Comparative cardiovascular safety of LABA/LAMA FDC versus LABA/ICS FDC in patients with chronic obstructive pulmonary disease: a population-based cohort study with a target trial emulation framework

This present study used data from the Taiwan National Health Insurance Research Database (NHIRD), which included de-identified data of approximately 23 million beneficiaries enrolled in a single-payer national health insurance system [16, 17]. See Additional file 1: eMethods for detailed data source description.

To our knowledge, no existing trials aimed to examine cardiovascular safety issues of LABA/LAMA FDC versus LABA/ICS FDC among patients with COPD. Therefore, we specified components of a target trial (i.e., a hypothetical trial) and emulated the trial using Taiwan NHIRD. In another word, we conducted a cohort study with a target trial framework in which ascertainment of study population, exposure, outcomes, and baseline covariates were anchored at the cohort entry date (i.e., T₀) and appropriate statistical analyses were applied to enhance causal inference estimation using real-world data [18‐22]. See Additional file 1: Table S1 for specification of each component and Fig. 1 for the graphic depiction of the design. Corresponding approaches applied in the present study were also mentioned in our previous work [23, 24].

Our study population comprised patients with COPD who initiated LABA/LAMA FDC or LABA/ICS FDC from the NHIRD between 2017/01/01 and 2020/12/31. COPD was defined as having ≥ 1 outpatient or inpatient relevant diagnosis code in any diagnosis position (see Additional file 1: Table S2 for codes [25]). Use of LABA/LAMA FDC or LABA/ICS FDC was defined based on outpatient pharmacy dispensing claims (see Additional file 1: Table S3 for codes). The cohort entry date (i.e., T₀) was the date of the first dispensing of a LABA/LAMA FDC or LABA/ICS FDC after a diagnosis of COPD. Further exclusion criteria were presented in Additional file 1: eMethods.

We defined the primary outcome, a composite of cardiovascular events, as the first hospitalization for acute myocardial infarction, unstable angina, congestive heart failure, cardiac dysrhythmia, or ischemic stroke. We ascertained each outcome based on inpatient diagnosis codes recorded in the primary positions. These claims-based algorithms have been shown to have positive predictive values of 76–100% (Additional file 1: Table S4) [26‐35].

The primary analysis applied an on-treatment approach, which followed patients from cohort entry to outcome occurrence, study FDC treatment discontinuation or change, death, or end of data (2020/12/31), whichever came first. Study FDC treatment discontinuation was defined using a grace period of 60 days between the end date of one dispensing and the start date of the next dispensing, if any; and the discontinuation date was 60 days after the end of the last dispensing. Study FDC treatment change was defined as a dispensing of ICS-containing regimens or another LABA/LAMA FDC for LABA/LAMA FDC initiators, and a dispensing of LAMA-containing regimens or another LABA/ICS FDC for LABA/ICS FDC initiators.

Trial and real-world data have shown that use of LABA/LAMA has a lower risk of pneumonia compared to use of LABA/ICS (relative risk of 0.57 [95% CI, 0.42–0.79] and 0.66 [95% CI, 0.50–0.87], respectively) [36, 37]. Therefore, we chose pneumonia as a positive control outcome and examined if our design could identify a decreased risk of pneumonia associated with LABA/LAMA FDC. This positive control outcome approach is commonly applied in observational studies for assessing internal validity [38, 39]. The pneumonia outcome was determined using inpatient diagnosis codes recorded in any positions, with the claims-based algorithm exhibiting a positive predictive value of 88% (Additional file 1: Table S4) [40].

We assessed > 80 potential baseline confounders including age at cohort entry, sex, COPD duration defined as the duration from the first recorded date of COPD diagnosis (looking back until 2014/01/01) to the cohort entry date, calendar year of the cohort entry date, as well as comorbidities (e.g., cardiovascular disease), other medication use (e.g., cardiovascular medication), influenza or pneumococcal vaccination, and outpatient visits or hospital admissions for any reasons, for COPD, or for cardiovascular disease that may be associated with both the use of study FDC treatment and the risks of study outcomes. These characteristics were determined using diagnosis files or pharmacy dispensing records derived from outpatient or inpatient claims of the NHIRD within 365 days before cohort entry. Additional file 1: Tables S5 and S6 present detailed covariate information.

Using all above predefined, claims-based covariates, we estimated the baseline propensity scores (PS), i.e., the probabilities of initiating LABA/LAMA FDC versus LABA/ICS FDC, with a logistic regression model. Because there were many more LABA/LAMA FDC initiators, up to five LABA/LAMA FDC initiators were matched to one LABA/ICS FDC initiator using a nearest-neighbor algorithm without replacement and with a maximum matching caliper of 0.025 on the PS scale (see Additional file 1: eMethods for detailed variable-ratio matching steps) [41]. We computed standardized differences for each covariate to evaluate covariate distribution before and after PS matching, with a value < 0.1 in absolute value indicating balance between treatment groups [42].

The incidence rates and 95% CIs for LABA/LAMA FDC and LABA/ICS FDC initiators were estimated according to a Poisson distribution. The cumulative incidence plots were derived by one minus the Kaplan–Meier estimate (i.e., complement of the Kaplan–Meier survival function). Using conventional Cox proportional hazards models, we estimated HRs and 95% CIs of composite cardiovascular events comparing LABA/LAMA FDC with LABA/ICS FDC before and after PS matching. We also examined the risks for the individual components of the composite outcome. All the analyses after PS matching took the variable-ratio matching into consideration.

For composite cardiovascular events, we conducted four pre-specified sensitivity analyses to examine the robustness of the results comparing LABA/LAMA FDC with LABA/ICS. First, we used an intention-to-treat approach which continued to follow patients regardless of study FDC treatment discontinuation or change during follow-up and censored patients at the first of outcome occurrence, death, or end of data. Second, to account for the influence of competing risk from overall death, we re-draw cumulative incidence plots and applied the Fine‐Gray subdistribution hazard model for subdistribution HR estimation, which remains patients who have previously experienced overall death in the risk sets (Additional file 1: eMethods) [43, 44]. Third, to mitigate potential unmeasured confounding, we conducted high-dimensional PS (hd-PS) estimation which identifies and includes additional 100 empirically-identified, claims-based covariates in the PS model (Additional file 1: eMethods) [45, 46].

Finally, besides controlling for claims-based covariates, we attempted to additionally adjust for several important clinical parameters captured from the Taiwan National Health Insurance (NHI) Laboratory Database and the Taiwan COPD pay-for-performance (P4P) Database, both of which were recently established nationwide databases under series of national health policies in Taiwan (see Additional file 1: eMethods for detailed database description [47, 48]). Specifically, we measured the following 11 clinical parameters potentially related to the risks of outcomes from outpatient and inpatient health encounter records in the NHI Laboratory Database or the COPD P4P Database within 365 days before cohort entry, including laboratory test results (eosinophil, c-reactive protein, low-density lipoprotein-cholesterol, glycated hemoglobin, and glomerular filtration rate [GFR] or estimated GFR), lung function test results (predicted post-dose forced expiratory volume in one second [FEV₁], post-dose FEV₁/forced vital capacity), COPD Assessment Test (CAT) score, systolic blood pressure, and health behavior (body mass index and smoking status) (Additional file 1: Table S7). Because not all the patients had information on these parameters, we conducted multiple imputation to handle missing data issues (Additional file 1: eMethods) [49‐51].

For composite cardiovascular events, we also conducted three pre-specified subgroup analyses (1) to evaluate potential effect measure modification by patient characteristic (age, sex, COPD duration, history of hospitalized COPD exacerbations, and history of cardiovascular diseases); (2) to assess whether the association was different comparing individual LABA/LAMA FDC with individual LABA/ICS FDC; (3) and to explore potential treatment duration-response relation for LABA/LAMA FDC (1–90, 91–180, 181–365, and > 365 days). We re-estimated the PS and re-matched patients within each subgroup [41, 52, 53]. See Additional file 1: Table S8 for details of each analysis. Besides comparing the subgroup-specific effect estimates, we applied the Wald test to formally test if the results differ materially across patient subgroups, and a p-value < 0.05 was considered statistically significant.

A total of 99,506 eligible patients who initiated a LABA/LAMA FDC (n = 61,221) or a LABA/ICS FDC (n = 38,285) were included. Vilanterol/umeclidinium (VIL/UME, n = 26,606) was the most commonly used LABA/LAMA FDC, followed by olodaterol/tiotropium (OLO/TIO, n = 19,189) and indacaterol/glycopyrronium (IND/GLY, n = 15,426). Formoterol/budesonide (FOR/BUD, n = 10,351) and formoterol/beclomethasone (FOR/BEC, n = 10,311) were the two most frequently used LABA/ICS FDC, followed by salmeterol/fluticasone (SAL/FLU, n = 9,605), vilanterol/fluticasone (VIL/FLU, n = 7,584), and formoterol/fluticasone (FOR/FLU, n = 434) (Additional file 1: Fig. S1).

Before PS matching, LABA/LAMA FDC initiators were slightly older (mean age in years: 70 versus 69) and more likely to be male (male %: 85 versus 65) than LABA/ICS FDC initiators. LABA/LAMA FDC initiators also tended to have a history of lung cancer, receive LABA or LAMA monotherapy, and have more frequent outpatient visits for COPD than LABA/ICS FDC initiators. In contrast, LABA/LAMA FDC initiators were less likely to receive ICS than LABA/ICS FDC initiators. The PS model yielded a c-statistic of 0.790. After 5:1 variable-ratio matching, a total of 75,926 patients were included (76% of the study cohort, 80% of LABA/LAMA FDC initiators, and 71% of LABA/ICS FDC initiators). PS matching achieved balance in all baseline claims-based characteristics (Table 1; Additional file 1: Table S9, Fig. S2).

LABA/LAMA FDC initiators had a lower rate of pneumonia compared to LABA/ICS FDC initiators during follow-up, corresponding to an HR after PS matching of 0.65 (95% CI, 0.58–0.74) (Additional file 1: Table S10).

The intention-to-treat approach yielded mean follow-up days of 651 days for LABA/LAMA FDC initiators and 676 days for LABA/ICS FDC initiators, corresponding to an HR after PS matching of 0.97 (95% CI, 0.90-1.05) (Table 4; Additional file 1: Table S11). After accounting for competing risk from overall death, the cumulative incidence plots tended to be slightly flatter (Additional file 1: Fig. S4) compared to that of the main analysis. However, the Fine-Gray subdistribution HR after PS matching (0.90 [95% CI, 0.79–1.03], see Table 4) did not change materially when compared to what was generated by the traditional Cox model. The hd-PS estimation also yielded a similar finding (HRs after PS matching, 0.88 [95% CI, 0.77-1.00], see Table 4).

A total of 83,025 patients (83% of the eligible patients) had at least one clinical parameter captured from either the NHI Laboratory Database or the COPD P4P Database Information from the NHI Laboratory Database tended to be more comprehensive (approximately 82% of patients having at least one laboratory test result). Information from the COPD P4P Database; however, was less complete (approximately 8% of patients having data of lung function test, respiratory symptoms, blood pressure, or health behavior) (Additional file 1: Table S12).When we applied multiple imputation to incorporate these additional parameters in the PS-matched analysis, the risk estimate did not change apparently from those in the main analysis (HRs of 0.73 [95% CI, 0.31–1.73] when only including data from the NHI Laboratory Database and 0.89 [95% CI, 0.77–1.03] when simultaneously including data from the NHI Laboratory Database and the COPD P4P Database) (Table 4). Distribution of these parameters between FDC treatment groups became balanced as well (Additional file 1: Table S13).

Overall, the risk of composite cardiovascular events associated with LABA/LAMA FDC did not change materially by patient characteristic (Figure 2; Additional file 1: Table S14), individual LABA/LAMA FDC and LABA/ICS FDC (Additional file 1: Table S15), and treatment duration (Figure 2; Additional file 1: Table S16), although the risk seemed to be lower among some subgroups. The p-values for test of homogeneity across patient subgroups were > 0.05.

One meta-analysis of clinical trials (n = 18,170) reported an increased risk of MACE comparing LABA/LAMA to LABA/ICS (1.6% vs 1.3%; risk ratio, 1.42 [95% CI, 1.11–1.81]) [6]. The results were mainly driven by one dual combination therapy trial (LABA/LAMA, LABA/ICS) and two triple combination therapy trials (LABA/LAMA/ICS, LABA/LAMA, LABA/ICS) (n = 13,817; 2.0% vs 1.5%; risk ratio, 1.40 [95% CI, 1.08–1.82]) [9‐11]. The discrepant findings generated from these trial data and our study should be elaborated carefully in terms of differences in patient characteristics and methodological concerns across data source.

Specifically, given the inclusion and exclusion criteria, patients enrolled in these efficacy trials tended to have more severe COPD disease but had no clinically significant cardiovascular abnormalities. As demonstrated in Additional file 1: Table S17, patients in these trials had mean COPD duration of 7–8 years and mean predicted FEV₁ of 44–56%; all the patients had moderate or severe exacerbation history (> 20% having at least one severe exacerbation episode); the mean CAT score was 17–20; and nearly 56–81% of enrollees had baseline ICS use and some may have prior asthma history. In terms of methodological perspectives, the trial design of forced ICS withdrawal at randomization may therefore exacerbate disease control for those who have benefited from ICS treatment but were allocated to the LABA/LAMA group [12, 13], potentially leading to subsequent cardiovascular consequences. Moreover, the primary endpoint of the three trials was an annual rate of COPD exacerbations rather than cardiovascular outcomes. This may raise concerns about adjudication and potential misclassification of cardiovascular events in the original trials and in the meta-analysis [54]. This may also result in lack of detailed data on time to cardiovascular events for HR estimation.

The present Taiwanese cohort study included patients with COPD who initiated LABA/LAMA FDC or LABA/ICS FDC in the daily practice, who tended to have mild COPD disease but may have major underlying cardiovascular disease compared to enrollees in aforementioned trials. For example, our study patients had mean COPD duration of 2 years (measured in the database) and mean predicted FEV₁ of 63% (mainly based on imputed data); 14% of patients had at least one severe exacerbation episode; the mean CAT score was 13 (mainly based on imputed data); and only 10% of patients ever used ICS at baseline (Additional file 1: Table S17). To mitigate misclassification of cardiovascular outcomes, we ascertained cardiovascular events according to validated algorithms. Our on-treatment and intention-to-treat follow-up approaches yielded similar results, which did not show a higher cardiovascular risk associated with LABA/LAMA FDC use. Our findings highlight that it deserves more caution when directly applying trial data for safety evaluation in the real-world settings.

Few real-world studies have evaluated cardiovascular risk of LABA/LAMA versus LABA/ICS. One cohort study using the US Truven Health MarketScan Commercial and Medicare Database (2004–2012, n = 19,078) did not observe an increased risk of hospitalized, composite cardiovascular events (HR, 0.85; 95% CI, 0.66–1.04) [7]. Use of LABA/LAMA or LABA/ICS in this study; however, included FDC and free-combination forms. Another cohort study conducted in the Taiwan NHIRD (2015–2016, n = 28,237) further showed that there was no excess risk of hospitalized, composite cardiovascular events between individual LABA/LAMA FDC and LABA/ICS FDC. The HRs ranged from 1.03 (95% CI, 0.83–1.29) to 1.29 (95% CI, 0.96–1.73) comparing two LABA/LAMA FDC (IND/GLY, VIL/UME) to three LABA/ICS FDC (FOR/BEC, FOR/BUD, SAL/FLU) [8].

Using the Taiwan NHIRD with more updated data (2017–2020, n = 75,926), the present cohort study yielded a larger sample size and included more FDC products (three LABA/LAMA FDC [IND/GLY, VIL/UME, OLO/TIO] and five LABA/ICS FDC [FOR/BEC, FOR/BUD, FOR/FLU, SAL/FLU, VIL/FLU]). Extending findings from prior cohort studies, we comprehensively presented the risk of various cardiovascular outcomes and the risk across patients with different characteristics (such as patients with hospitalized COPD exacerbations or with cardiovascular disease at baseline), with different comparisons of LABA/LAMA FDC versus LABA/ICS FDC, and with longer treatment durations (such as > 365 days). The results provide more informative messages and facilitate physicians for treatment decision in different clinical scenarios.

Our study has some notable strengths. First, to our knowledge, our study is the largest study which provides real-world evidence on cardiovascular safety of LABA/LAMA FDC versus LABA/ICS FDC. Second, under a target trial emulation framework, we defined study population, treatment initiation, outcome occurrence, and baseline covariates anchored at the cohort entry date (i.e., T₀). This prevented a common shortcoming of “looking forward the future” and accompanying bias in the real-world studies [18‐22]. Third, in alignment with the emulation framework, our main analysis extracted diagnosis and medication information from the Taiwan NHIRD derived from a single-payer health insurance system. We used validated claim-based algorithms to determine cardiovascular and pneumonia outcomes. We also replicated the known association between LABA/ICS FDC and pneumonia. To mitigate potential confounding by patient characteristic, we applied the state-of-the-art PS matching analysis accounting for > 80 pre-specified claims-based covariates that may be associated with the exposures and the outcomes. Collectively, these lend support to our study validity. Fourth, we further improved confounding control using the sophisticated hd-PS algorithm with additional empirical claims-based covariates and the multiple imputation approach with additional important clinical parameters. Fifth, the findings of several subgroup analyses enhance the generalizability and the application of our study.

We need to recognize some limitations. First, our study yielded a mean follow-up duration of 215 days (based on an on-treatment approach), which tended to be shorter than that in previous clinical trials (52-week follow-up duration based on an intention-to-treat approach) [9‐11]. However, this reflected actual use patterns of LABA/LAMA FDC in the real-world environment and was consistent with previous observational studies [7, 8]. Our sensitivity analysis based on an intention-to-treat approach (a mean follow-up duration of 660 days, Table 4 and Additional file 1: Table S11) and subgroup analysis with longer treatment duration (such as >365 days, Figure 2 and Additional file 1: Table S16) did not change the results materially. Second, we examine a broad spectrum of cardiovascular events to enhance clinical relevance. However, the current database did not provide information on cardiovascular death for further risk estimation. Third, while we controlled for a large number of potential confounders with several approaches, we could not fully exclude the possibility of residual or unmeasured confounding, which was an inherent limitation of most observational studies. Similarly, we attempted to integrate claims-based covariates and clinical parameters to strengthen confounding adjustment. For example, for adjusting for the influence of COPD severity, we simultaneously captured information on COPD duration, history of hospitalized COPD exacerbations, and lung function test results. However, proportions of patients with clinical parameters from COPD P4P Database tended to be low. This may be because the database was only available since 2017. The availability with corresponding information is expected to increase over time and future studies with accrual data would facilitate replication of our results.

This population-based cohort study with a target trial emulation framework provides important reassurance about comparative cardiovascular safety of LABA/LAMA FDC compared with LABA/ICS FDC among patients with COPD.