We comprehensively searched PubMed, EMBASE (Ovid SP), and Cochrane Central Register of Controlled Trials (Ovid SP) for studies published from the time when the databases were established to 26 January 2022. ClinicalTrial.gov was screened for unpublished studies. The reference lists of relevant published research studies investigating the risk of DKA of SGLT2 inhibitors in patients with T2DM were also screened for potentially relevant studies. The key terms searched in this study were based on the PICOS framework (Supplementary Table S1 and Supplementary Table S2). Duplicate records were removed with EndNote X9.

We included studies meeting the following criteria: 1) participants: adults (>18 years) with a diagnosis of T2DM; 2) interventions/comparisons: SGLT2 inhibitors, active antidiabetic drugs (we defined active antidiabetic drugs as antidiabetic drugs other than SGLT2 inhibitors), or placebo; 3) outcomes: reporting the risk of DKA; and 4) study design: published or unpublished randomized controlled trials (RCTs) limited to the English language. The exclusion criteria were as follows: 1) including pregnant participants; 2) animal experiments; 3) studies published in a language other than English; 4) published as abstract only; 5) including patients with prediabetes; and 6) DKA caused by T2DM.

All retrieved literature studies were identified by two independent reviewers (YL and SY), and data were extracted by a predefined form. Any discrepancies were resolved by discussion with a third reviewer (NS), as required. We extracted the data including the first author’s name, publication year, sample size, follow-up length, intervention and comparison, outcomes, and the characteristics of participants.

SGLT2 inhibitors with diverse doses were separated to several trials. If a study contained more than one SGLT-2 inhibitor or more than one dose of SGLT-2 inhibitors, we defined them as a different comparison.

Four reviewers (N.S., P.F., Y.L., and S.Y.) conceived the study. Two independent reviewers (Y.L. and S.Y.) assessed the risk of bias of all included studies according to ROB 2, a revised Cochrane risk-of-bias tool for randomized trials (Sterne et al., 2019), and the discrepancies were resolved by consulting the third reviewer (N.S.). The Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) was used to assess the certainty of the evidence for the outcome. Three reviewers (F.W., D.L., and S.Z.) conducted the analysis and interpreted the data. Two reviewers (Q.W. and H.Z.) checked the analysis data on the review. The members of the research team assessed the confidence rating for each comparison as high, moderate, low, or very low, based on the direct and indirect estimates. Discrepancies were resolved by discussions.

Treatment nodes were grouped by different kinds of active antidiabetic drugs and different doses of SGLT2 inhibitors. We drew network plots with the multinma package in R (version 4.1.3) (Multinma, 2020).

We conducted a network meta-analysis of randomized controlled trials that assessed the sparse network with a fixed-effect model (Efthimiou et al., 2019) and consistency model in the frequentist framework with a graph-theoretical method by the netmeta package in R (version 4.1.3) (Rücker et al., 2015). The effect size for assessing DKA safety was calculated as odds ratios (ORs) with accompanying 95% credible intervals (CIs). We calculated the consistency by node-splitting models (Van Valkenhoef et al., 2016). We calculated the P-score to rank treatments (Salanti et al., 2011). We assessed the global and local statistical heterogeneity with generalized Cochran’s Q. We estimated the variance in heterogeneity between studies using the DerSimonian–Laird random-effects model. We assessed transitivity using descriptive statistics from studies and population baselines (Cipriani et al., 2013).

We assessed publication bias using funnel plots and Egger’s test with the netmeta package in R. Multiple sensitivity analyses were carried out to assess the robustness of the final results, including the following: 1) this analysis was estimated in a Bayesian framework; 2) exclusion of studies with treatment duration <24 weeks; 3) exclusion of studies without a placebo control; 4) exclusion of the high risk of bias studies (exclusion of unblinded studies.); 5) exclusion of studies where the risk of DKA was 0 percent; and 6) exclusion of studies with fewer than 100 participants.

The literature search flow diagram is shown in Figure 1. After screening 534 articles and 6 registered clinical trials, a total of 36 studies between 2013 and 2021 were included in the meta-analysis according to predetermined criteria (35 articles and 1 registered clinical trial), involving 52,264 patients and 70 DKA events (Bode et al., 2013; Lavalle-González et al., 2013; NCT, 2013; Barnett et al., 2014; Rosenstock et al., 2014; Erondu et al., 2015; Haering et al., 2015; Roden et al., 2015; Rosenstock et al., 2015; Araki et al., 2016; Frías et al., 2016; Hadjadj et al., 2016; Mancia et al., 2016; Rodbard et al., 2016; Rosenstock et al., 2016; Ito et al., 2017; Neal et al., 2017; Søfteland et al., 2017; Aronson et al., 2018; Dagogo-Jack et al., 2018; Fioretto et al., 2018; Han et al., 2018; Hollander et al., 2018; Kawamori et al., 2018; Pratley et al., 2018; Ridderstrale et al., 2018; Rosenstock et al., 2018; Scott et al., 2018; Terauchi et al., 2018; Allegretti et al., 2019; Cho et al., 2019; Ji et al., 2019; Perkovic et al., 2019; Pollock et al., 2019; Cahn et al., 2020; Persson et al., 2021; Weng et al., 2021). Of the included studies, 21 were conducted in multinational country studies, and most studies were registered (35/36, 97%) and all published in English. The baseline characteristics of the included studies are presented in Table 1 and Supplementary Table S3, where 34 were two-arm studies and two were three-arm studies. Among the 36 studies (the retrieved SGLT2 inhibitors contain 1 study about bexagliflozin; 7 studies about canagliflozin; 8 studies about dapagliflozin; 10 studies about empagliflozin; 6 studies about ertugliflozin; 1 study about henagliflozin; 2 studies about ipragliflozin; and 1 study about tofogliflozin), 19 studies with different doses of SGLT2 inhibitors were compared; 27 studies compared SGLT2 inhibitors to placebo; 11 studies compared SGLT2 inhibitors to active antidiabetic drugs (active antidiabetic drugs retrieved include pioglitazone, exenatide, glimepiride, metformin, and sitagliptin); and two studies compared active drugs to placebo. The study population comprised 31,829 males (60.9%) and 20,435 females (39.1%), with the mean age being 59.3 years (ranging from 51.6 to 69.9 years); the mean HbA1c was 8.1% (ranging from 6.9% to 9.3%); the baseline mean BMI was 30.4 kg/m² (ranging from 25.4 to 35.0 kg/m²); the mean disease duration was 9.8 years (ranging from 3.3 to 17.7 years); and the mean duration of treatment was 61.7 weeks (ranging from 12.0 to 271.0 weeks). In addition, most trials were funded by pharmaceutical companies (34/36, 94%).

The overall risk of bias was low. The assessment of the risk of bias in the included studies is shown in Supplementary Table S4. Overall quality assessment indicated that more than half of the studies had a low risk of bias.

The network plots of each outcome are presented in Figure 2A and Figure 2B, presenting the results and quality of evidence for the different doses of SGLT2 inhibitors and the different active antidiabetic drugs. The inconsistency of the network meta-analysis is also evaluated in Supplementary Figure S1 and Supplementary Figure S2. Heterogeneity and intransitivity of the network meta-analysis were also evaluated (Supplementary Table S5–6, Supplementary Figures S3–S5).

In total, 36 studies including 52,264 patients were reported on the risk of DKA, with a total of 70 DKA events occurring at a rate of 0.13%. Intervention nodes included in this network meta-analysis were different doses of SGLT2 inhibitors, metformin, sitagliptin, glimepiride, pioglitazone, and placebo. The SGLT2 inhibitor was not associated with a statistically significant increase in the risk of DKA (Figure 3A). There was also no difference in the risk of DKA between different doses of SGLT2 inhibitors (Figure 3B). The global I² of pairwise was 0%, and the global I ² of the consistency model was 0%. The node split analysis showed that the results were consistent. The GRADE quality for the network meta-analysis is shown in Supplementary Table S7 and Supplementary Table S8.

The P-score of DKA for different kinds of active antidiabetic drugs is illustrated in Supplementary Table S9, and the P-score of DKA for different doses of SGLT-2 inhibitors is illustrated in Supplementary Table S10. A higher P-score indicated a higher risk of DKA. There were no significant differences in network estimates between different hypoglycemic agents, but the P-score of DKA suggested that different kinds of active antidiabetic drugs ranked for the risk of DKA. Supplementary Table S9 shows that the highest P-score was of glucagon-like peptide-1 agonists (GLP1RAs) (P-score = 0.7361). The SGLT2 inhibitor ranked the third (P-score = 0.5298). Likewise, Supplementary Table S10 shows that canagliflozin (100 mg) had the highest P-score (P-score = 0.7388), and the lowest was tofogliflozin (20 mg) (P-score = 0.3491). The P-score was not found to be dose-dependent.

Egger’s test showed that there is no publication bias for different doses of SGLT2 inhibitors (p = 0.10, Supplementary Figure S6), but there is a publication bias for different kinds of active antidiabetic drugs (p < 0.01, Supplementary Figure S7). The sensitivity analyses are presented in Supplementary Table S11 and Supplementary Table S12. Studies with a duration of less than 24 weeks were excluded in the sensitivity analysis. The results showed that all sensitivity analyses demonstrated consistency with the primary results, regardless of the inclusion or exclusion of studies lasting less than 24 weeks.

Our study is the first frequentist network meta-analysis of SGLT2 inhibitors investigating the risk of DKA; in addition, we included one unpublished trial from the ClinicalTrials database that provided additional DKA data. Apart from that, we performed the quality assessment on all the included literature studies, ensuring that the literature studies were of high quality. Our systematic review and network meta-analysis included a large pool of trials and patients retrieved through a comprehensive literature search, and we used up-to-date and rigorous methodological tools to assess the risk of bias for the outcome. We included clinical trials from inception through 26 January 2022, with access to additional long-term trials (19 long-term follow-up of 52 weeks or more) and recent studies. A major strength of this network meta-analysis was to compare the risk of DKA among different active antidiabetic drugs. Furthermore, our study supports the evidence of the risk of DKA among different doses of SGLT2 inhibitors, which was the first time to be reported.

Several potential limitations should be acknowledged. First of all, the duration of treatment varied widely among the included studies, ranging from 12 weeks to 271 weeks. Second, because the incidence of DKA was low, the included trials reported a relatively small number of DKA, even reported no DKA event, which caused too much sparse data. We assessed the sparse network with a fixed-effect model and used the sensitivity analysis which excluded the sparse data’ studies for reducing the impact of sparse data. Third, all included studies did not distinguish eDKA and DKA. Therefore, our outcome was DKA not eDKA. Fourth, due to the influence of data collection, the results of our meta-analysis are more relevant to a population that is predominantly male, white, older, longer duration of diabetes, higher body mass index (BMI), and higher HbA1c levels. Moreover, we did not conduct a subgroup analysis to explore the correlation between the risk of SGLT2i-associated DKA and the duration of diabetes because of the range of data variations. There were few studies about the relationship between DKA and the duration of T2DM. An analysis of data on the incidence of DKA in SGLT2 inhibitors from the FDA Adverse Event Reporting System examined that the range of duration was from 1 day to >8 years, and there was no significance due to the limited availability of treatment duration information in some reports (Fadini et al., 2017).