Association between glucagon-like peptide-1 receptor agonist therapy and respiratory illness in patients with type 2 diabetes: a retrospective observational cohort study

TriNetX (Cambridge, MA, USA) is a global health research consortium that includes health-care organizations, researchers, and biopharmaceutical enterprises. This platform facilitates clinical research and supports efforts to improve clinical outcomes by providing access to deidentified patient data from diverse health-care institutions, enabling comprehensive real-world data analysis²². TriNetX is designed to generate real-world evidence through the analysis of electronic health records, claims data, and other data sets. The TriNetX network includes more than 220 health-care organizations across over 30 countries, predominantly large academic medical centers but also community hospitals, most with both inpatient and outpatient services. For this study, only structured electronic health record (EHR) data (demographics, diagnoses, procedures, prescriptions, laboratories, and vital signs) were available; unstructured records and linkages to external claims or cancer registries were not accessible. Routine TriNetX datasets do not include claims data, and while some institutions may internally link EHR data with claims before contributing, this is neither standardized nor common across the network. Likewise, direct linkage to external cancer registries is uncommon, although some sites may connect EHR data to their own hospital-based registries; this practice is also not consistent network-wide. The platform allows investigators to identify and analyze patient populations for clinical trials, streamlining the recruitment process. Furthermore, the platform provides advanced data visualization and analytical tools, enabling users to explore trends, treatment modalities, and outcomes. All data within TriNetX are de-identified, ensuring that patient identifiers are not accessible. The platform aggregates data counts and summary statistics from multiple institutions without including individual-level data. Given these safeguards, the Western Institutional Review Board has waived the requirement for informed consent for studies using TriNetX data.

This study adhered to the International Conference on Harmonization Good Clinical Practice guidelines, the Declaration of Helsinki, and applicable laws for noninterventional and observational studies. The study protocol was approved by the Ethics Review Board of Shin Kong Wu Ho-Su Memorial Hospital, Taiwan (protocol number: 20250104R). The requirement for informed consent was waived because all personal information was deidentified.

This retrospective population-based cohort study used both new-user and active comparator designs to mitigate biases in observational studies and enhance validity by approximating randomized controlled trial (RCT) conditions²³. Dipeptidyl peptidase-4 inhibitors (DPP4is) were chosen as the comparator because they share an incretin-based mechanism with GLP-1 RAs and are used as second-line anti-diabetic agents²⁴. From the TriNetX platform, we obtained patient data (more than 114 million records). The present study focused on individuals aged ≥ 18 years with at least three clinical visits and a confirmed diagnosis of T2DM recorded at least twice (n = 3,389,059) from January 1, 2005, to December 31, 2020. The cohort was divided into two treatment groups: users of GLP-1 RAs (n = 300,732) and users of DPP4is (n = 436,972). Participants with prior neoplasms or recent (6-month) use of either drug class were excluded, resulting in 201,153 new GLP-1 RA users and 323,114 new DPP4i users (Fig. 1).

GLP-1 RA exposure was identified using the Anatomical Therapeutic Chemical (ATC) code A10BJ, which includes exenatide, liraglutide, lixisenatide, albiglutide, dulaglutide, and semaglutide. DPP-4 inhibitor exposure was identified using ATC code A10BH, which includes sitagliptin, vildagliptin, saxagliptin, linagliptin, and alogliptin.

The first medication exposure was defined as the index event, with a 3-month lag applied to all exposures to minimize protopathic bias and allow for sufficient latency after cohort entry. The outcomes of interest were evaluated within 90 days to 10 years after the index date, with follow-up extending until January 21, 2025. This study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology reporting guidelines for observational studies²⁵. Cohort definitions in TriNetX are detailed in eMethod 1 of the supplementary file.

Propensity score matching (PSM) was performed to compare new GLP-1 RA users with new DPP4i users. PSM was performed in a 1:1 ratio to mitigate the effects of potential confounding factors. The PSM full model was adjusted for pre-specified baseline covariates, including demographic factors (e.g., proportion of male patients and ethnic composition of the study cohort), comorbidities, medication use, and laboratory metrics (details are shown in Table 1). Diagnostic codes for baseline variables are listed in Supplementary Materials (eMethod 1 of supplementary file).

We used the PSM tool integrated within TriNetX to compute propensity scores and perform 1:1 matching. Propensity scores were generated through logistic regression, and matching was performed with a greedy nearest-neighbor algorithm, applying a caliper of 0.1 for pooled standardized mean differences (SMDs). The analysis was performed using Python (Python Software Foundation, Wilmington, DE, USA) and R (version 3.4.4; R Foundation for Statistical Computing, Vienna, Austria). This approach ensured the creation of comparable groups, thereby enhancing the accuracy of treatment effect analyses (eMethod 2 of supplementary file).

We evaluated several clinical outcomes related to lung health in patients treated with GLP-1 RAs and DPP4is. The primary outcome was the incidence of lung cancer, with additional analyses examining specific tumor sites, such as the trachea and bronchus. The secondary outcomes included influenza and pneumonia, other acute lower lung infection, suppurative lung disease, and pulmonary fibrosis. We included a broader range of respiratory conditions as secondary outcomes to capture more clinically relevant morbidities potentially influenced by the exposure. These encompassed both upper and lower airway diseases of varying acuity, from acute infections to chronic inflammatory disorders. This approach allowed us to assess the consistency of associations across related phenotypes and explore possible mechanism-specific effects. These outcomes were evaluated to compare risks between the two treatment groups (eMethod 1 of supplementary file).

To test the robustness of our findings, we performed analyses using both positive and negative outcome controls. GLP-1 RAs have been demonstrated to reduce cardiovascular and renal risks in patients with T2DM^26,27. Therefore, major adverse cardiovascular events (e.g., cerebral infarction, myocardial infarction, and death) and major adverse kidney events (e.g., acute and advanced kidney failure, dialysis initiation, and death) were used as positive outcome controls. By contrast, bone fractures and scleroderma, which have no reported associations with GLP-1 RAs, were included as negative outcome controls.

To validate our findings, we performed sensitivity analyses: (1) by extending the time lag for outcome analysis from 3 to 6 months and 12 months after the index date, (2) by using different PSM models adjusted for multiple confounders (PSM Model 1 was matched for age, sex, race, body mass index [BMI], lifestyle, and socioeconomic status; Model 2 was matched for baseline comorbidities in addition to Model 1 covariates; and Model 3 was further matched for baseline medication use), and (3) by analyzing the risk profile across different time frames (3 months to 3 years, 3 months to 5 years, and 3 months to 7 years).

The baseline characteristics of the treatment groups were analyzed both before and after PSM. Summary statistics for continuous variables are presented as mean ± standard deviation values, whereas those for categorical variables are presented as frequency (%) values. SMDs were calculated to determine the balance of baseline characteristics between the groups, with an SMD value of < 0.1 indicating minimal differences²⁸.

Study outcomes were analyzed through risk and survival analyses. A Cox proportional-hazards model was used to calculate the hazard ratio (HR) and corresponding 95% confidence interval (CI) values for comparing outcome incidence between groups. Kaplan–Meier curves were generated to visualize time-to-event data, and the log-rank test was used to compare survival distributions. The proportional-hazards assumption was assessed using the generalized Schoenfeld approach integrated into the TriNetX platform through R's Survival package (version 3.2-3). If this assumption was violated, HRs were calculated for different periods. We tested seven hypotheses with a false discovery rate of 0.05, correcting P values using the Benjamini–Hochberg procedure²⁹.

Subgroup analyses were performed to examine the consistency of treatment effects across key baseline characteristics, such as age (≥ 65 and < 65 years), sex (male and female), BMI (≥ 30 and < 30 kg/m²), estimated glomerular filtration rate (eGFR; ≥60 and < 60 mL/min/1.73 m²), and HbA1c (≥ 7% and < 7%). Further analyses compared lung conditions between GLP-1 RAs and other antidiabetic medications, including insulin, biguanides, sodium-glucose cotransporter 2 inhibitor, sulfonylureas, and thiazolidinediones.

All statistical analyses were performed using the TriNetX real-time analytics platform. A two-sided p value of < 0.05 indicated statistical significance.

After PSM incorporating all the baseline parameters, each drug group had 158,224 users. The baseline characteristics of the treatment groups before and after PSM are summarized in Table 1. Before PSM, significant differences were observed between the groups. GLP-1 RA users were younger than DPP4i users (mean age: 55.4 vs. 60.3 years; SMD: 0.398). Furthermore, the prevalence of obesity and insulin use was higher among GLP-1 RA users than among DPP4i users (obesity, BMI ≥ 35 kg/m²: 26.8% vs. 16.6% [SMD: 0.249]; insulin use: 36.1% vs. 24.7% [SMD: 0.250]). After PSM, these differences were minimized, with GLP-1 RA and DPP4i users having similar mean age (56.7 vs. 56.4 years; SMD = 0.023) and BMI (23.2% vs. 23.1%; SMD = 0.003) observed between the groups. Moreover, both groups exhibited similar sex distributions (44.5% men) and racial compositions. Comorbidities such as hypertension, dyslipidemia, and chronic kidney disease were prevalent in both groups, with minimal differences noted after PSM (e.g., hypertension: 51.9% vs. 51.7% [SMD: 0.005]). After PSM, between-group balance was observed also in laboratory results such as HbA1c (8.6% vs. 8.4%; SMD: 0.053) and eGFR (83.1 vs. 83.5 mL/min/1.73 m²; SMD: 0.015). Lifestyle factors such as nicotine dependence and alcohol-related disorders exhibited negligible differences between the groups after PSM. Medication use patterns (biguanides, sulfonylureas, and lipid-modifying agents) were similar between the two groups after PSM. Overall, PSM effectively reduced baseline differences between GLP-1 RA and DPP4i users, ensuring similarity for subsequent analyses.

The median follow-up was 1882 days (interquartile range 1153) for GLP-1 RA users and 2078 days (interquartile range 1682) for DPP4i user. Event-free survival rates of lung cancer differed significantly between GLP-1 RA and DPP4i users (p < 0.001; eFigure 1 of supplementary file). At 5-year follow-up, event-free survivals rate of lung cancer were approximately 99.4% and 99.2% for GLP-1 RA and DPP4i users, respectively. By 10 years, these rates were approximately 98.6% and 98.3% for GLP-1 RA and DPP4i users, respectively.

After PSM, GLP-1 RA users had a significantly lower risk of lung cancer than did DPP4i users (HR 0.86; 95% CI 0.80–0.94; Table 2). The risk reduction remained significant specifically for bronchial cancers (HR 0.91; 95% CI 0.84–0.99). These findings remained significant after Benjamini-Hochberg false discovery rate correction.

GLP-1 RA users consistently show higher event-free rates for influenza, pneumonia (log rank p < 0.001), acute lower lung infection (log rank p < 0.001), suppurative lung disease (log rank p < 0.001), and pulmonary fibrosis (log rank p = 0.01), indicating better outcomes compared to DPP4i users (eFigure 2 of supplementary file). Table 2 shows that GLP-1 RA users is associated with significantly lower risks of secondary outcomes compared to DPP4i users. For influenza and pneumonia, the HR was 0.94 (95% CI 0.92–0.96); for other acute lower lung infections, HR was 0.85 (95% CI 0.82–0.87); for suppurative lung disease, HR was 0.74 (95% CI 0.65–0.84); and for pulmonary fibrosis, HR was 0.92 (95% CI 0.87–0.98).

Subgroup analyses revealed that GLP-1 RAs use was associated with a reduced risk of lung cancer across multiple demographic groups (Fig. 2). The protective effect was observed in patients aged < 65 years as well as in those aged ≥ 65 years. The risk reduction was also significant among African American patients and White patients. Similar protective effects were observed across various eGFR groups and in individuals with BMI < 30 Kg/m² and HbA1c levels of ≥ 7%.

For secondary pulmonary outcomes comparing GLP‑1 RA and DPP4i treatments across various groups. Across nearly all subgroups, GLP‑1 RA is consistently associated with a reduced risk of influenza and pneumonia, acute lower lung infection, suppurative lung disease, and pulmonary fibrosis, with effect sizes generally favoring GLP‑1 RA over DPP4i (Fig. 3). These protective effects remain robust regardless of demographic or clinical factors, highlighting that GLP-1 RA use is associated with reduced risk of respiratory infection and lung fibrosis across diverse patient populations with T2DM (detail are shown in eFigs. 3, 4, 5 and 6 of supplementary file).

Figure 4 shows the risk comparison of lung cancer between GLP-1 RA and other antidiabetic medications. The users of GLP-1 RAs consistently had a lower risk of lung cancer than did the users of most other diabetes medications, such as insulin (HR 0.68; 95% CI 0.62–0.75), biguanides (HR 0.84; 95% CI 0.76–0.93), sulfonylureas (HR 0.84; 95% CI 0.78–0.91), and thiazolidinediones (HR 0.87; 95% CI 0.80–0.95). However, the risk was similar between GLP-1 RA users and sodium–glucose cotransporter 2 inhibitor (SGLT2i) users (HR 0.93; 95% CI 0.84–1.01).

Figure 5 shows that GLP-1 RA users had lower event rates and HRs for acute lower lung infections and suppurative lung disease versus most other diabetes drugs, especially insulin, biguanides, sulfonylureas, and thiazolidinediones, but not always SGLT2i. For influenza, pneumonia, and pulmonary fibrosis, GLP-1 RA showed benefits mainly over insulin. Overall, GLP-1 RA generally lowers pulmonary risks compared to most drugs except SGLT2i.

Positive outcome control analyses indicated that GLP-1 RA users had markedly lower risks of major adverse cardiovascular and kidney events than did DPP4i users. However, negative outcome control analyses revealed no significant between-group differences in the incidence of bone fractures or scleroderma (Supplementary Materials, eTable 1). Our findings remained consistent when the analysis time lag was extended from 3 to 6 or 12 months after the index date. However, pulmonary fibrosis lost significance with a 12-month lag (Supplementary Materials, eTable 2 and eTable 3). Furthermore, the results were consistent across various PSM models after adjustments were made for multiple confounders (Supplementary Materials, eTable 4) and across diverse temporal frameworks (Supplementary Materials, eTable 5).