odds ratios and 95% confidence intervals for direct comparisons between therapeutic regimens. Statistical heterogeneity was assessed with the I² statistic, with values over 50% indicating substantial heterogeneity [15]. Analyses were performed using Markov chain Monte Carlo methods in the network meta-analysis. We performed a Bayesian hierarchical network meta-analysis. Fixed-or random-effects models were selected for each outcome based on the deviance information criterion (DIC), using the model with the smallest value (Appendix 5 in S1 Text). The transitivity assumption underlying network meta-analysis was evaluated by comparing the distribution of clinical variables that could act as effect modifiers across treatment comparisons [16]. A design-by-treatment test [17] was used to check the assumption of consistency across the entire analytical network. The node splitting method was used to assess the local inconsistency of the model by separating evidence for a particular comparison into direct and indirect evidence. If P > 0.1 in the node-splitting analysis, there would be no significant inconsistency. The global heterogeneity was assessed with the I²-statistic. The surface under the cumulative ranking curve (SUCRA) and mean ranks were used to rank the treatments for each outcome, which were also built to increase the estimate precision of the relative effect sizes of comparisons, enabling conclusions to be drawn for each outcome of interest.

We compared treatment effects in the following subgroups: (1) Trials with event-driven cardiovascular outcome versus non-cardiovascular outcome; (2) Trials with chronic kidney disease (CKD) versus non-CKD (CKD trial was defined as the mean eGFR <90 ml/min/1.73 m² or the proportion of patients with CKD>60%); (3) Trials with cardiovascular disease (CVD) versus non-CVD (CVD trial was defined as the proportion of patients with CVD>50%); and (4) Trials with ACEi/ARB treatment versus non-ACEi/ARB treatment (ACEi/ARB treatment trial was defined as the proportion of background therapy ACEi/ARB >70%). The sensitivity analysis that was performed was restricted to trial duration (excluding trials with ≤24 weeks of follow-up) and risk of bias (excluding open-label trials). We also performed network metaregressions on BMI at baseline, age, HbA1c, duration of diabetes, and eGFR on the outcomes to better understand the influence of patient characteristics on drug efficacy.

The GeMTC package on R (version 3.6.3) will be used for statistical analysis on network meta-analysis, assessment of global heterogeneity, local inconsistency, and network meta-regression; STATA 14.0 software will be used for pairwise meta-analysis, estimation of global inconsistency, transitivity, and funnel plot.

Fig 1 shows the study selection process. We retrieved 17 445 studies, of which 98 articles enrolling 186 335 participants were included for the network meta-analysis. For direct comparisons, 39 trials compared an SGLT-2 inhibitor with control, 25 trials compared a GLP-1 agonist with control, 22 trials compared a DPP-4 inhibitor with control, and 12 trials compared between drug classes. Of the 98 included studies, 69 were designed as composite renal outcome trials, and 63 were designed as acute kidney injury outcome trials (Appendix 4 in S1 Text). Fig 2 shows the network of eligible comparisons for composite renal events and acute kidney injury events.

Appendix 2 in S1 Text shows the characteristics of the studies. For composite renal outcome trials, the mean sample size was 2467 (range, 47–17, 160), and the mean duration of follow-up ranged between 12 and 302.4 weeks (median, 52; interquartile range, 24–104). The participants’ mean age was 60.4 years, with a median duration of diabetes of 10.97 years.The mean baseline glycated hemoglobin level was 8.13%, and the mean body mass index was 31.19 kg/m². The baseline characteristics of studies were deemed sufficiently similar based on sex, age, hemoglobin A1c (HbA1c) levels, and body mass index (BMI) to permit network comparison.

Of the 98 included studies, 45 (45.9%) were at low risk of bias across all domains. 17 (17.3%) reported adequate random sequence generation; 37 (37.8%) reported adequate allocation concealment. 6 (6.1%) were at high risk of bias for blinding and 17 (17.3%) for attrition bias (Appendix 3 in S1 Text). All six high-risk blind bias trials are open-label, and there are deviations from intended interventions in all of them. All of the trials were funded by industry.

The analyses of composite renal events included 69 studies that had enrolled 162,959 participants who reported at least one event in any group. In all, there were 9163 events: 1342 (4.9%) of 27 208 participants treated with SGLT-2 inhibitors, 1947 (6.3%) of 30 803 treated with GLP-1 agonists, 1079 (3.7%) of 29 557 treated with DPP-4 inhibitors, and 4795 (6.4%) of 75 391 in the control groups. The results of the pairwise meta-analysis are presented in Appendix 4 in S1 Text. GLP-1 agonists were significantly associated with a lower risk of composite renal events than the control groups (OR 0.84, 95% CI 0.78–0.89, I² = 0.0%). However, there were no significant differences found between DPP-4 inhibitors and SGLT2 inhibitors and the control groups (OR 1.05, 95% CI 0.96–1.15, I² = 0.0%) and (OR 0.88, 95% CI 0.68–1.14, I² = 70.1%), respectively. GLP-1 agonists (OR 0.87, 95% CI 0.21–3.65, I² = 0.0%) and SGLT-2 inhibitors (OR 1.23, 95% CI 0.61–2.51, I² = 0.0%) had no significant difference compared with the DPP-4 inhibitors. Overall, there was no evidence of significant heterogeneity observed, with one exception found between SGLT2 inhibitors and the control groups (I² = 70.1%). The heterogeneity was reduced when we excluded one large study [18] and an outlying study [19] (I² = 39.3%), but the association was not substantially altered (OR 0.85, 95% CI 0.68 to 1.06).

The analyses of AKI events included data from 63 trials reporting 2002 events among 167,584 patients. The results of the pairwise meta-analysis are presented in Appendix 4 in S1 Text. SGLT2 inhibitors were significantly associated with a lower risk of acute kidney injury events than the control groups (OR 0.75, 95% CI 0.65–0.86, I² = 0.0%). No statistically significant difference was observed in other head-to-head comparisons. Overall, there was no evidence of significant heterogeneity observed (I² = 0.0%).

For our primary outcome, the network meta-analysis revealed no significant difference between drug classes (Fig 3). According to the network contribution plots (Appendix 3 in S1 Text), the comparison of placebo versus SGLT-2 inhibitors and GLP-1 agonists had the major impact in entire networks, with 33.7% and 32.8% for composite renal events, respectively.

Network meta-analysis suggested that, when compared with the control groups (OR 0.74, 95% CI 0.62–0.87), GLP-1 agonists (OR 0.76, 95% CI 0.59–0.96) and with DPP-4 inhibitors (OR 0.67, 95% CI 0.50–0.86), SGLT-2 inhibitors were associated with a lower risk of acute kidney injury events (Fig 4) The comparison of control groups versus SGLT-2 inhibitors, DPP-4 inhibitors and GLP-1 agonists had the major contribution in entire networks, with 32.9%, 32.1% and 32.8% for AKI events, respectively (Appendix 3 in S1 Text).

We estimated the ranking probabilities of each treatment according to SUCRA, GLP-1 agonists were most likely to rank best for reducing composite renal events, with a probability of 80%, followed by SGLT-2 inhibitors and DPP-4 inhibitors, both with a probability of 49%. For AKI events, SGLT-2 inhibitors were most likely to rank best (Table 1 and Fig 5).

We found that the mean BMI, mean age, mean baseline HbA1c, and mean duration of diabetes were all reasonably similar across treatment comparisons when we used box plots to measure transitivity (Appendix 6 in S1 Text). The global inconsistency test found no significant difference between the consistency and inconsistency models for two outcomes (p = 0.902 for composite renal events and p = 0.626 for acute kidney injury events in Appendix 7 in S1 Text). After constructing the node-splitting models, we found that no significant inconsistency existed in all networks except for acute kidney injury events (Control vs SGLT-2i and DPP-4i vs SGLT-2i; Appendix 7 in S1 Text). Heterogeneity (global I²)was low for both our primary and secondary outcomes (10% and 15%, respectively). At visual inspection, comparison-adjusted funnel plots for composite renal outcomes appeared to be symmetric, indicating that there was no significant risk of publication bias in our included studies. However, the funnel plots suggest that there might be small-study effects for AKI events (S3 Fig and Appendix 3 in S1 Text).

We did a series of subgroup analyses for our outcomes (Appendix 8 in S1 Text). The effect estimate for composite renal outcome events did not change between the CVD and non-CVD groups, nor did the ACEi/ARB treatment and non-ACEi/ARB treatment groups. However, in our cardiovascular outcome group and CKD group, SGLT-2 inhibitors and GLP-1 agonists both significantly lowered the renal composite outcome risk compared to the control groups. Subgroup analyses found that SGLT-2 inhibitors were more effective than the control groups and DPP-4 inhibitors in lowering AKI events in patients with CVD, CKD, and ACEi/ARB treatment. Sensitivity analyses did not affect the associations of SGLT-2 inhibitors with reduced AKI events compared with the control groups, DPP-4inhibitors, and GLP-1 agonists (Appendix 9 in S1 Text). Network meta-regression indicated that covariates had no effect on outcome except for DPP-4 inhibitor versus control groups, and that composite renal outcome events would be influenced by eGFR at baseline (Appendix 10 in S1 Text).

According to GRADE, the quality of evidence ranged from very low to moderate, but was classified as low or moderate in the majority of comparisons. The quality was low for GLP-1 agonists versus control groups, moderate for DPP-4 inhibitors versus control groups, and very low for SGLT-2 inhibitors versus control groups in composite renal events. For AKI events, the quality of SGLT-2 inhibitors against control groups, DPP-4 inhibitors, and GLP-1 agonists was low, low, and moderate, respectively. The overall ranking of treatments for our primary and secondary outcomes had a moderate quality of evidence (Appendix 11 in S1 Text).

Our study has some limitations, which are worth further discussion. First, there were no head-to-head kidney outcomes trials that directly compared the three new classes of antidiabetic drugs. The comparative effects were generated with indirect evidence. The clinical effect of the drug was evaluated by drug class rather than by individual drug type. Although this increases the power to detect treatment effects, and there could be different within-class treatments. Second, Some of the data for composite renal events come from adverse event reporting rather than trial data, of the 69 studies with composite renal outcome events, 11 (15.9%) were directly measured and 58 (84.1%) were evaluated based on adverse reports. In comparison to 8.04 percent in direct measurement trials, composite renal events accounted for 1.87 percent of adverse reports. None of the trials included were systematically designed to investigate AKI events. The majority of AKI event data comes from adverse event reporting rather than trial data. Such constraints reduce the validity of our meta-analysis. Third, there were differences in background treatments, patient characteristics, and outcome evaluation among studies that might contribute to heterogeneity. Finally, the limitations of our meta-analysis include the absence of patient-level data, restriction of subgroup analyses to the primary outcome, and our ability to examine only the secondary endpoints of particular interest reported by the investigators of the included trials.