GLP-1 receptor agonists and clinical outcomes in adults with Crohn’s disease and obesity: a propensity score–matched real-world cohort study

Abstract

Obesity affects 15%–40% of adults with Crohn’s disease (CD) and is associated with worse outcomes. GLP-1 receptor agonists (GLP-1RAs) have pleiotropic anti-inflammatory effects, yet their clinical impact in CD with comorbid obesity is incompletely characterized. We evaluated whether GLP-1RA initiation was associated with mortality, health care utilization, and TNF inhibitor use as a proxy measure of treatment intensity. Retrospective new-user cohort study using the TriNetX Global Collaborative Network (162 organizations, 2015–2024). Adults (≥ 18 years) with CD and obesity without prior GLP-1RA exposure were included. Nonusers received calendar-time–matched index dates to minimize immortal time bias. Propensity score matching (1:1) incorporated > 30 covariates including comorbidities, CD treatment intensity proxies, and inflammatory markers. Outcomes were assessed at 1 year (days 1–365 post-index) and 5 years (days 1–1,825 post-index); both time windows were analyzed as fixed horizons under an intention-to-treat framework using separate TriNetX query runs. Patients were censored at last recorded EHR encounter or study end (2024). Due to database constraints, detailed longitudinal exposure data (adherence, discontinuation, agent switching) were not available. Among 10,550 GLP-1RA users and 108,944 nonusers, 9,766 pairs were matched at 1 year and 9,320 at 5 years (the latter representing a subset of the base population with sufficient follow-up). After matching, covariates were well-balanced (all SMDs < 0.1 except BMI by design). At 1 year, GLP-1RA initiation was associated with lower mortality (0.7% vs. 4.2%; RR 0.17, 95% CI 0.13–0.22), hospitalization (10.0% vs. 24.7%; RR 0.40, 95% CI 0.38–0.43), ED visits (18.9% vs. 25.3%; RR 0.75, 95% CI 0.71–0.79), and TNF inhibitor use (7.1% vs. 10.9%; RR 0.65, 95% CI 0.60–0.72); corticosteroid rates were similar (RR 1.00, 95% CI 0.96–1.03). At 5 years, all utilization associations were sustained; corticosteroid use was modestly lower (RR 0.91, 95% CI 0.89–0.94). E-value for 1-year mortality upper CI: 8.56; healthy-user bias cannot be excluded. In adults with CD and obesity, GLP-1RA initiation was associated with lower health care utilization and lower TNF inhibitor use at 1 and 5 years; the latter should not be interpreted as evidence of reduced treatment escalation given inability to distinguish new initiation from continuation. The observed mortality difference is unlikely to reflect a true causal effect and should be treated as hypothesis-generating. These findings support GLP-1RA metabolic safety in this population and warrant prospective evaluation.

Introduction

Crohn’s disease is a chronic, relapsing inflammatory disorder of the gastrointestinal tract associated with substantial morbidity, progressive bowel damage, and high health care utilization^1,2. In parallel with global population trends, obesity has become increasingly prevalent among adults with Crohn’s disease: contemporary estimates suggest that 15%–40% of patients meet criteria for obesity, with an additional 20%–40% classified as overweight^3–5. Specifically, a population-based cohort study by Johnson et al. reported an obesity prevalence of approximately 25%–30% in patients with incident Crohn’s disease⁵, while comprehensive reviews have documented wide variation across regions and study periods, reflecting both global adiposity trends and the heterogeneity of Crohn’s disease phenotypes^4–6. This demographic shift has complicated disease management and introduced a clinically important patient subgroup that remains undercharacterized in the literature.

Obesity is increasingly recognized as a biologically active modifier of inflammatory bowel disease (IBD) rather than a coincidental comorbidity^3,7–11. Visceral adipose tissue—particularly mesenteric fat—functions as an immunometabolic organ that drives chronic immune activation through adipokine dysregulation, proinflammatory cytokine production (including TNF-α and IL-6), immune cell infiltration, and microbiome perturbation^7–11. In Crohn’s disease, mesenteric fat expansion (creeping fat) has been associated with transmural inflammation, fibrostenotic complications, impaired epithelial barrier function, increased corticosteroid exposure, and higher hospitalization rates^3,4,7–11. Obesity also appears to reduce response to biologic therapies, particularly TNF inhibitors, via altered pharmacokinetics and persistent low-grade systemic inflammation^4,12,13.

Glucagon-like peptide-1 receptor agonists (GLP-1RAs), initially developed for type 2 diabetes and later approved for obesity, have demonstrated pleiotropic effects beyond weight reduction. Preclinical evidence suggests GLP-1 signaling attenuates intestinal inflammation, enhances epithelial barrier integrity, and modulates innate and adaptive immune responses in experimental colitis models^17–19. Observational studies and meta-analyses in IBD cohorts have reported associations between GLP-1RA use and reduced corticosteroid exposure, fewer hospitalizations, and lower rates of IBD-related surgery, without evidence of disease exacerbation^14–16,20. However, patients with IBD were systematically excluded from pivotal GLP-1RA randomized controlled trials, and available clinical data remain retrospective and susceptible to residual confounding.

Despite growing interest, clinical outcomes of GLP-1RA therapy in adults with both Crohn’s disease and comorbid obesity remain incompletely characterized. We therefore conducted a large, multicenter, real-world cohort study to evaluate whether GLP-1RA initiation was associated with systemic corticosteroid use, health care utilization, TNF inhibitor use, and all-cause mortality among adults with Crohn’s disease and obesity.

Methods

Study design and data source

We conducted a retrospective, multicenter, new-user cohort study using data from the TriNetX Global Collaborative Network, a federated electronic health record (EHR) platform aggregating deidentified patient-level data from 162 healthcare organizations across the United States and internationally. Analyses were performed within the TriNetX analytics environment in October–November 2025. The platform provides built-in tools for cohort construction, propensity score matching (PSM), and time-to-event analysis, returning aggregated summary statistics without individual-level data access. All data are fully deidentified in compliance with the Health Insurance Portability and Accountability Act (HIPAA) Safe Harbor standard (45 CFR § 164.514(b)). This study was therefore exempt from formal Institutional Review Board (IRB) review and approval. Exemption was confirmed under the U.S. Common Rule (45 CFR § 46.104(d)(4)) by the Howard University Institutional Review Board (Howard University IRB, Washington, DC, USA), which determined that research involving exclusively pre-existing, fully deidentified data does not constitute human subjects research requiring IRB oversight. This research was conducted in accordance with the Declaration of Helsinki.

Because the TriNetX Global Collaborative Network includes organizations from multiple countries, GLP-1RA initiation during earlier years of the study period (2015–2020) may disproportionately reflect use for type 2 diabetes rather than obesity indications, as obesity-specific approvals for agents such as liraglutide (2014 in the US), semaglutide (2021 in the US), and tirzepatide (2023 in the US) occurred at varying timepoints and with regional variation. This potential confounding by indication is acknowledged as a limitation (see Limitations) and underscores the importance of interpreting findings as hypothesis-generating.

Eligibility criteria

We identified adults aged ≥ 18 years with Crohn’s disease and obesity. Crohn’s disease ascertainment required at least one ICD-10-CM K50.x diagnosis code recorded in the EHR at any time prior to or at the index date. We recognize that reliance on a single diagnosis code introduces potential for misclassification, a limitation common to EHR-based studies that is explicitly acknowledged below. Obesity ascertainment required either an ICD-10-CM E66.x code or a recorded BMI ≥ 30 kg/m² at any time in the EHR. Both conditions (Crohn’s disease and obesity) were required to be documented prior to or at the index date to ensure that treatment initiation occurred in the context of established comorbidity; patients in whom obesity documentation occurred exclusively after the index date were not eligible. BMI was ascertained from the most recent recorded measurement prior to the index date where available; the completeness of BMI data by group is described in the Limitations. We acknowledge that active Crohn’s disease can cause weight loss and that some patients classified as obese at an earlier timepoint may have had lower BMI proximate to GLP-1RA initiation; this represents a residual misclassification risk that cannot be fully excluded with EHR-based data and is discussed further in the Limitations.

Patients were excluded if they had a diagnosis of ulcerative colitis (K51.x) at any time, a pregnancy procedure code (ICD-10-PCS 10.x), or any prior GLP-1RA exposure—defined as a prescription for semaglutide (RxNorm 1991302), liraglutide (RxNorm 475968), tirzepatide (RxNorm 2601723), or any ATC class A10BJ agent—before the assigned index date. These criteria identified 10,550 GLP-1RA users and 108,944 eligible nonusers before matching. Cohort attrition is detailed in Online Resource 1 (Table S1) and illustrated in Fig. 1.

Fig. 1.

Cohort attrition and propensity score matching. Consolidated flow diagram showing patient identification, exclusions, propensity score matching steps, and final analytic cohort sizes. The two sections of the figure correspond to the 1-year (days 1–365) and 5-year (days 1–1,825) analytic time horizons, derived from separate TriNetX platform runs (Trinetx_4 and Trinetx_3, respectively). These do not represent two independent cohorts; both derive from the same base population, and the 5-year cohort is a subset of the 1-year cohort. Slight differences in pre-matching denominators reflect time window–specific data availability. CD = Crohn’s disease; UC = ulcerative colitis; GLP-1RA = glucagon-like peptide-1 receptor agonist; PSM = propensity score matching.

We note that immunomodulator use (thiopurines, methotrexate; ATC L04AX) was not captured as a covariate or outcome in the current analysis, representing a potential source of residual confounding for corticosteroid and TNF inhibitor outcomes, as detailed in the Limitations.

Exposure definition and new-user design

The exposure of interest was initiation of a GLP-1 receptor agonist (semaglutide, liraglutide, tirzepatide, dulaglutide, exenatide, or lixisenatide). To implement a new-user design, GLP-1RA initiators were required to have ≥ 12 months of observable EHR data prior to treatment initiation and no GLP-1RA prescriptions during that washout period. The index date was defined as the date of the first GLP-1RA prescription following the washout.

To minimize immortal time bias, nonexposed patients were assigned index dates using calendar-time risk-set sampling. For each exposed patient’s index date, eligible nonusers were required to be alive, actively observable within the network, and GLP-1RA naïve on that same calendar date. A minimum of ≥ 12 months of observable EHR data prior to the index date was also required for nonusers to ensure comparable opportunity for baseline covariate measurement. This approach ensures that both groups share a comparable time zero and equivalent at-risk follow-up from cohort entry.

Follow-up structure and intention-to-treat framework

Outcomes were analyzed under an intention-to-treat (ITT) framework, necessitated by the limitations of the TriNetX platform, which does not provide granular longitudinal prescription-level data (e.g., refill records, days of supply, or discontinuation dates) needed to implement per-protocol or as-treated analyses. Accordingly, patients were classified by their treatment group as assigned at the index date, regardless of subsequent changes in therapy.

Two fixed follow-up windows were pre-specified to capture distinct clinical timeframes: 1 year to assess early associations with health care utilization and mortality within the period most likely to reflect on-treatment effects, and 5 years to assess whether observed associations were sustained over a horizon consistent with the natural course of Crohn’s disease and the expected duration of metabolic benefit. These were analyzed as separate TriNetX query runs: (1) 1-year follow-up (days 1–365 after the index date, derived from the “Outcomes at 1 year” run [Trinetx_4]) and (2) 5-year follow-up (days 1–1,825 after the index date, derived from the “Outcomes at 5 years” run [Trinetx_3]). Both cohorts were drawn from the same base population of 10,550 GLP-1RA users and 108,944 nonusers; the smaller matched sample in the 5-year cohort (9,320 pairs vs. 9,766 pairs) reflects the subset of patients with ≥ 12 months of EHR activity beyond the 1-year horizon necessary to contribute to the longer follow-up window. There is accordingly substantial patient overlap between the two analytic cohorts: all patients contributing to the 5-year analysis are a subset of those in the 1-year analysis. Patients were censored at the last recorded EHR encounter, the administrative end of the study period (2024), or at outcome event occurrence, whichever occurred first. The two analyses are therefore not independent and should be interpreted as presenting the same cohort at different time horizons rather than separate populations.

Baseline covariates

Covariates were assessed during the 12 months prior to the index date and included: (1) demographics—age, sex, race/ethnicity, and tobacco use (Z72.0); (2) cardiometabolic comorbidities—diabetes mellitus (E08–E13), essential hypertension (I10), chronic kidney disease (N18), and COPD (J44); (3) Crohn’s disease treatment intensity and frailty proxies—prior systemic glucocorticoid use (ATC class HS051; serving as a surrogate for steroid dependence and cumulative inflammatory burden), prior TNF inhibitor use (ATC L04AB; serving as a surrogate for prior biologic or small molecule therapy exposure and disease severity), and prior digestive surgical procedures (ICD-10-PCS 1006964; serving as a surrogate for prior CD-related surgical morbidity and hospitalization history); and (4) laboratory markers—C-reactive protein, erythrocyte sedimentation rate, albumin, hemoglobin, BMI, ferritin, and complete metabolic panel components. These treatment intensity proxies served collectively as the available surrogates for Crohn’s disease severity and overall frailty within the TriNetX platform; direct measures of endoscopic activity, Montreal phenotype, functional status, or validated frailty scores were not available.

Missingness was substantial for several laboratory covariates, as is common in federated EHR data where testing reflects clinical indication rather than systematic measurement. Specifically, CRP was available for 51.1% of GLP-1RA users and 25.7% of nonusers before matching; ESR for 49.7% and 25.3%; albumin for 87.8% and 56.3%; and ferritin for 43.8% and 18.9%. Missing laboratory values were handled within the TriNetX platform using a missing indicator approach—patients without a recorded value for a given laboratory covariate were assigned a missing indicator, and this indicator was included as a covariate in the propensity score model. This approach assumes that missingness is adequately captured by the indicator itself; however, this assumption may not hold under informative missingness, where the probability of a laboratory value being recorded depends on the patient’s underlying health status or disease severity. In EHR-based studies, testing is frequently driven by clinical indication rather than systematic ascertainment, making informative missingness likely. This differential missingness pattern—where GLP-1RA users had more complete laboratory data—likely reflects greater healthcare engagement in the treated group and represents a potential source of residual measurement bias. The higher testing rates in GLP-1RA users may partially explain the higher albumin and hemoglobin observed in that group before matching, as sicker patients in the nonuser group may have been less likely to have these values captured. Sensitivity analyses using complete-case methods or multiple imputation would provide stronger evidence of robustness but were not feasible within the TriNetX platform’s aggregated output interface, which does not permit individual-level data export required for these approaches.

Diabetes status and GLP-1RA indication

Given that diabetes mellitus was substantially more prevalent in GLP-1RA users before matching (53.0% vs. 12.2%; SMD 0.97), and that diabetes is a strong predictor of mortality and health care utilization, diabetes was included as a covariate in the propensity score model. After matching, diabetes prevalence was balanced (51.0% vs. 50.6%; SMD 0.008). Baseline HbA1c data were not systematically available within the TriNetX platform and therefore could not be reported. Stratified analyses by diabetes status were not feasible within the platform’s aggregated output interface but represent an important avenue for future investigation. Similarly, while individual BMI values were available (mean BMI 37.1 vs. 31.4 kg/m² before matching), formal BMI category stratification (Class I: 30–34.9; Class II: 35–39.9; Class III: ≥40 kg/m²) and stratified analyses could not be conducted within the platform’s aggregate interface. Regarding GLP-1RA agent distribution: the cohort captured semaglutide (RxNorm 1991302), liraglutide (RxNorm 475968), tirzepatide (RxNorm 2601723), and ATC class A10BJ agents broadly (including dulaglutide and exenatide); however, the platform’s aggregate output does not stratify matched cohort characteristics or outcomes by individual agent. Given the predominance of semaglutide prescribing in the latter study period (2021–2024) and the diabetes-era dominance of liraglutide and dulaglutide (2015–2020), the cohort reflects a shifting mixture of agents whose relative proportions could not be extracted. Agent-specific analyses remain an important priority for future studies.

Propensity score matching

GLP-1RA users were matched 1:1 to nonusers using propensity scores derived from a logistic regression model incorporating all baseline covariates. Nearest-neighbor matching without replacement was applied using a caliper width of 0.1 standard deviations of the logit of the propensity score. Covariate balance was assessed using standardized mean differences (SMDs); values less than 0.1 indicate adequate balance (Fig. 2). Propensity score density plots before and after matching confirmed improved distributional overlap and are presented graphically in Fig. 2 (density curves) and in the TriNetX output reports (Online Resource 2).

Fig. 2.

Covariate balance and propensity score overlap. Upper panel: Covariate balance before (red circles) and after (blue squares) propensity score matching, shown as absolute standardized mean differences (SMDs). Lines connect before/after values for each covariate. Dashed orange vertical line = SMD 0.1 threshold. BMI, albumin, and hemoglobin remain imbalanced after matching by design or differential testing (see text). Lower panel: Propensity score density distributions before and after matching (GLP-1RA users: purple; nonusers: green), demonstrating improved distributional overlap following matching. CRP = C-reactive protein; ESR = erythrocyte sedimentation rate.

Outcomes

The primary outcome was new corticosteroid use during follow-up, captured using ATC class R01AD. As discussed in the Limitations, this ATC class encompasses both systemic glucocorticoids and inhaled/intranasal preparations within the TriNetX classification; findings should be interpreted as reflecting corticosteroid prescribing broadly rather than as a specific marker of IBD disease activity. This primary outcome is presented first in all results tables, with secondary outcomes following in sequence.

Secondary outcomes included: all-cause mortality; all-cause hospitalization (CPT 1013659); emergency department (ED) visits (CPT 1013711); and TNF inhibitor use (ATC L04AB). TNF inhibitor use captures any recorded prescription or administration in the follow-up window and cannot distinguish new initiation from continuation of pre-existing biologic therapy; this limitation is relevant to interpretation of this outcome as a proxy for treatment escalation. Outcomes were evaluated at 1 year (365 days) and 5 years (1,825 days). For mortality, patients with a recorded death prior to the analysis time window were excluded (~ 22–23 GLP-1RA users and ~ 32–43 nonusers per run). Follow-up was otherwise censored after the last recorded EHR encounter or study end (2024).

Statistical analysis

Risk ratios (RRs) with 95% confidence intervals were calculated for each outcome at fixed time horizons. Kaplan–Meier analyses with log-rank testing and Cox proportional hazards models were used for time-to-event comparisons, with hazard ratios (HRs) and 95% CIs reported. All tests were two-sided; P < .05 defined statistical significance. E-values were calculated for all outcomes to quantify robustness to unmeasured confounding (Fig. 3; Online Resource 1, Table S2). Hospitalization and ED visit counts are presented as means; given the expected skewness of utilization count outcomes, these should be interpreted with appropriate caution, and future analyses should consider median and distributional measures. Additional sensitivity analyses—including inverse probability of treatment weighting (IPTW), negative control outcome analyses, complete-case analyses restricted to patients with fully observed laboratory data, and multiple imputation for missing laboratory covariates—were not feasible within the TriNetX platform’s aggregated output interface, which does not permit the individual-level data export required for these approaches. These analyses are recommended as important components of future replication studies using platforms that provide individual-level data access. In particular, complete-case and multiple imputation analyses are needed to evaluate whether the missing-indicator approach adequately controlled for laboratory-based confounding under informative missingness.

Fig. 3.

E-value sensitivity analysis for all outcomes at 1-year and 5-year follow-up. Filled bars = E-value for the point estimate; open bars = E-value for the upper 95% confidence limit. Dashed orange line = E-value 3.0 (moderate robustness threshold). These values quantify robustness to a specific magnitude of unmeasured confounding but do not exclude qualitative biases. Corticosteroid use at 1 year (RR = 1.00) has E-value = 1.0 (null result; not shown). ED = emergency department; TNF = tumor necrosis factor inhibitor.

All language in this manuscript is explicitly framed in terms of association rather than causation. The study design, while incorporating several pharmacoepidemiologic best practices, cannot establish causal effects due to the potential for residual confounding from unmeasured variables.

Results

Study population and baseline characteristics

Among 10,550 adults with Crohn’s disease and obesity who initiated GLP-1RA therapy and 108,944 eligible nonusers, 9,766 matched pairs were included in the 1-year analytic cohort and 9,320 matched pairs in the 5-year cohort. Both cohorts were derived from the same base population; the 5-year cohort represents the subset with sufficient follow-up (see Methods and Table S1). Baseline characteristics are presented in Table 1.

Table 1.

Baseline characteristics of the propensity score–matched 1-year analytic cohort (n = 9,766 per group).

Characteristic	GLP-1RA Users (n = 9,766)	Nonusers (n = 9,766)	SMD
Demographics
Age at index, mean ± SD, years	54.0 ± 13.4	54.6 ± 16.9	0.036
Female sex, n (%)	6,629 (67.9)	6,809 (69.7)	0.040
White race, n (%)	7,596 (77.8)	7,641 (78.2)	0.011
Black or African American, n (%)	1,307 (13.4)	1,283 (13.1)	0.007
Tobacco use, n (%)	699 (7.2)	663 (6.8)	0.014
Comorbidities, n (%)
Diabetes mellitus (E08-E13)	4,983 (51.0)	4,946 (50.6)	0.008
Diabetes prevalence pre-matching	53.0%	12.2%	0.966
Essential hypertension (I10)	6,367 (65.2)	6,351 (65.0)	0.003
Chronic kidney disease (N18)	1,423 (14.6)	1,368 (14.0)	0.016
COPD (J44)	1,093 (11.2)	1,071 (11.0)	0.007
CD Treatment Proxies, n (%)
Prior systemic corticosteroid (ATC HS051)	7,758 (79.4)	7,829 (80.2)	0.018
Prior TNF inhibitor (ATC L04AB)	1,573 (16.1)	1,551 (15.9)	0.006
Prior digestive surgical procedures	5,071 (51.9)	5,050 (51.7)	0.004
Laboratory Values, mean ± SD
BMI, kg/m² *	37.1 ± 7.4	31.4 ± 6.7	0.810
C-reactive protein, mg/L	19.0 ± 36.4	25.4 ± 49.5	0.148
Albumin, g/dL *	4.1 ± 0.4	3.9 ± 0.6	0.322
Hemoglobin, g/dL *	13.3 ± 1.7	12.6 ± 2.1	0.372
Erythrocyte sedimentation rate, mm/h	25.4 ± 22.2	27.5 ± 25.8	0.088
Ferritin, ng/mL	151.3 ± 362.0	241.1 ± 1390.0	0.088

SMD standardized mean difference. Values < 0.1 indicate adequate covariate balance. COPD, chronic obstructive pulmonary disease; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; TNF, tumor necrosis factor.

*Expected imbalance by design (see text). Pre-matching diabetes prevalences are shown for transparency regarding confounding-by-indication.

Before matching, GLP-1RA users were substantially older (mean 54.0 vs. 49.6 years; SMD 0.29), more frequently female (68.8% vs. 59.3%; SMD 0.20), and had markedly higher prevalences of diabetes mellitus (53.0% vs. 12.2%; SMD 0.97), essential hypertension (66.2% vs. 30.7%; SMD 0.76), prior systemic corticosteroid use (80.2% vs. 44.0%; SMD 0.81), and prior TNF inhibitor use (17.5% vs. 8.7%; SMD 0.26), consistent with expected confounding-by-indication.

After propensity score matching, all key clinical covariates achieved adequate balance (Table 1; Fig. 2). SMDs for comorbidities (diabetes 0.008; hypertension 0.003; CKD 0.016), Crohn’s treatment proxies (prior corticosteroid 0.018; prior TNF inhibitor 0.006; prior surgery 0.004), and inflammatory markers (CRP 0.148; ESR 0.088) were substantially reduced. Residual imbalances in BMI (SMD 0.810), albumin (SMD 0.322), and hemoglobin (SMD 0.372) were pre-specified expected findings: BMI reflects the mandatory obesity indication for GLP-1RA prescribing; the albumin and hemoglobin differences likely reflect both the higher BMI in GLP-1RA users (which is associated with higher hemoglobin and albumin in the absence of inflammatory states) and differential completeness of laboratory testing between groups. These residual imbalances suggest that complete adjustment for disease severity was not fully achieved by propensity score matching alone and should be considered when interpreting outcomes, particularly hospitalization and mortality.

Figure 1 flow diagram presents the cohort attrition and matching steps. The two sections of the figure correspond to the 1-year and 5-year analytic time horizons, respectively, derived from separate TriNetX query runs; they do not represent two independent cohorts but rather the same underlying population assessed at different follow-up windows.

Clinical outcomes at 1 year and 5 years

All outcomes are summarized in Table 2; Figs. 4 and 5. Results are presented with the primary outcome (corticosteroid use) first, followed by secondary outcomes.

Table 2.

Clinical outcomes at 1-year and 5-year follow-up in propensity score–matched cohorts.

Outcome	GLP-1RA Users n/N (%)	Nonusers n/N (%)	RR (95% CI)	P	HR (95% CI)	P
1-Year Follow-Up (n = 9,766 matched pairs)
Corticosteroid use (ATC R01AD)‡	3,593/9,766 (36.8)	3,607/9,766 (36.9)	1.00 (0.96–1.03)	0.836	0.99 (0.95–1.04)	0.685
All-cause mortality†	69/9,744 (0.7)	407/9,728 (4.2)	0.17 (0.13–0.22)	< 0.001	0.17 (0.14–0.22)	< 0.001
All-cause hospitalization	975/9,766 (10.0)	2,412/9,766 (24.7)	0.40 (0.38–0.43)	< 0.001	0.37 (0.35–0.40)	< 0.001
Emergency department visits	1,847/9,766 (18.9)	2,466/9,766 (25.3)	0.75 (0.71–0.79)	< 0.001	0.73 (0.69–0.77)	< 0.001
TNF inhibitor use	692/9,766 (7.1)	1,060/9,766 (10.9)	0.65 (0.60–0.72)	< 0.001	0.65 (0.59–0.71)	< 0.001
5-Year Follow-Up (n = 9,320 matched pairs)
Corticosteroid use (ATC R01AD)‡	4,574/9,320 (49.1)	5,011/9,320 (53.8)	0.91 (0.89–0.94)	< 0.001	1.05 (1.01–1.10)	0.014
All-cause mortality†	223/9,299 (2.4)	865/9,288 (9.3)	0.26 (0.22–0.30)	< 0.001	0.35 (0.30–0.40)	< 0.001
All-cause hospitalization	1,563/9,320 (16.8)	3,243/9,320 (34.8)	0.48 (0.46–0.51)	< 0.001	0.51 (0.48–0.54)	< 0.001
Emergency department visits	2,504/9,320 (26.9)	3,516/9,320 (37.7)	0.71 (0.68–0.74)	< 0.001	0.79 (0.75–0.83)	< 0.001
TNF inhibitor use	814/9,320 (8.7)	1,343/9,320 (14.4)	0.61 (0.56–0.66)	< 0.001	0.64 (0.59–0.70)	< 0.001

RR, risk ratio; HR, hazard ratio; CI, confidence interval; ED, emergency department; TNF, tumor necrosis factor.

^†Patients with a recorded death prior to the analysis time window excluded (~ 22–23 GLP-1RA users and ~ 32–43 nonusers per run).

^‡ATC R01AD encompasses corticosteroids across all routes of administration (systemic, inhaled, intranasal) in the TriNetX platform; see Limitations. The 1-year and 5-year cohorts are derived from the same base population; the 5-year cohort is a subset of the 1-year cohort. Denominator variation reflects outcome-specific data availability.

Fig. 4.

Outcome rates at 1-year and 5-year follow-up, GLP-1RA users vs. propensity score–matched nonusers. Grouped bar chart showing absolute outcome rates (%) for each group at both time horizons. Primary outcome (corticosteroid use) shown first. * P < .001 for all outcomes except corticosteroid use at 1 year (P = .836). ^‡ATC R01AD encompasses systemic and inhaled/intranasal corticosteroids (see Limitations). ED = emergency department; TNF = tumor necrosis factor inhibitor.

Fig. 5.

Forest plot of risk ratios (log scale) for all outcomes at 1-year (blue squares) and 5-year (green diamonds) follow-up. Horizontal lines represent 95% confidence intervals. Dashed vertical line = null (RR 1.0). Primary outcome (corticosteroid use) listed first. ‡ ATC R01AD outcome (see Limitations). ED = emergency department; TNF = tumor necrosis factor inhibitor.

New corticosteroid use (ATC R01AD)—primary outcome

At 1 year, corticosteroid prescription rates were similar between matched groups: 3,593 of 9,766 GLP-1RA users (36.8%) versus 3,607 of 9,766 nonusers (36.9%) (RR 1.00, 95% CI 0.96–1.03; HR 0.99, 95% CI 0.95–1.04; P = .836). At 5 years, a modest reduction was observed: 4,574 of 9,320 GLP-1RA users (49.1%) versus 5,011 of 9,320 nonusers (53.8%) (RR 0.91, 95% CI 0.89–0.94; P < .001). As detailed in the Limitations, the ATC R01AD code captures corticosteroids across all routes of administration in the TriNetX platform, including systemic, inhaled, and intranasal preparations; this outcome should therefore not be interpreted as specifically reflective of CD disease activity or systemic steroid burden.

All-cause mortality

At 1 year, mortality occurred in 69 of 9,744 GLP-1RA users (0.7%) versus 407 of 9,728 nonusers (4.2%) (RR 0.17, 95% CI 0.13–0.22; HR 0.17, 95% CI 0.14–0.22; P < .001). At 5 years, mortality occurred in 223 of 9,299 GLP-1RA users (2.4%) versus 865 of 9,288 nonusers (9.3%) (RR 0.26, 95% CI 0.22–0.30; HR 0.35, 95% CI 0.30–0.40; P < .001). The magnitude of these associations requires prominent cautious interpretation, as detailed below.

All-cause hospitalization

At 1 year, hospitalization occurred in 975 of 9,766 GLP-1RA users (10.0%) versus 2,412 of 9,766 nonusers (24.7%) (RR 0.40, 95% CI 0.38–0.43; HR 0.37, 95% CI 0.35–0.40; P < .001). Among hospitalized patients, GLP-1RA users had fewer admissions per patient (mean 7.4 vs. 11.4; P < .001). At 5 years, hospitalization occurred in 1,563 of 9,320 GLP-1RA users (16.8%) versus 3,243 of 9,320 nonusers (34.8%) (RR 0.48, 95% CI 0.46–0.51; HR 0.51, 95% CI 0.48–0.54; P < .001).

Emergency department visits

At 1 year, ED visits occurred in 1,847 of 9,766 GLP-1RA users (18.9%) versus 2,466 of 9,766 nonusers (25.3%) (RR 0.75, 95% CI 0.71–0.79; HR 0.73, 95% CI 0.69–0.77; P < .001). Among those with visits, GLP-1RA users had fewer encounters per patient (mean 0.42 vs. 0.74; P < .001). At 5 years, ED visits occurred in 2,504 of 9,320 GLP-1RA users (26.9%) versus 3,516 of 9,320 nonusers (37.7%) (RR 0.71, 95% CI 0.68–0.74; P < .001).

TNF inhibitor use

At 1 year, TNF inhibitor use occurred in 692 of 9,766 GLP-1RA users (7.1%) versus 1,060 of 9,766 nonusers (10.9%) (RR 0.65, 95% CI 0.60–0.72; P < .001). At 5 years, TNF inhibitor use occurred in 814 of 9,320 GLP-1RA users (8.7%) versus 1,343 of 9,320 nonusers (14.4%) (RR 0.61, 95% CI 0.56–0.66; P < .001). Because this outcome captures any TNF inhibitor record in the follow-up window, it cannot differentiate new initiation from continuation of prior therapy, limiting interpretation as a specific marker of treatment escalation.

E-value sensitivity analysis

E-values for all outcomes are presented in Fig. 5 and Online Resource 1 (Table S2). For the 1-year mortality association (RR 0.17), the E-value for the upper confidence limit (RR 0.22) was 8.56, indicating that an unmeasured confounder would need to be associated at least 8.56-fold with both GLP-1RA initiation and mortality—above all measured covariates—to fully explain the null. For 5-year mortality (upper CI RR 0.30), the E-value was 6.12. E-values for hospitalization at 1 year (4.27) and 5 years (3.59) exceed the moderate robustness threshold of 3.0. E-values for ED visits and TNF inhibitor use ranged from 2.03 to 2.85. While these E-values indicate mathematical robustness to a specific magnitude of unmeasured confounding, they do not exclude qualitative biases such as healthy-user bias, differential healthcare engagement, or frailty selection, which are known to operate in GLP-1RA observational studies and cannot be fully captured by E-value calculations.

Discussion

In this large, multicenter, propensity score–matched cohort study of adults with Crohn’s disease and obesity, GLP-1RA initiation was consistently associated with substantially lower health care utilization—including fewer hospitalizations and ED visits—and lower TNF inhibitor use at both 1-year and 5-year follow-up. A striking all-cause mortality difference was observed across both time horizons; however, this association requires prominent interpretive caution and should be treated as hypothesis-generating rather than reflecting a true causal effect. The primary outcome of corticosteroid prescribing was equivalent at 1 year, with a modest reduction at 5 years, though interpretation is limited by ATC coding constraints.

These findings extend prior observational data on GLP-1RAs in IBD to a clinically distinct and growing subpopulation: adults with both Crohn’s disease and comorbid obesity. Published meta-analyses have reported associations between GLP-1RA use and reduced corticosteroid exposure, hospitalization, and surgery in broader IBD cohorts^14–16,20. Our study contributes several methodologic refinements important for pharmacoepidemiologic validity: an explicit new-user design requiring washout confirmation, calendar-time risk-set sampling for index date alignment, comprehensive propensity score adjustment incorporating Crohn’s treatment intensity proxies, and transparent reporting of follow-up structure and denominator variation.

The reductions in hospitalization and ED utilization are biologically plausible. GLP-1 receptor activation in preclinical models attenuates NF-κB–mediated intestinal inflammation, promotes mucosal healing, reduces intestinal permeability, and modulates mucosal immune responses^17–19. Weight reduction and cardiometabolic risk factor improvement associated with GLP-1RA therapy may independently reduce hospitalizations in a population with high comorbidity burden. The sustained 5-year reduction in TNF inhibitor use is potentially clinically meaningful, though causality cannot be established and continuation of pre-existing therapy cannot be excluded.

Residual confounding from incomplete disease severity adjustment warrants explicit consideration for the hospitalization, ED, and TNF inhibitor associations. Despite propensity score matching, post-matching imbalances persist in CRP (SMD 0.148), albumin (SMD 0.322), and hemoglobin (SMD 0.372). Higher albumin and hemoglobin in GLP-1RA users likely reflect a combination of higher BMI (which independently elevates these values in the absence of active inflammation) and greater healthcare engagement with more complete laboratory documentation; however, they may also reflect lower baseline inflammatory burden in GLP-1RA users. Critically, endoscopic disease activity, Montreal phenotype, fecal calprotectin, and cumulative steroid burden were not captured and could not be included as covariates. If GLP-1RA users had less severe CD at baseline—because metabolically healthier patients are preferentially selected for GLP-1RA therapy, or because channeling biases clinicians away from initiating these agents in the sickest patients—this would produce spuriously lower rates of hospitalization, ED utilization, and TNF inhibitor use independent of any pharmacologic benefit. The directional bias for all three outcomes would be toward overestimating the apparent association with GLP-1RA initiation. Future studies with access to endoscopic, histologic, and biomarker-level disease activity data would be necessary to adequately control for this source of confounding.

The null 1-year corticosteroid finding warrants explicit discussion. Both matched cohorts had approximately 79%–80% baseline prevalence of prior systemic glucocorticoid use. The absence of a short-term corticosteroid-sparing association may reflect: (1) ongoing corticosteroid courses at the time of GLP-1RA prescription; (2) insufficient time for anti-inflammatory effects to manifest as reduced prescribing; (3) dilution of a true systemic steroid signal by the ATC R01AD code capturing inhaled and intranasal preparations; or (4) absence of a true effect on CD disease activity at this time horizon. The emergence of a modest but statistically significant reduction at 5 years (RR 0.91) is consistent with either a delayed effect on inflammatory disease course or cumulative weight and metabolic benefit, though selection bias and differential follow-up cannot be excluded.

The mortality associations—a 6-fold lower 1-year mortality risk and 4-fold lower 5-year mortality risk—require prominent interpretive caution and are unlikely to reflect a true causal benefit of GLP-1RA therapy alone. While E-values indicate substantial mathematical robustness to unmeasured confounding, healthy-user bias remains a well-characterized threat to validity that structured EHR data cannot fully mitigate. GLP-1RA initiators differ systematically from noninitiators in unmeasured ways, including medication adherence, functional status, frailty trajectory, and healthcare engagement. Additionally, mortality capture in the TriNetX platform relies on in-hospital death records and administrative death records from participating organizations; out-of-hospital deaths may be incompletely ascertained, potentially resulting in differential underascertainment between groups with different healthcare utilization patterns. The mortality findings should therefore be treated as exploratory and hypothesis-generating, and de-emphasized relative to the utilization outcomes. Future studies using platforms with individual-level data access should prioritize IPTW analyses and negative control outcome analyses to probe the specific contribution of healthy-user bias and differential mortality ascertainment to this association.

These findings could inform clinical practice in several ways: they provide reassurance regarding the metabolic safety of GLP-1RAs in patients with Crohn’s disease and obesity, support monitoring strategies that track both metabolic and IBD-related endpoints in patients initiating these agents, and provide a rationale for hypothesis-driven prospective studies. They do not, however, support clinical decision-making regarding GLP-1RA initiation for Crohn’s disease activity in the absence of metabolic indications, given the observational design and the potential for substantial residual confounding. A future active-comparator design—comparing GLP-1RA initiators with initiators of other anti-obesity or diabetes pharmacotherapies—would substantially strengthen causal inference by reducing confounding by indication and channeling bias.

Limitations

Several limitations merit consideration. First, Crohn’s disease was defined by a single ICD-10-CM K50.x code, without requirement for repeat diagnoses, treatment proxies, or procedural confirmation, introducing potential misclassification. Obesity was defined by E66.x or BMI ≥ 30 kg/m² recorded prior to the index date; however, completeness of BMI documentation was uneven (available in 79.7% of GLP-1RA users and 60.1% of nonusers before matching), and patients whose obesity was first documented after treatment initiation were excluded, though some misclassification cannot be fully excluded. The temporal requirement that both Crohn’s disease and obesity be documented prior to the index date is described in the Methods; however, the platform does not permit verification that obesity predated GLP-1RA initiation by a specified minimum interval.

Second, the study period (2015–2024) spans substantial variation in GLP-1RA approvals and indications across participating healthcare systems, which are globally distributed. GLP-1RA initiation in earlier calendar years likely reflects use for type 2 diabetes rather than obesity, introducing potential confounding by indication and heterogeneity in treatment effect across calendar periods. Agent-specific and calendar period–stratified analyses could not be conducted within the platform’s aggregate output interface.

Third, immunomodulator use (thiopurines, methotrexate) was not captured as a covariate or outcome, representing a potentially important source of residual confounding for corticosteroid-sparing and TNF inhibitor outcomes. Fourth, TriNetX returns aggregated analytic outputs without access to individual patient data, precluding verification of proportional hazards assumptions, analysis of time-varying exposures including GLP-1RA adherence and discontinuation, or agent-specific subgroup analyses. Outcomes were analyzed under an ITT framework without ability to assess per-protocol effects; adherence, switching, and dose changes over the extended follow-up are unaccounted for.

Fifth, the missing-indicator method used to handle absent laboratory values in the propensity score model warrants explicit consideration as a limitation. This approach assumes that the fact of missingness—rather than the missing value itself—is sufficient to account for the information absent from unmeasured observations. However, in EHR data where laboratory testing is driven by clinical indication, this assumption is unlikely to hold: patients who are sicker or have more active Crohn’s disease may be systematically more likely to have inflammatory markers measured, meaning that missing values are not missing at random (MNAR). Under informative missingness, the missing indicator approach may fail to adequately adjust for the underlying severity differences between GLP-1RA users and nonusers, resulting in residual confounding that persists even after propensity score matching. Sensitivity analyses that would more robustly address this limitation—including complete-case analyses restricted to patients with fully observed laboratory data, and multiple imputation approaches that model the missing data mechanism explicitly—were not feasible within the TriNetX platform’s aggregated output interface, which does not permit the individual-level data export required for these methods. Future studies using platforms with individual-level data access should prioritize these sensitivity analyses, particularly for the hospitalisation, ED utilisation, and mortality outcomes where inflammatory marker values are most likely to be prognostically informative. Sixth, standardized clinical disease activity measures—including endoscopic findings, Montreal phenotype, fecal calprotectin, and longitudinal weight trajectories—were unavailable. The corticosteroid outcome uses ATC R01AD, which encompasses all routes of administration including inhaled and intranasal preparations; future studies should use more specific medication codes to isolate systemic steroid use. The TNF inhibitor outcome cannot distinguish new initiation from continuation of pre-existing therapy. Seventh, residual confounding from unmeasured frailty, functional status, and health-seeking behavior remains a significant threat to validity, particularly for the mortality outcome. Eighth, mortality ascertainment relies on in-hospital and administrative death records; out-of-hospital deaths may be incompletely captured, with potential differential underascertainment between groups.

Finally, because the 1-year and 5-year cohorts are derived from the same base population and conducted as separate TriNetX platform runs, they are not independent analyses; the 5-year cohort represents a subset of the 1-year population, and direct formal comparison across time horizons should be made cautiously²¹.

Conclusions

In adults with Crohn’s disease and obesity, GLP-1RA initiation was consistently associated with lower health care utilization and lower TNF inhibitor use at both 1-year and 5-year follow-up. The observed mortality difference is hypothesis-generating and demands cautious interpretation given the inherent limitations of observational EHR-based data and the strong potential for healthy-user confounding. Corticosteroid prescribing was broadly similar at 1 year with a modest difference at 5 years; the clinical significance of this finding is constrained by ATC coding limitations. These findings support the metabolic safety of GLP-1RAs in adults with Crohn’s disease and obesity and provide a rationale for prospective randomized controlled trials to establish whether GLP-1RAs confer disease-modifying benefits beyond their established metabolic indications.

Author contributions

Conceptualization and study design: Ganju N, Ssebambulidde K, Aduli F. Data acquisition and analysis: Ganju N, Ssebambulidde K, Michael M. Data interpretation: all authors. Original draft preparation: Ganju N, Ssebambulidde K. Critical revision: Gupta S, Rao Adidam S, Michael M, Kibreab A, Aduli F. All authors reviewed and approved the final manuscript.

Data availability

Data supporting these findings are accessible within the TriNetX Global Collaborative Network to authorized users. Individual-level patient data cannot be shared publicly.

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

This study utilized exclusively deidentified patient-level data accessed through the TriNetX Global Collaborative Network, a federated electronic health record platform operating under institutional data use agreements with all participating healthcare organizations. All data are fully deidentified in compliance with the Health Insurance Portability and Accountability Act (HIPAA) Safe Harbor standard (45 CFR § 164.514(b)). This study was reviewed by the Howard University Institutional Review Board (Howard University IRB, Washington, DC, USA), which confirmed exemption from formal IRB oversight under the U.S. Common Rule (45 CFR § 46.104(d)(4)) on the grounds that the research involves exclusively pre-existing, fully deidentified data and does not constitute human subjects research requiring IRB approval. This research was conducted in accordance with the principles of the Declaration of Helsinki.

Informed consent

This study involved exclusively pre-existing, fully deidentified retrospective electronic health record data accessed through the TriNetX Global Collaborative Network under institutional data use agreements. No direct contact with human participants occurred, and no individually identifiable information was accessed at any point in the study. The Howard University Institutional Review Board (Howard University IRB, Washington, DC, USA) confirmed that the requirement for individual informed consent was waived in accordance with 45 CFR § 46.116(f), which provides for waiver of consent requirements for research involving only pre-existing deidentified data that could not practicably be carried out otherwise and that poses no more than minimal risk to participants.