Association of adverse respiratory events with sodium-glucose cotransporter 2 inhibitors versus dipeptidyl peptidase 4 inhibitors among patients with type 2 diabetes in South Korea: a nationwide cohort study

We conducted a retrospective cohort study using data from the Health Insurance Review and Assessment Service (HIRA) database of South Korea (data number: M20210607316). The HIRA database includes nationwide health insurance claims data on reimbursed healthcare utilization information, including but not limited to diagnoses (International Classification of Diseases, 10^th Revision [ICD-10] diagnostic codes, physician’s specialty), procedures (domestic codes), and prescriptions (domestic codes based on the active ingredient, date of prescription, days’ supply, dose) from all settings (inpatient, emergency department, outpatient, and nursing home) [29]. As South Korea has a universal, single-payer health insurance system that is operated and managed exclusively by the National Health Insurance, all residents, domestic or foreign, are registered in the HIRA database [30]. While the HIRA database has information on in-hospital deaths only, 75.6% of all deaths in 2020 were from medical institutions, with 15.6% from homes and 8.8% from other places [31]. The “end of enrollment” date is however not available in the HIRA database, which could prolong the duration of follow-up for non-hospital deaths, although these people would be naturally dis-enrolled from the national claims system. Yet, any misclassified follow-up time due to this issue is likely to be nondifferential between exposure groups to have any impacts on the study.

We identified all patients newly prescribed a SGLT2i or DPP4i between 1 Jan 2016 and 31 Dec 2020, given that these medications in this period were indicated and reimbursed exclusively for type 2 diabetes. Inclusion was restricted to new users of SGLT2is or DPP4is, defined as patients prescribed one of these drug classes with no prior record of prescription for either drug class during the prior year. We defined cohort entry as the date of the first prescription for a SGLT2i or a DPP4i. Patients meeting any of the following criteria were excluded: (1) aged less than 18 years on cohort entry to restrict to adults with T2D, (2) initiated both a SGLT2i and a DPP4i on cohort entry to ensure that our exposure groups were mutually exclusive, (3) recorded diagnosis of end-stage renal disease or recipient of dialysis any time before or on the date of cohort entry as these conditions are contraindications for SGLT2is, and (4) recorded diagnosis of any of the respiratory outcomes of interest (acute pulmonary edema, ARDS, pneumonia, and respiratory failure) in the year before or on the date of cohort entry to ensure that outcomes were newly-occurring rather than the sequelae of respiratory outcomes that occurred shortly before cohort entry.

Patients were classified into either the SGLT2i or DPP4i group depending on the drug class prescribed at cohort entry. Exposure was defined using an as-treated approach in which patients were considered exposure to their initial drug from cohort entry until an event or censoring due to switching to or adding the comparator drug, discontinuation of the cohort-entry defining drug, in-hospital death, or end of the study, whichever occurred first. Discontinuation was defined as a treatment gap of at least 30 days between the end of one prescription (defined by the days supply) and the start of the next prescription. We chose DPP4is as the comparator drug to SGLT2is as this drug class has similar indications, formulations, and is used at comparable stages for the treatment of T2D (e.g., second or third-line treatments), with no previous known associations with respiratory events.

The primary outcome of interest was respiratory events, defined as a composite endpoint that included acute pulmonary edema, ARDS, pneumonia, and respiratory failure (Additional file 1: Table S1 for ICD-10 diagnostic codes); these endpoints were selected based on acute, serious events related to abnormal lung function [32–34]. Secondary outcomes were the individual components of the composite respiratory outcome and in-hospital death. Study outcomes, or respiratory events, were defined by the presence of a relevant primary or secondary diagnostic code in an inpatient setting, and the event date was defined by the date of diagnosis. For the primary outcome, the event date was defined as the date of the first occurrence of any component of the composite endpoint. The overall positive predictive value for diagnostic codes from insurance claims was reported to be 82% overall, and 85% from tertiary hospitals for all general conditions, when compared to those from electronic medical records [35].

We assessed sociodemographic characteristics on the date of cohort entry and clinical covariates of comorbidities, co-medication use, level of antidiabetic treatment (proxy for diabetes severity), and healthcare utilization during the year before cohort entry. We then used logistic regression to estimate a propensity score (PS) [36] for the probability of receiving an SGLT2i versus DPP-4, with potential confounders included as independent variables (Additional file 1: Table S2). Level of antidiabetic treatment was categorized into three levels based on the type and number of different antidiabetic drugs classes prescribed in the year prior to cohort entry: level 1, patients not prescribed an antidiabetic drug or treated with only one antidiabetic drug; level 2, patients treated with ≥ 2 different classes of non-insulin antidiabetic drugs; level 3, patients treated with ≥ 1 insulin either alone or in combination with other antidiabetic drugs. SGLT2i and DPP4i users were matched 1:1 using the estimated PS and the nearest neighbor matching algorithm without replacement with a caliper width of 0.05 on the PS scale.

Baseline characteristics were summarized using descriptive statistics, with frequencies and percentages for categorical variables and means and standard deviations (SD) for continuous variables. Potential imbalances in patients’ characteristics were assessed before and after PS matching by using the absolute standardized mean difference (aSMD); aSMD > 0.1 was considered as a meaningful imbalance. Incidence rates of study outcomes per 1000 person-years and the rate differences with 95% confidence interval (CIs) between treatment groups were estimated assuming a Poisson distribution. For the primary analysis, we used cause-specific Cox proportional hazards models within the PS-matched cohort to estimate hazard ratios (HRs) with 95% CIs for the study outcomes associated with SGLT2is versus DPP4is, with in-hospital death treated as a competing event (individuals censored when this competing event occurred). We also evaluated the proportional hazards assumptions using Schoenfeld residuals, which revealed no violations.

In secondary analyses, we subclassified SGLT2i users into four subgroups (dapagliflozin, empagliflozin, ertugliflozin, and ipragliflozin) and compared each individual SGLT2i to DPP4i to estimate the relative effect of each SGLT2i on respiratory events. We also conducted stratified analyses according to age (< 65 and ≥ 65 years), sex, history of chronic obstructive pulmonary disease (COPD), history of asthma, baseline insulin use, and duration of follow-up (< 1 and ≥ 1 year) to assess any effect modifications of these factors on the comparisons between SGLT2is and DPP4is using the Wald test for heterogeneity. We further stratified by the history of cardiovascular disease (CVD), history of heart failure, and the history of chronic kidney disease (CKD) to examine whether the risk of adverse respiratory events associated with SGLT2is differed across these categories.

To examine the robustness of our main findings, we performed several sensitivity analyses. First, we repeated the main analysis using an intention-to-treat approach to define follow-up. In this setting, the initial exposure status is carried forward and treatment discontinuation and switching during follow-up were disregarded, thus, patients were followed from cohort entry until the earliest of outcome occurrence, in-hospital death, 365^th day of follow-up, or end of the study; the maximum duration of follow-up was limited to 365 days to minimize any bias arising from potential exposure misclassification when there is too lengthy follow-up. Second, we used another method for handling competing risk, the Fine–Gray subdistribution hazard model, to estimate HRs, with in-hospital death treated as a competing event [37]. Third, we varied the grace period duration used to define the exposure effect window from 30 days (primary analysis) to 0 and 60 days. Fourth, we applied a stricter definition for new users of SGLT2is or DPP4is by limiting the study cohort to patients with no record of prescription for the study and comparator drugs in the 2 years prior to cohort entry. Fifth, we defined the exposure of interest as use of a SGLT2i as monotherapy or in combination with metformin. Sixth, to avoid any potential confounding arising from the coronavirus disease 2019 (COVID-19) pandemic or related changes in clinical practice, we shortened the end of the study period from 31 December 2020 to 20 January 2020 (the first positive case of COVID-19 reported in Korea was on 20 January 2020). Seventh, we restricted events to those with primary (or principal) diagnosis codes only to minimize potential outcome misclassification. Eighth, we excluded patients who used insulin during the baseline period to restrict to patients with T2D with similar disease severity. Ninth, we conducted a negative (herpes zoster virus reactivation [38]) and a positive control outcome (hospitalization for heart failure [18]) analysis to assess the possibility of residual confounding in our data. Finally, we calculated E-values to assess the strength of association an unmeasured or unaccounted confounder would need to have to disregard the observed treatment-outcome association [39]. All analyses were performed with SAS Enterprise Guide version 7.1 (SAS Institute).

From 1,610,006 adult patients prescribed SGLT2i and DPP4i, we identified 205,556 users of SGLT2i and 1,264,730 users of DPP4i. After PS matching, 205,534 patients were included in each treatment group; 22 users of a SGLT2i were not matched to users of a DPP4i and thus, were excluded from further analyses (Fig. 1; c-statistic was 0.693 for the PS estimation model). The overall median follow-up duration in the PS-matched cohort was 0.66 years, where the mean duration was 1.14 (SD 1.18) years. All baseline characteristics pre- and post-PS-matching are shown in Table 1. The mean age of the entire PS-matched cohort was 53.8 (SD 13.2) years, and 59.4% were male. All baseline characteristics were well balanced between the treatment groups after PS-matching, with all aSMD < 0.1.

We found lower incidence rates of adverse respiratory events among SGLT2i users than among DPP4i users (incidence rate 4.54 versus 7.54 events per 1000 person-years; rate difference: − 3.00, 95% CI − 3.44 to − 2.55) in the PS-matched cohort. Within the PS-matched cohort, SGLT2i users had lower risks of the composite respiratory endpoint (HR 0.60, 95% CI 0.55 to 0.64) as well as the secondary outcomes of acute pulmonary edema (HR 0.35, 95% CI 0.23 to 0.55), pneumonia (HR 0.61, 95% CI 0.56 to 0.66), and respiratory failure (HR 0.49, 95% CI 0.31 to 0.76), than DPP4i users. The analysis for ARDS suggested that SGLT2i use was beneficial, but 95% CIs were wide due to sparse data (HR 0.44, 95% CI 0.18 to 1.05) (Table 2). Use of SGLT2i versus DPP4i was also associated with a reduced risk of in-hospital mortality (HR 0.46, 95% CI 0.41 to 0.51) (Additional file 1: Table S3). The analyses for individual SGLT2is showed a similar trend of a lower risk of the composite respiratory event across individual SGLT2i molecules versus DPP4is (Fig. 2; Additional file 1: Table S4).

Similar trends for reduced risk of the composite respiratory endpoint were observed among SGLT2i users in each age, sex, and history of asthma, history of COPD, history of CVD, history of CKD, and history of baseline insulin use subgroup. Greater benefits were observed among patients with previous insulin use (HR 0.46, 95% CI 0.39 to 0.55) than among those without previous insulin use (HR 0.62, 95% CI 0.57 to 0.67) (Fig. 2; Additional file 1: Tables S5-S12 for results of each of the stratified analyses).

Sensitivity analyses that examined study assumptions produced results that were similar to those of our primary analysis (Fig. 3; Additional file 1: Tables S13-22); for instance, the intention-to-treat analysis that did not censor follow-up at treatment interruption or the as-treated analyses that varied the grace period length were consistent to the main analysis (Additional file 1: Table S13). Moreover, the negative control outcome (HR 1.02, 95% CI 0.98 to 1.06) and positive control outcome (HR 0.69, 95% CI 0.61 to 0.78) analyses suggested minimal residual confounding (Additional file 1: Table S15). The E-value for point estimates in the as-treated analysis within the PS-matched cohort for respiratory events was 2.72, and 2.50 for CIs; the E-values for the secondary outcomes of acute pulmonary edema, ARDS, pneumonia, and respiratory failure were 5.16, 3.97, 2.66, and 3.5 for point estimates, and 3.04, 1.00, 2.40, and 1.96 for CIs, respectively.

In this nationwide retrospective cohort study, patients with T2D who received a SGLT2i had a 40% lower risk of adverse respiratory events than those receiving a DPP4i in a real-world setting. Similar lower risks were observed for acute pulmonary edema, pneumonia, and respiratory failure. Moreover, similar trends were observed across individual SGLT2i molecules, suggesting a class effect of SGLT2is. Our study findings were supported from a broad range of sensitivity analyses.

To date, only a few studies have investigated respiratory events associated with SGLT2is, despite its biological plausibility [11]. The lower risk of respiratory events with SGLT2is observed in the present study is consistent with prior meta-analyses that compared adverse respiratory events using data from placebo-controlled randomized trials of SGLT2is; these meta-analyses found a 25% lower risk of overall respiratory disorders [21], a 16% lower risk of pneumonia [22], a 48–60% lower risk of acute pulmonary edema [22, 23], a 29% lower risk of respiratory failure [23], and a 41% lower risk of asthma [24]. Another trial that compared dapagliflozin versus placebo in patients with COVID-19 also hinted at potential respiratory benefits albeit showing an inconclusive lower rate of respiratory decompensation (9.3% versus 11.2%; HR 0.85, 95% CI 0.60 to 1.20) [40]; network meta-analyses in patients with diabetes and COVID-19 also suggested respiratory benefits with SGLT2is by reporting lower COVID-19 mortality [41, 42]. Nevertheless, translating the results of these placebo-controlled trials of SGLT2is with respiratory events to routine clinical practice warrants further studies, given the highly dynamic treatment of T2D in a real-world setting.

No observational study, to our knowledge, has investigated the association between overall respiratory events and SGLT2is using real-world data. We however identified two previous observational cohort studies that examined pneumonia-related outcomes with SGLT2is versus DPP4is by using primary care data from the UK and electronic health records from Hong Kong [26, 27]. Both studies reported a lower risk of pneumonia associated with SGLT2is compared to DPP4is (Hong Kong: HR 0.71, 95% CI 0.62 to 0.81 [27]; United Kingdom: HR 0.48, 95% CI 0.28 to 0.82) [26], which was consistent with our findings (HR 0.60, 95% CI 0.55 to 0.64); subgroup analyses according to the etiology of pneumonia were also consistent (Additional file 1: Tables S23). Moreover, while indirect, our study (HR 59, 95% CI 0.47 to 0.75) was also consistent with a recent cohort study, which used primary care data from the United Kingdom, reporting a lower risk of severe exacerbations of COPD (HR 0.62, 95% CI 0.48 to 0.81) with SGLT2is versus sulfonylureas in patients with T2D and COPD [28]. While few data were available on the real-world, but partial, respiratory effects of SGLT2is, we came across no study, to our knowledge, that made a more comprehensive investigation into several respiratory events simultaneously in this patient population. In the meantime, based on the available evidence from randomized trials and observational studies, our findings provide novel and clinically meaningful real-world evidence by demonstrating that SGLT2is may have respiratory benefits by being associated with a lower risk of non-chronic respiratory events in patients with T2D, given that respiratory mortality, which includes pneumonia, remains a significant public health concern not only in South Korea, but also in Western populations [43].

Pulmonary function deficiency is an underrecognized issue in patients with T2D although these patients have lower lung diffusion capacity than those without T2D, and it has also been associated with an increased risk of hospitalization for pneumonia and subsequent mortality [44, 45]. Consequently, optimal glycemic control in this patient population is essential to improving clinical prognosis, including respiratory disease. While the exact biological mechanism for the favorable effects of SGLT2is on respiratory events remain unknown, several hypotheses exist. Preclinical studies showed that SGLT2is have direct vasodilatory effects on pulmonary circulation that may influence the subsequent risk of pulmonary disease [46, 47]; empagliflozin exhibited pulmonary protective effects after pulmonary ischemia or reperfusion injury in vivo.[48] In addition, SGLT2is facilitate glucose homeostasis of the ASL by reducing glucose movement into the ASL by transcellular pathways via an insulin-independent glucose-lowering mechanism [6, 49]; insulin treatment was previously reported to stimulate cellular glucose uptake [6]. This is important as low ASL glucose concentration helps prevent infections and exacerbation of pulmonary diseases [6]. Furthermore, SGLT2is, via their unique glucose-lowering mechanism that cause energy and salt loss [50], may achieve specific metabolic adaptations and activate aestivation-like hypometabolisms to improve organ’s cellular lifespan [50, 51] to possibly offer cardiorenal benefits [15, 18, 52] and potential hepatic benefits [53–56]. Another possibility that cannot be ruled out is the pleiotropic effects of SGLT2is (e.g., weight and blood pressure control) [57, 58], which could have contributed to lowering the risk of respiratory events in patients with T2D. Although further studies are needed to explore the exact mechanism underlying SGLT2is against respiratory diseases, SGLT2i may indeed have beneficial respiratory effects based on our findings and the existing hypotheses.

Strengths of this nationwide cohort study is the use of a large-scale real-world data with information on several potential confounders that were used to estimate PS to minimize confounding bias, further accompanied with a robust methodology (e.g., use of an active comparator to minimize confounding by indication bias) and consistent findings in multiple sensitivity analyses. We used DPP4is as an active comparator to SGLT2is because these two antidiabetic treatments are used at a similar stage of T2D treatment and share a common route of administration (e.g., oral regimen). Moreover, we were able to estimate both the class and molecule-specific associations of SGLT2is owing to the large sample size of this study (> 500,000 patients after PS matching).

Nevertheless, some potential limitations exist. First, owing to the inherent limitations of insurance claims data, we could not determine the exact reasons for treatment switches and discontinuations during follow-up (Additional file 1: Tables S24). Furthermore, our findings may be affected by exposure misclassification as we were not able to measure adherence to the medications prescribed. However, results of sensitivity analyses that varied the grace period to define treatment discontinuation in the as-treated analysis or applied the intention-to-treat definition showed consistent findings to that of the main results, implying that any bias arising from exposure misclassification was unlikely to have impacted our findings. Second, while outcome misclassification may have been possible, this is expected to occur non-differentially between the treatment groups and would bias the estimates toward the null hypothesis. To examine this issue, we redefined the study outcomes using only diagnosis codes recorded in the primary position and obtained results that were consistent with those the primary analysis. Third, because our data were claims-based, we did not have information on patients’ lifestyle (e.g., smoking) and detailed laboratory test results (e.g., glycated hemoglobin level). Thus, some residual confounding is possible. However, based on the E-value for our primary outcome, an unmeasured confounder would need to be associated by at least 2.7-fold with both the exposure and outcome to nullify our finding, which we believe is unlikely under reasonable assumptions. Fourth, the median (0.66 years) and mean (1.14 years) duration of follow-up was relatively short to assess the long-term disease-modifying effects of SGLT2is on adverse respiratory events in this study. However, in addition to our mean duration of follow-up being longer than prior studies of SGLT2is [59, 60], similar beneficial effects of SGLT2is were observed in the first year of life as well as in subsequent years in the subgroup analyses that stratified on the duration of follow-up (eTable 17). Future studies using more recent and accumulated real-world data are warranted to examine the long-term respiratory safety of SGLT2is. Fifth, as this study did not investigate into all types of respiratory disease, there are limitations in generalizing the study findings beyond the endpoints studied. Sixth, as an intrinsic limitation of observational studies, our findings only investigated an association of respiratory with SGLT2is versus DPP4is, not its causation. Finally, our study used data from HIRA in South Korea, and more studies using real-world data sources from other regions or countries are needed to confirm the generalizability of our findings, given that analyses examining individual endpoints of the primary composite endpoint, particularly ARDS, produced imprecise estimates.

The study was approved by the Institutional Review Board (IRB) of Sungkyunkwan University (SKKU 2021–06-005) and obtaining informed consent from the study population was waived by the IRB.

Not Applicable

Dr. Shin reports receipt of research funding from the Ministry of Food and Drug Safety, the Ministry of Health and Welfare, the National Research Foundation of South Korea, and from Daiichi Sankyo, GSK, and Pfizer, outside the submitted work. No other disclosures were reported.

The data that support the findings of this study are available from the Health Insurance Review and Assessment Service of South Korea but restrictions apply to the availability of these data due to domestic laws and regulations that prohibit the distribution or release of individual’s data to the public, and so are not publicly available. Data are however available from the authors upon reasonable request and with permission of the Health Insurance Review and Assessment Service of South Korea.