Glucagon-like Peptide 1 Receptor Agonists and Asthma Exacerbations: Which Patients Benefit Most?

Abstract

Rationale

Although recent evidence suggested that glucagon-like peptide 1 receptor agonists (GLP1RAs) might reduce the risk of asthma exacerbations, it remains unclear which subpopulations might derive the most benefit from GLP1RA treatment.

Objectives

To identify characteristics of patients with asthma that predict who might benefit the most from GLP1RA treatment using real-world data.

Methods

We implemented an active-comparator, new-user design analysis using commercially ensured patients 18–65 years of age from MarketScan data for 2007–2019 and identified two cohorts: GLP1RAs versus thiazolidinediones and GLP1RAs versus sulfonylureas. The outcome was acute exacerbation of asthma (hospital admission or emergency department visit for asthma) within 180 days after initiation. We applied iterative causal forest, a novel causal machine learning subgrouping algorithm, to assess heterogeneous treatment effects. In identified subgroups, we predicted propensity score, conducted propensity score trimming, and then estimated adjusted risk differences for the effect of GLP1RAs relative to comparators on asthma exacerbation using inverse probability treatment weighting in the propensity score–trimmed subpopulation.

Results

Among 10,989 patients initiating GLP1RAs or thiazolidinediones and 17,088 patients initiating GLP1RAs versus sulfonylurea, GLP1RA initiators had fewer exacerbations, with adjusted risk differences of −0.5% (95% confidence interval [CI], −1.1% to 0.1%) and −1.6% (95% CI, −2.2% to −1.1%), respectively. In the GLP1RA versus sulfonylurea cohort, in which we observed a beneficial effect, our iterative causal forest analysis identified five subgroups with different treatment effects, defined by the number of emergency department visits, the number of prescriptions for short-acting β₂-agonists, the number of prescriptions for inhaled steroids and long-acting β-agonists (either combination therapy or concurrent use), and age ≥ 50 years. Among these, patients with two or more emergency department visits during the 12-month baseline period had the largest absolute exacerbation risk reduction, with a decrease of 2.8% for GLP1RAs (95% CI, −4.8% to −0.9%).

Conclusions

GLP1RAs demonstrated a beneficial effect on reducing asthma exacerbation relative to sulfonylureas. Patients with asthma with two or more emergency department visits (a proxy for disease severity) benefit most from GLP1RAs. Emergency department visit frequency, the number of maintenance and reliever inhalers, and age might help individualize prediction of the short-term benefit of GLP1RAs on asthma exacerbation.

Diabetes is associated with microvascular complications that can damage the kidneys, eyes, and peripheral nerves (1). The lung is also a target organ for diabetic microangiopathy, and decreased lung function has been reported in patients with diabetes (2, 3). Patients with diabetes are at increased risk of several pulmonary diseases (asthma, chronic obstructive pulmonary disease [COPD], fibrosis, and pneumonia), but it is unclear how much of this is due to common risk factors versus a direct effect (4). Among patients with diabetes, approximately 5% (∼2 million in the United States) have asthma (5).

Glucagon-like peptide 1 receptor agonists (GLP1RAs), such as dulaglutide, liraglutide, and semaglutide, are approved by various health authorities for the treatment of type 2 diabetes (T2D) and to reduce the risk of major adverse cardiovascular events in adults with T2D (6). GLP1RAs increase insulin secretion by activating the GLP1 receptor on pancreatic β cells. GLP1 receptors are also found in multiple organs, including the stomach, heart, lung, and kidney (7). Highly expressed in lung epithelial and endothelial cells (8), the GLP1 receptor might play a role in pulmonary disease (9). Ex vivo and preclinical studies show that GLP1RAs inhibit airway inflammation (10–12), reduce airway eosinophilia and mucus production, and attenuate bronchial hyperresponsiveness (8, 11, 13). Furthermore, a recent study showed that liraglutide inhibits the aeroallergen-induced activation of lung group 2 innate lymphoid cells and neutrophilic airway inflammation in obese mice (14). As obesity is an important comorbidity in a subset of patients with asthma (15, 16), and the U.S. Food and Drug Administration has approved GLP1RAs for obesity treatment, liraglutide and semaglutide have been suggested to have a potential role as pharmacotherapeutic agents for obese persons with asthma (14).

An electronic health record–based cohort study by Foer and colleagues showed that GLP1RA initiators had fewer asthma exacerbations during the 6-month study period compared with initiators of other classes of glucose-lowering drugs, including sodium-glucose cotransporter-2 inhibitors, dipeptidyl peptidase-4 inhibitors, sulfonylureas (SUs), and basal insulin (17). Specifically, the incidence rate ratio was 0.55 (95% confidence interval, 0.36–0.83) when comparing 448 GLP1RA initiators with 2,253 SU initiators. A cohort study by Albogami and colleagues using employer-sponsored health insurance data involving 4,150 GLP1RA initiators suggested fewer asthma exacerbations during a 12-month follow-up, with an adjusted hazard ratio of 0.54 (95% confidence interval, 0.23–1.27), compared with 12,540 dipeptidyl peptidase-4 inhibitor initiators (18). Notably, these studies focused on the average treatment effects of GLP1RAs in the population. However, treatment effects might vary across subpopulations with T2D because of heterogeneity among individuals (19), and little is known about how these effects might vary according to clinical characteristics that can be identified in real-world data. Identifying these can lead to improved patient responses and target randomized studies.

Recent developments in machine learning for causal inference, especially causal forest analyses, allow us to better estimate heterogeneous treatment effects (HTEs) (20–22). We recently developed a causal forest–based subgrouping algorithm, iterative causal forest (iCF), which pinpoints important subgroups with HTEs without requiring prior knowledge of treatment effect modifiers and often outperforms other subgrouping methods (23, 24). Leveraging large databases with algorithms such as iCF presents a pivotal opportunity to uncover previously unknown heterogeneity and suggest personalized therapies. The aim of this study was to assess GLP1RA’s potential HTEs on asthma exacerbations by implementing iCF to identify patients with diabetes and asthma who benefit most from treatment with GLP1RA.

Methods

Data Source

We used MarketScan Research Databases, among the largest and longest running proprietary U.S. claims databases used for healthcare research, from January 2007 to December 2019. The MarketScan databases contain longitudinal, individual-level data, including demographics, inpatient, outpatient, and pharmacy claims, and encounters, primarily for adults younger than 65 years from approximately 350 employers across the United States. The study protocol was approved by the University of North Carolina Institutional Review Board (#21-2340).

Study Population

The eligible population consisted of MarketScan enrollees 18–65 years of age. Previous studies showed that thiazolidinediones, a class of medications for treating patients with T2D that improve insulin resistance, might improve asthma control compared with nonthiazolidinedione glucose-lowering drugs (25). A randomized trial demonstrated that rosiglitazone, a thiazolidinedione, improved lung function relative to inhaled beclomethasone in smokers with asthma (26). Insulin resistance is closely linked to the risk of asthma in adults, and insulin potentially activates airway immune and structural cells, leading to inflammation and narrowing (27). Therefore, we chose thiazolidinediones and SUs (which increase insulin secretion) as comparators to increase the clinical relevance by addressing the question “Which treatment is more effective?” We first independently identified all new-use periods from January 1, 2008, to December 30, 2019, of each the three therapies of interest—GLP1RA, thiazolidinedione, and SU—simultaneously (i.e., the three groups of new users are not mutually exclusive) on the basis of the first dispensing of a prescription in a given drug class after a 12-month washout period when the drug of interest was not prescribed (28). Then, we constructed two comparison cohorts from the three groups of new users—GLP1RAs versus thiazolidinediones and GLP1RAs versus SUs—also requiring no evidence of use of the comparator drug in the 12 months before initiation (i.e., baseline period as shown in Figure E1 in the data supplement). With this selection, the number GLP1RA initiators differed in the two comparisons. In addition, we required patients to have a second prescription dispensing claim within the same drug class and the sum of the first prescription’s days’ supply plus a 90-day grace period to increase the probability that patients actually took the medication. The use of active comparators helps reduce bias by selecting comparator cohorts of similar disease severity and with an indication for initiating a second-line glucose-lowering drug, and the new-user design ensures appropriate temporal ordering of baseline confounders, treatment, and outcome (28, 29). In the baseline period, we required patients to have at least one inpatient or two outpatient encounters with asthma, defined on the basis of previously used diagnostic codes (International Classification of Diseases, Ninth Revision, Clinical Modification codes and International Classification of Diseases, Tenth Revision, Clinical Modification codes) or medication dispensing (17), and we excluded patients who had conditions necessitating the use of systemic steroids (to increase the probability that observed steroid prescribing was associated with asthma severity/exacerbations) or with chronic congestive heart failure, vocal cord dysfunction, COPD, and other respiratory diseases (see Tables E1–E3) (17, 18).

Outcome

The outcome was asthma exacerbation, defined on the basis of a previously used algorithm (17, 18, 30) as the first hospital admission with a primary diagnosis of asthma or with a primary diagnosis of respiratory symptoms (dyspnea, shortness of breath, wheezing, and cough) and a secondary diagnosis of asthma or an emergency department visit for asthma (17, 18) (diagnostic codes are listed in Table E1).

Risk Period

We performed intention-to-treat analyses, in which we ignored treatment changes. Patients were followed starting from the index date (second prescription date) until the end of MarketScan enrollment. Patients were followed for a maximum of 0.5 years after the first prescription, which was informed by adherence patterns of GLP1RAs in real-world clinical practice (31, 32). We required patients to enter the cohort no later than July 1, 2019, to increase the probability of 0.5-year follow-up (see Figure E1).

Covariates

We controlled for a variety of covariates defined on the basis of claims (by International Classification of Diseases, Ninth Revision, Clinical Modification/International Classification of Diseases, Tenth Revision, Clinical Modification diagnosis and procedure codes, Healthcare Common Procedure Coding System codes, Current Procedural Terminology codes, and Anatomical Therapeutic Chemical codes) during the 12 months before the index prescription, including demographics, comorbidities, comedications, and healthcare use. For each comparison, we used these covariates to estimate propensity score (PS; the predicted probability of receiving the index treatment conditional on baseline covariates [33]) for each patient in each comparison to control for factors that might influence the decision to prescribe a given treatment.

Statistical Analysis

We implemented the iCF approach (23, 24) on the original cohort to classify patients into subgroups (Figures 1 and E2). At a high level, the iCF algorithm has the following steps: 1) it predicts both the outcome and the PS, which are then used to formulate a raw causal forest, followed by a homogeneity test and the selection of important variables (with a variable importance [VI] value greater than the mean VI value of all variables, where VI is a summary weighted sum of how many times a feature was split on at each depth in the raw causal forest); 2) varying depths of the causal forest are grown iteratively, using the chosen variables; and 3) from each forest, a decision tree that best represents a treatment heterogeneity signal is selected. Thereafter, a voting system determines the most consistent tree structure, considering it as a subgroup decision (defined by the set of leaves in a tree) at every respective depth; 4) models are then constructed with the purpose of predicting the treatment effect on the basis of subgroup decisions; and 5) the algorithm identifies the cross-validated subgroup decision that offers the most accurate prediction of the treatment effect, which is crucial in pinpointing both important and stable subgroups. We ran iCF for fivefold cross-validation and grew each causal forest with 1,000 trees and 100 iterations to obtain subgroup decision.

Figure 1.

Iterative causal forest (iCF). If the heterogeneity test from the raw CF is significant at P ≤ α, a family of shallow CFs are grown in the training set by tuned minimum leaf (subgroup) size to grow trees with varying depths at depth 2 (D2), D3, D4, and D5, respectively. Then, a set of subgroup decisions are derived from the voted best trees that best represent treatment heterogeneity signal in each CF at different depths. Next, four models are constructed to predict treatment effect, including interaction terms of treatment W and subgroup decisions G_D. Last, the most precise model (with the smallest error) is selected as the best model among the four models. The G_D in the best model is chosen as the final subgroup decision, G_iCF, from the iCF algorithm. The whole process is cross-validated. Further details are available in Figure E2. CF = causal forest.

Next, in each subgroup identified by iCF of the study sample, we quantified the conditional average treatment effect (per-subgroup treatment effect) of the new-user cohort by 0.5-year adjusted risk difference (aRD) versus comparators, which provided a clinically meaningful measure of treatment effect (34) and was also the scale assessed by the causal forest. We estimated PSs conditional on baseline covariates using logistic regression. To improve confounding control, we then applied asymmetric PS trimming to the subpopulation to avoid treatment comparisons with patients who were treated most contrary to prediction (i.e., in the “tails” of the PS distribution [35]) and excluded all patients with estimated PSs below the 0.1st percentile of the GLP1RA or above the 99.9th percentile of the comparator group. Subsequent analyses were performed on this subset of the population. We controlled measured confounding by inverse probability treatment weighting and balance was assessed using absolute standardized mean differences (ASMDs) (29).

We assessed the average treatment effect on the original cohort by aRD for overall population, similar to our approach in each subgroup but without PS trimming. In addition, we conducted bias analyses to evaluate how much unmeasured confounding would be necessary to nullify the observed 0.5-year risk difference (RD) considering two factors (36–38): the confounder’s estimated RD (confounder = 1 vs. confounder = 0) for the outcome in the comparator cohort and its observed prevalence difference (PD) across the treatment cohorts. An unmeasured confounder could nullify the observed beneficial effect on asthma exacerbation (negative RD) if it exhibits both a beneficial effect on the outcome (negative RD) and a higher prevalence in the GLP1RA group (positive PD) or both a harmful effect (positive RD) and a lower prevalence in the GLP1RA group (negative PD) (i.e., a negative product of RD and PD). Therefore, we assessed the top three most important confounders with the most negative product of RD in the comparator (untreated) cohort and PD between the treated and comparator groups, together with known important confounders including emergency department visit for asthma (39) and obesity (15). All computations for iCF subgroup identification were conducted in R version 4.1 (R Foundation for Statistical Computing). R codes for the iCF algorithm are available at https://github.com/tianshengwang/iCF. Data management and estimation of conditional average treatment effects were performed in SAS version 9.4 (SAS Institute).

Results

Average Treatment Effect

We identified 10,989 initiators for the GLP1RA versus thiazolidinedione comparison and 17,088 initiators for GLP1RA versus SU comparison (see Figure E3). The crude distribution of important variables and a full list of baseline characteristics are shown in Table 1 and Table E4, respectively. Approximately 41–45% of patients were 50–60 years of age, with 25.8–39.9% being men. Patients were more likely have been treated with GLP1RAs compared with thiazolidinediones and SUs since 2013 and 2016, respectively. Overall, the prevalence of baseline characteristics were similar in the comparison cohorts, except that GLP1RA initiators were more likely to 1) be slightly younger and have no hospital admissions; 2) have diabetic neuropathy, diabetic circulatory and other complications, and obstructive sleep apnea; and 3) take sodium-glucose cotransporter-2 inhibitors and undergo glycated hemoglobin and lipid tests. The well-balanced distribution of the covariates in the PS-weighted GLP1RA versus comparator cohorts (with the weighted median ASMDs being 0.028 and 0.038, respectively) indicates our success in eliminating measured confounding (see Table E4). Among the two comparison cohorts (see Figure E4; Table 2), the number of exacerbation events ranged from 86 to 408 in 0.5 years, and exacerbation risk ranged from 1.8% to 3.7%. The 0.5-year aRD was −0.5% (95% confidence interval, −1.1% to 0.1%) compared with thiazolidinediones and −1.6% (95% confidence interval, −2.2% to −1.1%) compared with SUs.

Table 1.

Distributions of selected important variables for comparisons of glucagon-like peptide 1 receptor antagonists versus thiazolidinediones and versus sulfonylureas before propensity score weighting in patients with baseline asthma and without chronic obstructive pulmonary disease

	GLP1RAs vs. TZDs			GLP1Ras vs. SUs
Characteristic	Overall Population			Overall Population			Subgroup 1: ≥2 Emergency Department Visits
	GLP1RAs (n = 7,939)	TZDs (n = 3,050)	ASMD*	GLP1RAs (n = 6,084)	SUs (n = 11,004)	ASMD†	GLP1RAs (n = 1,009)	SUs (n = 2,141)	ASMD‡
Demographic characteristic
Age group
18 ≤ age ≤ 30	214 (2.7)	61 (2.0)	0.046	181 (3.0)	484 (4.4)	0.076	53 (5.3)	164 (7.7)	0.098
30 < age ≤ 40	932 (11.7)	297 (9.7)	0.065	743 (12.2)	1,557 (14.1)	0.057	186 (18.4)	416 (19.4)	0.025
40 < age ≤ 50	2,241 (28.2)	767 (25.1)	0.070	1,759 (28.9)	2,728 (24.8)	0.093	298 (29.5)	570 (26.6)	0.065
50 < age ≤ 60	3,359 (42.3)	1,393 (45.7)	0.068	2,524 (41.5)	4,471 (40.6)	0.017	366 (36.3)	756 (35.3)	0.02
60 < age ≤ 65	1,193 (15.0)	532 (17.4)	0.066	877 (14.4)	1,764 (16.0)	0.045	106 (10.5)	235 (11.0)	0.015
Sex, male§	2,096 (26.4)	1,216 (39.9)	0.289	1,568 (25.8)	3,749 (34.1)	0.182	223 (22.1)	571 (26.7)	0.107
Cardiovascular disorders
Arrhythmia disorders	453 (5.7)	152 (5.0)	0.032	341 (5.6)	612 (5.6)	0.002	121 (12.0)	247 (11.5)	0.014
Obesity§	2,484 (31.3)	278 (9.1)	0.575	1,916 (31.5)	1,322 (12.0)	0.486	420 (41.6)	346 (16.2)	0.585
Respiratory disorders
Bronchitis	848 (10.7)	352 (11.5)	0.027	634 (10.4)	1,303 (11.8)	0.045	166 (16.5)	402 (18.8)	0.061
Smoking and smoking cessation	477 (6.0)	202 (6.6)	0.025	352 (5.8)	807 (7.3)	0.063	107 (10.6)	326 (15.2)	0.138
Other comorbidity
Electrolyte disorderǁ	610 (7.7)	178 (5.8)	0.074	450 (7.4)	858 (7.8)	0.015	189 (18.7)	409 (19.1)	0.009
Iron deficiency anemia	965 (12.2)	274 (9.0)	0.103	752 (12.4)	1,051 (9.6)	0.090	187 (18.5)	292 (13.6)	0.134
Depressionǁ	1,457 (18.4)	370 (12.1)	0.174	1,147 (18.9)	1,558 (14.2)	0.127	299 (29.6)	455 (21.3)	0.193
Obstructive sleep apnea	2,156 (27.2)	483 (15.8)	0.278	1,648 (27.1)	1,916 (17.4)	0.234	325 (32.2)	445 (20.8)	0.261
Psychosis¶	451 (5.7)	169 (5.5)	0.006	354 (5.8)	721 (6.6)	0.030	94 (9.3)	231 (10.8)	0.049
Metabolic disorders¶	5,597 (70.5)	2,013 (66.0)	0.097	4,232 (69.6)	6,644 (60.4)	0.193	722 (71.6)	1,339 (62.5)	0.193
Hypothyroidismǁ	1,553 (19.6)	372 (12.2)	0.203	1,248 (20.5)	1,573 (14.3)	0.165	207 (20.5)	329 (15.4)	0.134
Mild liver disorders¶	737 (9.3)	225 (7.4)	0.069	519 (8.5)	870 (7.9)	0.023	149 (14.8)	297 (13.9)	0.026
Medications for asthma or COPD
Number of prescriptions for ICS + LABA products**
0	4,830 (60.8)	1,810 (59.3)	0.031	3,631 (59.7)	7,037 (63.9)	0.088	646 (64.0)	1,443 (67.4)	0.071
1–5	1,215 (15.3)	459 (15.0)	0.007	927 (15.2)	1,595 (14.5)	0.021	148 (14.7)	340 (15.9)	0.034
6–10	1,009 (12.7)	414 (13.6)	0.026	813 (13.4)	1,174 (10.7)	0.083	113 (11.2)	186 (8.7)	0.084
≥11	885 (11.1)	367 (12.0)	0.028	713 (11.7)	1,198 (10.9)	0.026	102 (10.1)	172 (8.0)	0.072
ICSǁ
0	6,767 (85.2)	2,572 (84.3)	0.025	5,184 (85.2)	9,240 (84.0)	0.034	875 (86.7)	1,826 (85.3)	0.041
1–5	733 (9.2)	311 (10.2)	0.033	551 (9.1)	1,172 (10.7)	0.054	93 (9.2)	212 (9.9)	0.023
≥6	439 (5.5)	167 (5.5)	0.002	349 (5.7)	592 (5.4)	0.016	41 (4.1)	103 (4.8)	0.036
Ipratropium	597 (7.5)	242 (7.9)	0.016	461 (7.6)	864 (7.9)	0.010	117 (11.6)	233 (10.9)	0.023
Number of prescriptions for SABAs
0	2,257 (28.4)	910 (29.8)	0.031	1,731 (28.5)	3,088 (28.1)	0.009	295 (29.2)	642 (30.0)	0.016
1–5	2,606 (32.8)	1,069 (35.0)	0.047	2,017 (33.2)	3,682 (33.5)	0.007	281 (27.8)	618 (28.9)	0.023
6–10	1,673 (21.1)	591 (19.4)	0.042	1,267 (20.8)	2,309 (21.0)	0.004	230 (22.8)	456 (21.3)	0.036
≥11	1,403 (17.7)	480 (15.7)	0.052	1,069 (17.6)	1,925 (17.5)	0.002	203 (20.1)	425 (19.9)	0.007
Number of prescriptions for albuterol–ipratropium	482 (6.1)	180 (5.9)	0.007	382 (6.3)	688 (6.3)	0.001	91 (9.0)	176 (8.2)	0.028
Systemic steroid
0	6,152 (77.5)	2,473 (81.1)	0.089	4,689 (77.1)	8,820 (80.2)	0.075	740 (73.3)	1,612 (75.3)	0.045
1 or 2	820 (10.3)	274 (9.0)	0.046	645 (10.6)	1,013 (9.2)	0.047	115 (11.4)	232 (10.8)	0.018
≥3	967 (12.2)	303 (9.9)	0.072	750 (12.3)	1,171 (10.6)	0.053	154 (15.3)	297 (13.9)	0.039
History of medication use Insulin	2,362 (29.8)	421 (13.8)	0.394	1,728 (28.4)	994 (9.0)	0.513	371 (36.8)	294 (13.7)	0.55
SUs¶	2,082 (26.2)	1,141 (37.4)	0.242	NA	NA	NA	NA	NA	NA
Calcium channel blockers¶	2,095 (26.4)	684 (22.4)	0.092	1,561 (25.7)	2,941 (26.7)	0.024	236 (23.4)	544 (25.4)	0.047
Loop diuretics¶	2,966 (37.4)	1,059 (34.7)	0.055	2,250 (37.0)	3,631 (33.0)	0.084	158 (15.7)	258 (12.1)	0.105
Aspirin	2,966 (37.4)	1,059 (34.7)	0.055	2,250 (37.0)	3,631 (33.0)	0.084	104 (10.3)	163 (7.6)	0.094
Oral contraceptivesǁ	780 (9.8)	266 (8.7)	0.038	652 (10.7)	961 (8.7)	0.067	150 (14.9)	237 (11.1)	0.113
Estrogen	681 (9.6)	202 (8.7)	0.03	546 (10.3)	838 (8.8)	0.05	113 (11.2)	209 (9.8)	0.047
Measures of healthcare use
A1C tests¶	7,136 (89.9)	2,559 (83.9)	0.178	5,383 (88.5)	8,862 (80.5)	0.221	902 (89.4)	1,606 (75.0)	0.383
Flu shots¶	3,071 (38.7)	1,008 (33.0)	0.118	2,324 (38.2)	3,861 (35.1)	0.065	366 (36.3)	718 (33.5)	0.057
Number of hospital admission
0	6,982 (87.9)	2,634 (86.4)	0.047	5,391 (88.6)	9,262 (84.2)	0.130	673 (66.7)	1,320 (61.7)	0.105
1	793 (10.0)	342 (11.2)	0.040	573 (9.4)	1,407 (12.8)	0.107	250 (24.8)	586 (27.4)	0.059
≥2	164 (2.1)	74 (2.4)	0.024	120 (2.0)	335 (3.0)	0.069	86 (8.5)	235 (11.0)	0.083
Days of hospitalization¶
0	6,982 (87.9)	2,634 (86.4)	0.047	5,391 (88.6)	9,262 (84.2)	0.130	673 (66.7)	1,320 (61.7)	0.105
1 or 2	166 (2.1)	76 (2.5)	0.027	126 (2.1)	273 (2.5)	0.027	45 (4.5)	142 (6.6)	0.095
3 or 4	390 (4.9)	153 (5.0)	0.005	271 (4.5)	654 (5.9)	0.067	127 (12.6)	268 (12.5)	0.002
≥5	401 (5.1)	187 (6.1)	0.047	296 (4.9)	815 (7.4)	0.106	164 (16.3)	411 (19.2)	0.077
Number of emergency department visits
0	4,998 (63.0)	1,986 (65.1)	0.045	3,884 (63.8)	6,613 (60.1)	0.077	NA	NA	NA
1	1,564 (19.7)	596 (19.5)	0.004	1,191 (19.6)	2,250 (20.4)	0.022	NA	NA	NA
2	692 (8.7)	230 (7.5)	0.043	533 (8.8)	1,005 (9.1)	0.013	533 (52.8)	1,005 (46.9)	0.118
≥3	685 (8.6)	238 (7.8)	0.030	476 (7.8)	1,136 (10.3)	0.087	478 (47.1)	1,136 (53.1)	NA
Number of emergency department visits for asthma
0	7,342 (92.5)	2,768 (90.8)	0.062	5,652 (92.9)	9,629 (87.5)	0.182	739 (73.2)	1,251 (58.4)	0.316
1	447 (5.6)	219 (7.2)	0.063	322 (5.3)	1,027 (9.3)	0.156	160 (15.9)	542 (25.3)	0.236
2	110 (1.4)	44 (1.4)	0.005	82 (1.3)	233 (2.1)	0.059	82 (8.1)	233 (10.9)	0.094
≥3	40 (0.5)	19 (0.6)	0.016	28 (0.5)	115 (1.0)	0.068	28 (2.8)	115 (5.4)	0.132

Definition of abbreviations: A1C = glycated hemoglobin; ASMD = absolute standardized mean difference; COPD = chronic obstructive pulmonary disease; GLP1RA = glucagon-like peptide 1 receptor agonist; ICS = inhaled cortico steroid; LABA = long-acting β₂-agonist; NA = not applicable; SABA = short-acting β₂-agonist; SU = sulfonylurea; TZD = thiazolidinedione.

Data are expressed as n (%). All covariates (except for the number of days between first and second prescriptions) are measured 12 months before the first prescription date. A full list of covariates is provided in Table E4.

^*For all covariates after propensity score weighting: range, 0–0.104; median, 0.028; mean, 0.034.

^† For all covariates after propensity score weighting: range, 0.002–0.129; median, 0.038; mean, 0.044.

^‡ For all covariates after propensity score weighting: range, 0–0.128; median, 0.017; mean, 0.027.

^§ Sex and obesity are not selected as important variables.

^ǁ Selected important variables by raw causal forest for the GLP1RA versus SU cohort but not in the GLP1RA versus TZD cohort.

^¶ Selected important variables (with a variable importance value greater than the mean value of all variables) by raw causal forest for the GLP1RA versus TZD cohort but not in the GLP1RA versus SU cohort.

^**Including ICS/LABA combination product or using both ICS and LABA simultaneously.

Table 2.

Crude and adjusted risk differences for asthma exacerbation associated with the use of glucagon-like peptide 1 receptor antagonists in the population by initial treatment analysis in a maximum of 6-month follow-up by subgroup decisions from iterative causal forest algorithm

Comparison	Subpopulation Identified by iCF Algorithm	Cohort	n	Patient-Year	Number of Events	Risk (%)	Censor (%)	Crude Risk Difference (%)	IPTW Risk Difference (%)
GLP1RAs vs. thiazolidinediones	Overall population	GLP1RAs	7,939	3,632	151	1.9	14.6	−0.9 (−1.6 to −0.3)	−0.5 (−1.1 to 0.1)
		Thiazolidinediones	3,050	1,388	86	2.8	14.8
GLP1RAs vs. SUs	Overall population	GLP1RAs	6,084	2,797	107	1.8	13.7	−1.9 (−2.4 to −1.5)	−1.6 (−2.2 to −1.1)
		SUs	11,004	4,936	408	3.7	15.9
	Subgroup 1: ≥2 emergency department visits	GLP1RAs	888	395	35	3.9	16.7	−4.5 (−6.2 to −2.7)	−2.8 (−4.8 to −0.9)
		SUs	1,904	817	160	8.4	18.0
	Subgroup 2: <2 emergency department visits, >10 SABA prescriptions	GLP1RAs	787	361	19	2.4	12.8	−0.9 (−2.3 to 0.6)	−0.7 (−2.2 to 0.8)
		SUs	1,309	585	43	3.3	15.7
	Subgroup 3: <2 emergency department visits, ≤10 SABA, ICS + LABA prescriptions	GLP1RAs	1,576	736	17	1.1	12.8	−1.1 (−1.9 to −0.4)	−1.4 (−2.3 to −0.5)
		SUs	2,478	1,128	55	2.2	15.2
	Subgroup 4: <2 emergency department visits, ≤10 SABA prescriptions, no ICS + LABA, >50 yr of age	GLP1RAs	1,457	673	16	1.1	13.9	−0.6 (−1.4 to 0.1)	−0.5 (−1.3 to 0.3)
		SUs	2,669	1,227	46	1.7	14.2
	Subgroup 5: <2 emergency department visits, ≤10 SABA prescriptions, no ICS + LABA, ≤50 yr of age	GLP1RAs	1,114	514	14	1.3	12.6	−1.4 (−2.4 to −0.4)	−1.4 (−2.5 to −0.4)
		SUs	1,997	905	53	2.7	16.3

Definition of abbreviations: GLP1RA = glucagon-like peptide 1 receptor agonist; iCF = iterative causal forest; ICS = inhaled cortico steroid; IPTW = inverse probability treatment weight; LABA = long-acting β₂-agonist; SABA = short-acting β₂-agonist; SU = sulfonylurea.

All patients were required to enter the cohort no later than July 1, 2019, so that we could potentially follow patients for 0.5 years (the study ended on December 31, 2019). Patients were censored for insurance disenrollment. Propensity score trimming was applied in each identified subgroup. ICS + LABA includes the use of a combination product or the use of both medications simultaneously.

Figure 2 illustrates the extent to which a single potential unmeasured confounder would have to influence risk and be unbalanced across cohorts to nullify the observed 0.5-year RD observed in our study. The confounder would have to surpass the confounding bias of known factors such as emergency department visit for asthma (39). Note that such an unmeasured confounder would need to be this strong even after PS weighting using measured confounders. The strength of an unmeasured confounder in the GLP1RA versus thiazolidinedione comparison that, when controlled for, would lead to a null or harmful effect is smaller than that in the GLP1RA versus SU comparison. In the latter comparison, measured obesity has an RD of −1.8% and a PD of 17.2%, and emergency department visit for asthma has the highest RD of 10.1% and a PD of −3.8% (see Table E5).

Figure 2.

Bias analyses for unmeasured confounding. (A) Comparison of glucagon-like peptide 1 receptor antagonists (GLP1RAs) versus thiazolidinediones. (B) Comparison of GLP1RAs versus sulfonylureas. Index year is the year a GLP1RA or comparator drug was initiated. The shaded area indicates the strength of confounding implied if the true risk difference (RD) is null or harmful. The dotted line indicates the strength of confounding implied if true RD is −0.3% in the GLP1RA versus thiazolidinedione cohort (A) and −1.0% in the GLP1RA versus sulfonylurea cohort (B). A negative RD suggests a beneficial effect in the untreated cohort. A negative prevalence difference (PD) suggests a lower frequency in the GLP1RA group. Labels with solid circles indicate the observed RD and PD of confounders displayed on the absolute scale x-axis and y-axis. Note that only the following two scenarios involving a confounder can nullify our observed beneficial effect (negative RD): 1) the confounder has a negative RD (beneficial effect) and a positive PD (higher frequency in the GLP1RA group), or 2) the confounder has a positive RD (harmful effect) and a negative PD (lower frequency in the GLP1RA group). ED = emergency department.

Conditional Average Treatment Effect

The raw causal forest produced P values for homogeneity tests of 1.000 and 0.357 for the GLP1RA versus thiazolidinedione cohort and GLP1RA versus SU cohort, respectively. As a P value ≥0.05 does not imply homogeneity (22), it seems reasonable to further detect potential treatment heterogeneity for GLP1RAs versus SUs. We ignored P values and implemented iCF in the GLP1RA versus SU cohort using the selected 31 important covariates (Table 1; see Figures E6 and E7) to obtain subgroup decisions. The summary decision tree split the cohort by variation in exacerbation RD between GLP1RAs and SUs.

In the GLP1RA versus SU cohort, the first split of the tree was defined by two or more emergency department visits during the baseline period. For patients with fewer than two emergency department visits, the next split was defined by ≥10 short-acting β₂-agonists (SABAs), patients with <10 SABAs were further split by inhaled cortico steroids (ICSs) and long-acting β-agonists (LABAs), that is, ICS plus LABA (either a combination product or using both simultaneously), and those without ICS plus LABA were split by age >  50 years (during the baseline period) (Figure 3). We present subgroup-specific treatment effects in Table 2. In each subgroup, after PS trimming (decreasing sample size by ∼10%), the PS distributions of GLP1RAs and SUs are well overlapped (see Figure E8), and the covariates between the GLP1RA and SU groups have the same distribution after inverse probability treatment weighting adjustment (the median ASMD ranged from 0.017 to 0.031; see Tables E6 and E7). Subgroup 1, characterized by two or more emergency department visits (∼10% had emergency department visits for asthma during the 1-year baseline period) included 2,792 patients after PS trimming (Table 2). This group exhibited a higher proportion of comorbidities (diabetic complications, cardiovascular, and bronchitis), asthma medication use, and hospitalization (Tables 1, E6, and E7). GLP1RA initiators and SU initiators had exacerbation risks of 3.9% and 8.4%, respectively, with approximately 17% censored because of insurance disenrollment. This subgroup showed the largest absolute exacerbation risk reduction, with a decrease of 2.8% for GLP1RAs (95% confidence interval, −4.8% to −0.9%). Subgroup 2 (fewer than two emergency department visits and >10 SABA prescriptions) and subgroup 4 (fewer than two emergency department visits, ≤10 SABA prescriptions, no ICS plus LABA prescriptions, and age > 50 years) exhibited similar treatment effects, with aRDs of −0.7% (95% confidence interval, −2.2% to 0.8%) and −0.5% (95% confidence interval, −1.3% to 0.3%), respectively. Subgroup 3 (fewer than two emergency department visits, ≤10 SABA prescriptions, and ICS plus LABA prescriptions) and subgroup 5 (fewer than two emergency department visits, ≤10 SABA prescriptions, no ICS plus LABA prescriptions, and age ≤ 50 years old) also showed similar treatment effects, with aRDs of −1.4% (95% confidence interval, −2.3% to −0.5%) and −1.4% (95% confidence interval, −2.5% to −0.4%), respectively.

Figure 3.

Subgroup decisions from the iCF algorithm on the basis of baseline characteristics for the glucagon-like peptide 1 receptor antagonist (GLP1RA) versus sulfonylurea (SU) cohort. Negative values indicate decreased risk of exacerbation (benefit from GLP1RAs), whereas positive values indicate increased risk of exacerbation (harm from GLP1RAs). The sample size, event, and aRD are for the propensity score–trimmed population within each subgroup. The treatment effect is also detailed for the subpopulation before splitting in the dotted box. ICS + LABA includes the use of a combination product or the use of both medications simultaneously. aRD = adjusted risk difference (percentage) by inverse probability treatment weight; ED = emergency department; ICS = inhaled cortico steroid; LABA = long-acting β-agonist; SABA = short-acting β₂-agonist.

Discussion

We aimed to identify patients with T2D and asthma who might benefit more from GLP1RAs on asthma exacerbation in a large cohort using claims data and the iCF machine learning subgrouping algorithm. Overall, our finding on the average treatment effect of GLP1RA is consistent with the study by Foer and colleagues (17) that compared GLP1RAs with SUs. The smaller treatment effect with GLP1RAs relative to thiazolidinediones could possibly be due to the described beneficial effect of thiazolidinediones on asthma, although such effect has been observed only in one real-world study (thiazolidinediones were associated with decreased risk of asthma exacerbation with an odds ratio of 0.79 [95% confidence interval, 0.62–0.99] relative to nonthiazolidinedione during 1-year follow-up [25]) and a few small studies (26, 40, 41).

An unmeasured confounder could nullify the observed beneficial effect if it exhibits both a beneficial effect and a lower prevalence in the GLP1RA group or both a harmful effect and a higher prevalence in the GLP1RA group, that is, either a double negative in RD and PD (Figure 2) or a double positive in both. For emergency department visit for asthma, the residual confounding could conceal an even greater beneficial effect (a positive RD and a negative PD). Although we adjusted for codes for obesity, such claims clearly do not capture all the confounding due to obesity, as codes for obesity have been shown to have sensitivity less than 40% (42, 43). If obesity were accurately measured, given the direction of measured confounding by codes for obesity, the residual confounding from actual obesity would tend to reduce the observed beneficial effect. This aligns with previous evidence indicating a potential protective effect of obesity on respiratory conditions such as COPD (44). However, obesity is generally associated with a higher risk of asthma and asthma exacerbation (15), and the inconsistency could be due partly to measurement error.

HTEs are attributable to interactions between clinical features and treatment relative to comparator, which is likely to vary when compared among different comparators. In the GLP1RA versus SU cohort, in which we identified heterogeneity, the rate of emergency department visits for asthma among the SU initiators was 12.5%; thus, absolute aRDs for exacerbation of 1.6%, 2.8%, 1.4%, and 1.4% (from the overall population and subgroups 1, 3, and 5, respectively) are clinically meaningful (45).

When a clinical feature is not (accurately) measured, a decision tree tends to split population by a proxy of the feature (46); thus, subgroups identified need to be explained with caution. When comparing with SUs, emergency department visit is a proxy for asthma severity (the proportions of emergency department visits for asthma in the subpopulations with two or more and fewer than two emergency department visits during the baseline period are 37% and 5%, respectively), and a previous exacerbation is the strongest predictor for future events (47). Thus, subgroup 1, with the largest treatment effect, could be explained by GLP1RAs’ beneficial effect on asthma (10–13) and its exacerbation (17, 18), leading to the most pronounced benefit of GLP1RAs in those at highest risk. Notably, this is consistent with Foer and colleagues’ (17) findings. They observed that in the moderate/severe asthma subgroup with 211 GLP1RA initiators and 1,052 SU initiators, the beneficial effect was more pronounced (with an incidence rate ratio of 0.45 [95% confidence interval, 0.27–0.75]) than in the overall population during 0.5-year follow-up. Such consistency reinforces the reliability and strength of the iCF method. Although HTEs are less pronounced, treatment effects vary in subgroups defined by number of maintenance medications (ICSs plus LABAs) and as-needed medications (SABAs), as well as age, and further studies are needed to investigate the potential interaction between GLP1RAs and other medications as proxies of asthma severity (48) as well as age.

To the best of our knowledge, our study is the first large real-world study looking at the HTEs of GLP1RAs on asthma. The subpopulation with the largest treatment effect identified by iCF was not highlighted in the study of Foer and colleagues (17), likely because of limited sample size, nor were they detected in the study by Albogami and colleagues (18) using conventional analyses. Our finding, if confirmed, might support precision medicine strategies whereby treatments are tailored according to the patient characteristics and direct randomized studies toward specific target population (19). Furthermore, our bias analysis assessed the strength of a single confounder necessary to negate a beneficial effect of GLP1RAs versus SUs and revealed that unmeasured confounding from such a single confounder would need to be stronger than confounding from individual measured confounders in order for our finding to be consistent with a null or harmful true RD. This strengthens the credibility of our findings regarding treatment heterogeneity.

Strengths and Limitations

Our study has limitations. First, although we adopted pharmacoepidemiologic methods to reduce unmeasured confounding (e.g., an active-comparator, new-user design [28] and PS trimming), we cannot exclude the possibility of residual confounding that might also differ among subgroups, leading to spurious heterogeneity. We did not account for forced expiratory volume in 1 second, because lung function records are absent for the vast majority of patients in claims data (18, 49). Second, we assessed only severe asthma exacerbations. Third, we did not conduct external validation using a separate commercial claims dataset, so further research is required to evaluate our subgroup findings. Fourth, the choice of comparators aimed to assess whether GLP1RAs offer superior control of asthma exacerbations against therapeutic alternatives in people living with asthma. However, there is a possibility of residual confounding, as patients prescribed the more expensive GLP1RAs might also access higher quality care. Last, our findings are based on MarketScan enrollees younger than 65 years and likely employed. Therefore, extrapolation to older patients or those from lower socioeconomic backgrounds might not be appropriate.

Conclusions

Our study suggests that the short-term real-world use of GLP1RAs decreases the risk of asthma exacerbation compared with SUs. Furthermore, we assessed HTEs of GLP1RA relative to SUs and pinpointed that patients with at least two emergency department visits benefit most from GLP1RAs, using an iCF subgrouping algorithm. Clinicians might use previous emergency department visit (a proxy for asthma disease severity), number of maintenance and reliever inhalers, and age to identify subpopulation and provide precision treatment to maximize the benefit of GLP1RAs among patients with T2D and asthma to reduce the risk for asthma exacerbations.