GLP-1 receptor agonists at immune checkpoint inhibitor initiation with immune-related and supportive-care outcomes in patients with cancer and overweight or obesity without diabetes: a target trial emulation

Abstract

Aims

Sarcopenia, cachexia, and malnutrition are common in patients on immune checkpoint inhibitors (ICIs). GLP-1 receptor agonists (GLP-1 RAs) are increasingly used for weight management in patients with obesity, including those with cancer. Whether GLP-1 RA use at ICI initiation relates to wasting-related outcomes in patients without diabetes is unknown. We examined whether GLP-1 RA supply at ICI start was linked to wasting-related diagnoses and acute-care use.

Materials and methods

We used target-trial emulation with TriNetX US data. Adults with cancer and obesity/overweight starting an ICI were included. Baseline diabetes in the prior 12 months were excluded, with a 90-day GLP-1 RA washout. Exposure was a GLP-1 RA prescription or administration within a prespecified 30-day peri-initiation window before or after ICI initiation, versus none. Cohorts were 1:1 propensity score matched and followed up to 36 months; associations were estimated with intention-to-treat Cox models.

Results

After matching, 1974 patients were analyzed (987 per group). Over 36 months, GLP-1 RA overlap was associated with fewer immune-related adverse events (HR 0.63, 95% CI 0.50–0.79); hospitalization (HR 0.67, 0.51–0.89), ICU (HR 0.69, 0.53–0.90), and emergency department visits (HR 0.68, 0.53–0.89) were also lower.

Conclusions

GLP-1 RA add-on at ICI initiation was associated with fewer recorded wasting-related diagnoses and less acute-care use. The all-cause mortality reduction is exploratory, likely reflecting channeling bias rather than causation. These findings support prospective evaluation of GLP-1 RAs as supportive-care add-on therapy in ICI-treated patients with obesity.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s00262-026-04488-8.

Introduction

Immune checkpoint inhibitors (ICIs) are used across an expanding range of malignancies and are accompanied by immune-related adverse events (irAEs) affecting multiple organ systems [1, 2]. As ICI therapy becomes routine care, clinical priorities often extend beyond tumor control to include treatment tolerance, symptom burden, and supportive-care needs [3].

Cancer-associated wasting commonly recorded as sarcopenia, cachexia, and malnutrition remains frequent during systemic therapy and is linked to functional decline, unplanned acute-care use, and interruptions in treatment delivery [4, 5]. Importantly, obesity does not preclude wasting phenotypes [6]; people with obesity may have limited muscle reserve or develop clinically meaningful nutritional deterioration, and body mass index alone may not reflect vulnerability to these complications [7, 8]. In the immunotherapy setting, where toxicity management can involve treatment delays, corticosteroids, or hospitalization, wasting-related outcomes have direct relevance to multidisciplinary care and health-system burden [9, 10].

Glucagon-like peptide-1 receptor agonists (GLP-1 RAs) are increasingly prescribed for chronic weight management in people with obesity, including those without type 2 diabetes (T2D) [11, 12]. Evidence from clinical trials and real-world cohorts indicates substantial weight loss with favorable cardiometabolic profiles, while gastrointestinal symptoms and uncertainty regarding lean-mass trajectories remain concerns in susceptible individuals [13–16]. As a result, oncology clinics increasingly encounter ICI initiators with obesity who are already receiving GLP-1 RAs or are considered for continuation of antiobesity pharmacotherapy during treatment [17, 18].

However, evidence describing how GLP-1 RA use around ICI initiation relates to wasting-related outcomes is limited, particularly in populations without baseline T2D. To address this gap, we emulated a target trial using a large federated electronic health record network to evaluate associations between GLP-1 RA overlap at ICI initiation and subsequent wasting-related diagnoses in adults with cancer and obesity without diabetes. Sarcopenia/cachexia/malnutrition was prespecified as the primary endpoint, with acute-care utilization and selected safety outcomes examined as key secondary measures to contextualize supportive-care burden during follow-up.

Methods

Ethics

The study protocol was approved by the Institutional Review Board of Taipei Tzu Chi Hospital (14-IRB154) and conducted in accordance with the principles of the Declaration of Helsinki. Data were obtained from the TriNetX US Collaborative Network, a federated platform that integrates de-identified electronic health records from 70 healthcare organizations across the USA.

Target-trial emulation

We used a target-trial emulation framework to specify eligibility criteria, treatment strategies, time zero, follow-up, endpoints, and the analytic approach. The emulated strategies compared initiation or reinitiation of a GLP-1 RA at ICI initiation versus ICI initiation without GLP-1 RA (Table S1). To approximate a new-user design, we applied a 90-day GLP-1 RA washout defined as the absence of any GLP-1 RA prescription or administration record during the 90 days before time zero. Individuals with a prior occurrence of each endpoint were excluded from that endpoint’s risk set [19, 20]. Eligible participants were adults aged 18 years or older with malignant neoplasms and overweight or obesity who initiated an ICI between January 1, 2017, and October 31, 2025, in the TriNetX US Collaborative Network (Fig. 1). Individuals with baseline diabetes diagnoses or any antidiabetic medication exposure during the 12-month baseline period were excluded. Additional exclusions included solid organ transplantation, bariatric surgery, documented human immunodeficiency virus infection, or end-stage renal disease. Overweight or obesity was identified by an ICD-10-CM E66 diagnosis or a baseline body mass index of 27 kg/m² or higher shown in Fig. 1. We restricted the cohort to patients without diabetes to isolate the obesity (weight management) indication and avoid confounding by antidiabetic treatment, because insulin resistance attenuates GLP-1 RA-induced weight loss and would otherwise blur the exposure, and because diabetes is itself associated with cachexia-related pathways, cardiovascular risk, and acute-care use and would act as a strong prognostic confounder.

Fig. 1.

Flowchart and target-trial emulation design. Exposure was defined as a GLP-1 RA prescription or administration within a prespecified 30-day peri-initiation window before or after ICI initiation

Time zero was defined as the date of the first qualifying ICI administration after all eligibility criteria were met and the washout window was satisfied. Exposure status was assigned at time zero: The exposed cohort had a GLP-1 RA prescription or administration within a prespecified 30-day peri-initiation window before or after ICI initiation (initiation or reinitiation of GLP-1 RA as add-on to ongoing ICI), whereas comparators had no GLP-1 RA record within this window and history. This 30-day grace period was applied to account for variation in prescription timing and dispensing relative to the ICI administration date within the federated database. Because exposure was ascertained in a prespecified 30-day peri-initiation window anchored to time zero and follow-up for both cohorts began on the day after time zero, no post-baseline survival time was classified as exposed, which minimizes the potential for immortal-time (guarantee-time) bias; exposure status was not redefined by events occurring during follow-up. To confirm that the brief post-index portion of this window did not bias the results, we performed a prespecified 90-day landmark analysis restricted to patients alive and event free at the landmark, together with a per-protocol analysis, both of which were consistent with the primary intention-to-treat estimates. ICI regimens included antibodies targeting programmed cell death-1 (PD-1), programmed death-ligand 1 (PD-L1), or cytotoxic T-lymphocyte-associated protein 4 (CTLA-4), used alone or in combination. Follow-up began on the day after time zero and continued until the earliest of an outcome event, death, last recorded encounter, or administrative censoring at 36 months. The primary analysis used an intention-to-treat approach, with prespecified sensitivity analyses (Tables S1 and S2). The GLP-1 RA agents comprised semaglutide, liraglutide, dulaglutide, exenatide, and tirzepatide (a dual GIP and GLP-1 RA), identified by RxNorm ingredient codes (Supplementary code list); their distribution in the exposed cohort is summarized in Table S3; semaglutide was the most frequently prescribed agent (616 patients, 55.6 percent), followed by dulaglutide (245 patients, 22.1 percent) and liraglutide (80 patients, 7.2 percent), with the remaining agents (including exenatide and tirzepatide) accounting for 357 patients (32.2 percent); proportions sum to more than 100 percent because some patients received more than one agent during follow-up (reported for the original exposed cohort, n = 1,108).

Covariates

Baseline covariates were captured in the twelve months before the index date and entered into the propensity-score model. They comprised demographics (age at index, sex, race or ethnicity, and available socioeconomic proxies); cancer characteristics (primary site [non-small cell lung, breast, urothelial, melanoma, or other solid tumors] and recent cancer-directed surgery or radiotherapy); ICI regimen (PD-1, PD-L1, or CTLA-4 class; monotherapy versus combination); prior healthcare utilization as care-intensity and performance surrogates (admissions, emergency department visits, and critical care stays); metabolic measures (BMI category, weight, blood pressure, HbA1c or fasting glucose, and lipids); kidney function (serum creatinine and eGFR category); and concomitant medications (dose-weighted systemic corticosteroids and commonly used cardio-renal and metabolic agents) (Table S4). The distribution of cancer types by arm before matching is shown in Table S5. After 1:1 propensity-score matching, all categorical covariates were balanced (standardized mean differences below 0.10); continuous baseline body mass index and HbA1c retained residual imbalance (standardized mean differences above 0.10), reflecting the higher baseline body mass index of GLP-1 RA recipients, as addressed in the Limitations.

Outcomes

Supportive-care endpoints were prespecified as the primary focus of this analysis. The primary endpoint was a wasting-related diagnosis composite (sarcopenia, cachexia, or malnutrition) recorded within 36 months after index. Key secondary endpoints captured acute-care burden and related safety outcomes, including all-cause hospitalization, intensive care use, emergency department visits, dehydration or orthostatic hypotension, and acute kidney injury. Secondary contextual endpoints reflected immune-related toxicity and overall prognosis, including a composite immune-related adverse event outcome and prespecified subtypes (interstitial lung disease or pneumonitis, colitis or enteritis, hepatitis, and endocrinopathies), as well as all-cause mortality. Additional safety outcomes included neutropenia or agranulocytosis, peripheral neuropathy, and ocular surface toxicity. Outcomes were analyzed as time to first event using prespecified ICD-10-CM, CPT or HCPCS, and visit-type definitions (Table S6). This composite is referred to as the wasting-related composite hereafter.

Statistical analysis

The primary comparison used an intention-to-treat framework in 1:1 propensity-score-matched cohorts. Propensity scores were estimated from the baseline covariates described above, and nearest-neighbor matching was performed with a caliper on the logit of the propensity score. Time zero was the index date; follow-up began on day 1 and continued through 36 months, with censoring at 36 months or the last recorded encounter. Associations were summarized as hazard ratios with 95% confidence intervals from Cox proportional-hazards models in the matched sample. Proportional hazards were assessed using Schoenfeld residuals and log-survival plots on exported datasets. Kaplan–Meier curves and log-rank tests described between-group differences over time. For the tabulated absolute-risk summaries, crude event proportions through 36 months were used to calculate absolute-risk differences, risk-based relative risk reductions, and numbers needed to treat within endpoint-specific risk sets. These absolute-risk summaries were descriptive and were not interpreted as causal treatment effects. For prespecified key secondary and secondary endpoints, p values were adjusted using the Benjamini–Hochberg false-discovery rate procedure. E-values were calculated for the primary wasting-related endpoint (sarcopenia/cachexia/malnutrition) to quantify the minimum strength of association that an unmeasured confounder would need to have with both exposure and the outcome, beyond measured covariates, to move the estimate to the null; E-values for selected contextual endpoints, including all-cause mortality, were reported as exploratory.

Prespecified subgroup analyses assessed whether associations varied across strata, including sex, age, race, baseline body mass index, baseline eGFR below 60 versus at least 60 ml/min/1.73m², cardiovascular history, cancer type, ICI agent or class, baseline systemic prednisone equivalent use, and baseline exposure to other antiobesity pharmacotherapies. Each subgroup model included the exposure term and one multiplicative interaction term; interaction tests were interpreted as exploratory. Subgroup analyses were conducted in the original matched cohort assembled before adding prediabetes to the primary propensity-score model, so overall subgroup estimates differ modestly from the prediabetes-adjusted primary estimates, and all subgroup and interaction results are exploratory and hypothesis-generating.

Sensitivity analyses included a per-protocol definition based on prescription fill windows with grace periods and censoring at discontinuation, switching, or initiation of the alternative exposure, a 90-day landmark analysis among those alive and event free, replication in the TriNetX Global Network using the same protocol, and restriction of the exposure group to semaglutide (oral or injectable). Structured unknown categories were retained as indicator levels, partially available laboratory and vital sign covariates were modeled with an unavailable category, and no imputation was performed. Primary analyses were run within TriNetX, with de-identified aggregates exported to R and Python for diagnostics, figures, and supplementary analyses. This study adhered to the STROBE reporting standards for observational research.

Results

Cohort assembly and baseline characteristics

Of patients with cancer and obesity who initiated an ICI and met eligibility, the primary analysis, in which the propensity-score model was additionally adjusted for prediabetes, comprised 1115 GLP-1 RA recipients and 45,072 comparators, of whom 987 per group were retained after 1:1 matching (Fig. 1, Table 1, Table S7, and Fig. S1), with categorical baseline covariates balanced (standardized mean differences below 0.10), although the continuous baseline body mass index and HbA1c retained residual imbalance and are addressed in the Limitations. The prespecified subgroup, wasting-component, organ-specific immune-related, and active-comparator analyses were instead conducted in the original matched cohort, which was assembled before prediabetes was added to the primary propensity-score model and comprised 1108 GLP-1 RA recipients and 45,485 comparators before matching and 988 per group after matching; matched and unmatched cohort sizes therefore differ modestly between the primary analysis and these secondary or sensitivity analyses. Residual differences across demographics, comorbidities, prior healthcare utilization, and concomitant therapies were otherwise minimal.

Table 1.

Associations of concomitant GLP-1 receptor agonist use with mortality, healthcare utilization, immune-related adverse events during ICI therapy among individuals with neoplasms and overweight or obesity

	Exposure Cohort	Comparator Cohort	Exposure event proportion through 36 months (%)	Comparator event proportion through 36 months (%)	Risk difference (95% CI)	ARD (%)
ICI + GLP-1 RA versus ICI
All-cause mortality	36/ 927 (3.88)	90/ 907 (9.92)	3.88	9.92	− 0.060 (− 0.083, − 0.037)	− 6.04
Immune-related adverse event	120/ 392 (30.61)	190/ 404 (47.03)	30.61	47.03	− 0.164 (− 0.231, − 0.097)	− 16.42
Endocrinopathies	107/ 641 (16.69)	141/ 667 (21.14)	16.69	21.14	− 0.044 (− 0.087, − 0.002)	− 4.45
Neutropenia/agranulocytosis	39/ 896 (4.35)	66/ 885 (7.46)	4.35	7.46	− 0.031 (− 0.053, − 0.009)	− 3.10
Peripheral neuropathy	48/ 905 (5.30)	48/ 908 (5.29)	5.30	5.29	0.000 (− 0.020, 0.021)	+ 0.02
All-cause hospitalization	83/ 345 (24.06)	119/ 330 (36.06)	24.06	36.06	− 0.120 (− 0.189, − 0.051)	− 12.00
ICU/critical care use	89/ 866 (10.28)	134/ 853 (15.71)	10.28	15.71	− 0.054 (− 0.086, − 0.023)	− 5.43
Emergency department visits	96/ 529 (18.15)	139/ 520 (26.73)	18.15	26.73	− 0.086 (− 0.136, − 0.036)	− 8.58
Sarcopenia/Cachexia/Malnutrition	68/ 925 (7.35)	124/ 874 (14.19)	7.35	14.19	− 0.068 (− 0.097, − 0.040)	− 6.84
NCOs	77/ 715 (10.77)	89/ 742 (11.99)	10.77	11.99	− 0.012 (− 0.045, 0.020)	− 1.23

	NNT (3y)	Risk-based RRR (%)	HR (95% CI)	P value	Adjusted P value (BH Method)	E-value for HR	E-value for Lower CI of HR
ICI + GLP-1 RA versus ICI
All-cause mortality	17	60.86	0.41 (0.28, 0.61)	< 0.001	< 0.001	4.3	2.69
Immune-related adverse event	7	34.91	0.63 (0.50, 0.79)	< 0.001	< 0.001	2.57	1.85
Endocrinopathies	NA	NA	0.81 (0.63, 1.04)	0.094	0.105	NA	NA
Neutropenia/agranulocytosis	33	41.63	0.60 (0.40, 0.89)	0.011	0.014	2.71	1.48
Peripheral neuropathy	NA	NA	1.08 (0.73, 1.62)	0.699	0.699	NA	NA
All-cause hospitalization	9	33.28	0.67 (0.51, 0.89)	0.005	0.009	2.35	1.51
ICU/critical care use	19	34.58	0.69 (0.53, 0.90)	0.007	0.010	2.25	1.45
Emergency department visits	12	32.11	0.68 (0.53, 0.89)	0.004	0.008	2.29	1.51
Sarcopenia/Cachexia/Malnutrition	15	48.18	0.54 (0.40, 0.72)	< 0.001	< 0.001	3.14	2.13
NCOs	NA	NA	0.96 (0.71, 1.30)	0.783	NA	NA	NA

Outcomes were evaluated from day 1 through 36 months after index. Patients with a recorded occurrence of a given outcome prior to index were excluded from that outcome’s risk set. Event counts and proportions are shown for each group. Hazard ratios (HRs) with 95% confidence intervals (CIs) were estimated from Cox proportional-hazards models in the matched cohorts with robust variance clustered by healthcare organization. Proportional-hazards assumptions were reviewed using Schoenfeld residuals. Risk-based relative risk reduction (RRR) was calculated from the crude event proportions as (Fcontrol(t) minus Ftreat(t)) divided by Fcontrol(t), expressed as a percentage, and is derived from absolute event risks rather than from the hazard ratio; absolute-risk differences (ARDs) are reported as percentage points. Numbers needed to treat are computed as NNT(t) = 1/|F _treat(t) − F_control(t)∣, where F_group(t) is the crude event proportion through 36 months for that group, and are rounded upward to whole numbers; negative ARD indicates NNT for benefit, positive ARD indicates NNT for harm. F treat(t) and F control(t) are the 36-month crude event proportions shown in the exposure and comparator event-proportion columns, so the absolute-risk difference and NNT are reproducible directly from the table. Absolute-risk differences were calculated from crude event proportions among endpoint-specific risk sets. NNT is presented as an absolute-risk summary and should not be interpreted as a causal treatment effect, and is reported as not applicable when the hazard ratio is not significant after Benjamini–Hochberg adjustment or when the direction of the absolute-risk difference differs from the direction of the hazard ratio. Multiplicity across prespecified secondary endpoints was addressed using the Benjamini–Hochberg false-discovery rate procedure. Negative-control outcomes were not included in the Benjamini–Hochberg multiplicity adjustment and are reported for calibration only. E-values for primary and key secondary endpoints to contextualize potential unmeasured confounding. Denominators may vary across endpoints because of prior-event exclusions and data availability. GLP-1 RA, glucagon-like peptide-1 receptor agonist; ICI, immune checkpoint inhibitor; PSM, propensity-score matching; SMD, standardized mean difference; PD-1, programmed cell death-1; PD-L1, programmed death-ligand 1; ARD, absolute-risk difference; NNT, number needed to treat; irAE, immune-related adverse event; ICU, intensive care unit

Primary outcome: wasting-related diagnoses

Over 36 months (Table 1 and Fig. S2), GLP-1 RA overlap at ICI initiation was associated with a lower recorded incidence of the wasting-related composite (7.35% versus 14.19%; HR 0.54, 0.40–0.72). Acute-care use also differed across groups. By 36 months, all-cause hospitalization occurred in 24.06% versus 36.06% (HR 0.67, 0.51–0.89), ICU/critical care use in 10.28% versus 15.71% (HR 0.69, 0.53–0.90), and emergency department visits in 18.15% versus 26.73% (HR 0.68, 0.53–0.89). Corresponding associations at 12 and 24 months were directionally consistent. Decomposing the wasting-related composite into its components in a separate matched cohort, cachexia had a hazard ratio of 0.33 (95% CI 0.17 to 0.64) and malnutrition 0.49 (95% CI 0.37 to 0.67), whereas sarcopenia had fewer than 10 events and was suppressed under the TriNetX cell-size policy (Fig. S3). Sarcopenia, cachexia, and malnutrition are coded inconsistently and overlap clinically, so these component estimates are subject to misclassification and are reported alongside the composite.

For contextual endpoints at 36 months, the composite irAE outcome was 30.61% versus 47.03% (HR 0.63, 0.50–0.79), with a lower hazard for neutropenia/agranulocytosis (HR 0.60, 0.40–0.89), whereas endocrinopathies (HR 0.81, 0.63–1.04) and peripheral neuropathy (HR 1.08, 0.73–1.62) were not significantly different. All-cause mortality was 3.88% versus 9.92% (HR 0.41, 0.28–0.61); this large reduction is regarded as exploratory and most likely reflects channeling bias rather than a causal effect of GLP-1 RA, as detailed in the Discussion. Negative-control outcomes were not materially different (HR 0.96, 0.71–1.30).

Cumulative outcome probability curves in the matched cohort showed early and persistent separation between groups over 36 months (Fig. 2). Patients treated with ICI plus GLP-1 RAs had lower cumulative probabilities of composite irAEs (log-rank p < 0.001), endocrinopathies (p = 0.094), neutropenia or agranulocytosis (p = 0.011), peripheral neuropathy (p = 0.699), all-cause hospitalization (p = 0.005), ICU or critical care use (p = 0.007), emergency department visits (p = 0.004), and the wasting-related composite (p < 0.001), compared with patients receiving ICI alone. In contrast, probability curves for the negative-control composite largely overlapped between groups (p = 0.714).

Fig. 2.

Cumulative probabilities of wasting-related diagnoses and acute-care use outcomes in the matched cohort. Cumulative incidence curves over 36 months compare the ICI + GLP-1 RA and ICI-only groups for composite immune-related adverse events (irAEs), endocrinopathies, neutropenia or agranulocytosis, peripheral neuropathy, all-cause hospitalization, intensive or critical care use, emergency department visits, and sarcopenia, cachexia, or malnutrition. Curves represent Kaplan–Meier estimates; p values from log-rank tests assess group differences. The negative-control composite showed overlapping trajectories, supporting specificity of observed associations

In a focused organ-system re-analysis of immune-related adverse events using a separate matched cohort, HR were 0.63 (95% CI 0.50 to 0.80) for the irAE composite, 0.53 (95% CI 0.38 to 0.75) for gastrointestinal (colitis or enteritis), 0.71 (95% CI 0.56 to 0.92) for endocrine, and 1.02 (95% CI 0.58 to 1.78) for pulmonary (pneumonitis or interstitial lung disease) irAEs (Fig. S4); hepatic irAEs had fewer than 10 events and were suppressed, and this re-analysis was directionally consistent with the main-analysis composite (HR 0.63, 95% CI 0.50 to 0.79 in the primary matched cohort). Peripheral lymphocyte percentage was preserved with GLP-1 RA add-on but declined with ICI alone (21.5 versus 17.8 percent at 12 months; between-group p < 0.001), whereas eosinophil percentage did not differ (p = 0.93) (Fig. S5). The distribution of cancer types by arm before matching is reported in Supplementary Table S5(non-small cell lung cancer 27.6 versus 33.3 percent, breast cancer 22.7 versus 11.0 percent, and other cancers 59.1 versus 57.4 percent; proportions exceed 100 percent because some patients had more than one recorded cancer diagnosis). Cancer-type subgroups were examined but were limited by small strata and are reported as exploratory; tumor stage was not reliably ascertainable in the federated data and is noted as a limitation.

Safety outcomes associated with GLP-1 receptor agonists

ITT analyses showed that GLP-1 RA use during ICI therapy was not associated with an excess of prespecified safety endpoints (Table S8). Acute kidney injury occurred in 12.44% versus 20.92% (ARD 8.48% lower, NNT 12; HR 0.59, 95% CI 0.46 to 0.75; adjusted p < 0.001). Dehydration or orthostatic hypotension occurred in 17.31% versus 29.68% (ARD 12.37% lower, NNT 9; HR 0.54, 95% CI 0.43 to 0.67; adjusted p < 0.001). Additional gastrointestinal outcomes were explored descriptively but are not included in Table S8.

Kaplan–Meier survival

In an active-comparator analysis comparing ICI plus GLP-1 RA with ICI plus usual antidiabetic care (metformin or sulfonylurea) in a separate 1:1 propensity-score-matched cohort, overall survival over 36 months was higher in the ICI plus GLP-1 RA group, with the curves separating within the first year and remaining separated thereafter (Fig. 3; log-rank p = 0.003). This active-comparator comparison is a supportive sensitivity analysis only: Because the comparator group received antidiabetic therapy, it differs from the diabetes-free primary cohort, and it should not be interpreted as resolving residual confounding. In the primary comparison against ICI alone, overall survival was also higher with GLP-1 RA add-on, consistent with the all-cause mortality estimate (Fig. S6).

Fig. 3.

Kaplan–Meier curves. Active-comparator sensitivity analysis of overall survival comparing ICI plus GLP-1 RA with ICI plus usual antidiabetic care, defined as metformin or sulfonylurea, in a separate 1:1 propensity-score-matched cohort distinct from the diabetes-free primary cohort. Participants were followed until death, the last recorded encounter, or administrative censoring at 36 months. Group differences were assessed using the log-rank test. This cohort differs from the diabetes-free primary cohort and is supportive only

Subgroup analysis

For the immune-related outcomes (Fig. 4), subgroup estimates are exploratory and hypothesis-generating and were derived in the original matched cohort. For the composite immune-related adverse event outcome, heterogeneity was most apparent by age (younger than 65 years: HR 0.56, 95% CI 0.41–0.76; 65 years or older: 1.15, 0.82–1.63; p for interaction = 0.002), cancer type (breast cancer: 0.41, 0.22–0.75; non-small cell lung: 1.23, 0.74–2.04; other: 1.03, 0.76–1.39; p for interaction = 0.014), and ICI regimen (pembrolizumab: 0.61, 0.46–0.81; non-pembrolizumab: 0.96, 0.67–1.37; p for interaction = 0.050). Endocrinopathies showed heterogeneity by body mass index (30 to 35: 0.67, 0.44–1.03; greater than 35: 0.80, 0.51–1.27; p for interaction = 0.023), whereas peripheral neuropathy showed no statistically supported interaction (all p for interaction greater than 0.10). Prediabetes did not modify any immune-related outcome (p for interaction = 0.352, 0.109, and 0.540 for the immune-related adverse event composite, endocrinopathies, and peripheral neuropathy).

Fig. 4.

Subgroup analyses of immune-related outcomes. Forest plots show hazard ratios with 95% confidence intervals from Cox proportional-hazards models for the composite immune-related adverse event outcome, endocrinopathies, and peripheral neuropathy, comparing immune checkpoint inhibitor (ICI) plus concomitant GLP-1 receptor agonist (GLP-1 RA) versus ICI alone, across prespecified subgroups with tests for interaction. Markers to the left of the reference line (hazard ratio = 1) favor ICI plus GLP-1 RA. Subgroup analyses were conducted in the original matched cohort assembled before adding prediabetes to the primary propensity-score model, so overall estimates differ modestly from the prediabetes-adjusted primary estimates, and all subgroup and interaction results are exploratory and hypothesis-generating. GLP-1 RA, glucagon-like peptide-1 receptor agonist; ICI, immune checkpoint inhibitor

Across the prespecified subgroups (Fig. S7), the wasting-related association was modified by prediabetes status (p for interaction = 0.024): The estimate was pronounced in patients without prediabetes (HR 0.35, 95% CI 0.25–0.50) and attenuated to the null in those with prediabetes (HR 0.74, 95% CI 0.43–1.27). This single positive interaction among many subgroups is exploratory.

For the acute-care outcomes, prediabetes did not modify the associations (p for interaction = 0.753 for all-cause hospitalization, 0.647 for ICU or critical care use, and 0.800 for emergency department visits), with estimates directionally consistent in patients with and without prediabetes.

Sensitivity analysis

In prespecified sensitivity analyses, restriction to semaglutide and alternative design choices yielded estimates close to the primary analysis (Fig. S8). In the semaglutide cohort, ICI plus semaglutide related to lower all-cause hospitalization (HR 0.64, 95% CI 0.43 to 0.95, p = 0.026), emergency department visits (HR 0.66, 95% CI 0.45 to 0.95, p = 0.025), and sarcopenia, cachexia or malnutrition (HR 0.30, 95% CI 0.20 to 0.47, p < 0.001). With a 12-month washout (Fig. S9), all-cause mortality (HR 0.26, 95% CI 0.15 to 0.44, p < 0.001) and sarcopenia or cachexia (HR 0.44, 95% CI 0.32 to 0.60, p < 0.001) showed similar patterns.

Landmark analysis excluding the first three months of follow-up produced comparable estimates for composite irAEs (HR 0.75, 95% CI 0.56 to 0.99, p = 0.040; Fig. S10 and sarcopenia or cachexia (HR 0.62, 95% CI 0.44 to 0.89, p = 0.008). In the global network analysis (Fig. S11, all-cause mortality (HR 0.35, 95% CI 0.21 to 0.58, p < 0.001) and sarcopenia or cachexia (HR 0.47, 95% CI 0.35 to 0.62, p < 0.001) again aligned with these findings.

HbA1c, BMI, and body weight

Follow-up measures of body weight, BMI, and HbA1c were examined in the matched cohort (Fig. S12). Over follow-up, both groups showed modest decreases. By 12 months, mean body weight was 96.1% of baseline in the ICI plus GLP-1 RA group and 96.4% in the ICI-only group, and mean BMI was 96.3% and 97.2% of baseline, respectively (p < 0.001). Absolute body-weight change is reported in kilograms, the standard metric in obesity trials. Mean body weight decreased by 1.3, 2.8, and 3.7 kg with GLP-1 RA add-on and by 2.6, 3.3, and 4.4 kg with ICI alone at 3, 6, and 12 months, indicating a comparable and modest loss in both groups. The absolute between-group difference in mean body weight at each time point (4.8, 4.0, and 4.2 kg; p = 0.007, 0.024, and 0.018, testing the between-group comparison rather than change from baseline) primarily reflects the slightly higher baseline weight of the exposed group rather than greater weight loss. Dose titration to the maximally tolerated level could not be verified, because the database does not reliably record dose changes.

Discussion

Concomitant use of GLP-1 RA during ICI therapy in people with cancer and obesity without T2D was associated with a lower risk of wasting-related diagnoses and acute-care utilization compared with ICI treatment alone.

The observed pattern aligns with accumulating evidence that GLP-1 RA confer cardiometabolic and renal benefits in people with obesity, including those without diabetes, and may be associated with improved long-term safety profiles in large-scale real-world cohorts [11, 21]. Earlier work in populations with obesity but without cancer described favorable associations between GLP-1 RA exposure and composite cardiometabolic outcomes, with low rates of classic safety concerns such as pancreatitis or gallbladder disease [22, 23]. Other observational studies involving patients with cancer receiving ICIs have suggested that GLP-1 RA use may relate to lower cardiovascular event rates or improved clinical outcomes, although these reports have usually pooled individuals with and without diabetes and have provided limited detail on immune-related adverse event subtypes or on non-cardiovascular safety endpoints [24–26]. Within this context, the present analysis extends the literature by focusing on adults with obesity and without T2D, by incorporating a target-trial emulation framework with explicit washout periods and exclusion of prior endpoint events. A companion target-trial emulation in patients with cancer and type 2 diabetes reported the same direction of association, with lower all-cause mortality, hospitalization, and immune-related adverse events during GLP-1 RA use, alongside higher diabetic retinopathy progression and non-arteritic anterior ischemic optic neuropathy, ophthalmic signals specific to the diabetes setting that warrant monitoring [18]. We excluded diabetes here to isolate the obesity indication, and whether the wasting-related associations differ by glycemic status is best addressed in a dedicated cohort.

Several pathways may plausibly contribute to the associations between GLP-1 RA add-on and the lower risks of immune-related events and healthcare utilization observed in this cohort. The all-cause mortality finding is discussed separately below in the context of residual confounding. GLP-1 RA induces sustained weight loss and improves insulin sensitivity, glycemic indices, blood pressure, and lipid profiles in people with obesity, changes that could translate into better tolerance of cancer therapies and fewer decompensations requiring hospitalization [27, 28]. Weight loss and improved metabolic control may also mitigate frailty, preserve functional status, and interact with the risk of sarcopenia or cachexia, which was notably lower in GLP-1 RA users in this analysis [29, 30]. Renal and hemodynamic benefits described in T2D may help explain the lower rates of acute kidney injury and dehydration or orthostatic hypotension, though confounding by indication and differential clinical monitoring cannot be excluded [31]. These mechanisms are hypotheses: No immune profiling, cytokine, or tumor-response data were available in this dataset to confirm them, and the proposed inflammatory, metabolic, immune-modulatory, and body-composition pathways require prospective and translational testing.

Beyond systemic metabolic effects, preclinical and translational data support the possibility that GLP-1 pathway modulation directly or indirectly reshapes the tumor microenvironment during immune checkpoint blockade. In murine tumor models, GLP-1 RA and related incretin agonists have been associated with increased intratumoral CD8 positive T cell infiltration, altered expression of exhaustion or activation markers, reduced frequencies of FoxP3 positive regulatory T cells, and enrichment of dendritic cells and antigen presenting macrophage subsets [32]. In some settings, treatment has been linked to reductions in myeloid derived suppressor cells, shifts in macrophage polarization, and attenuation of neutrophil extracellular trap formation, together with evidence of more durable antitumor immune responses when combined with PD-1 blockade [33, 34]. At the level of tumor stroma, GLP-1 pathway activation has been reported to dampen stromal activation and extracellular matrix deposition in pancreatic models, changes that may ease T cell access to tumor nests [35]. These findings are preclinical and correlative. They suggest, but do not establish, that GLP-1 RA could support effective ICI responses without increasing immune-related toxicity; we present these mechanisms as hypotheses only.

The subgroup analyses suggested some heterogeneity in associations across age, race, body mass index categories, obstructive sleep apnea status, cancer type, and ICI regimen. Lower composite irAE and intensive care use rates associated with GLP-1 RA use were more apparent in people younger than 65 years and in non-white patients, and endocrinopathy patterns differed between moderate and more severe obesity categories. Signals by cancer type were most pronounced in breast cancer, with less clear differences in non-small cell lung cancer and other tumors, and there were indications that pembrolizumab treated patients might derive greater benefit for certain immune-related and utilization outcomes than those receiving other regimens. However, they should be interpreted cautiously considering multiple testing, limited sample sizes within strata, and the exploratory nature of the interaction analyses. All subgroup and interaction analyses are exploratory and hypothesis-generating; the Benjamini–Hochberg procedure controlled the false-discovery rate across prespecified secondary endpoints but not across the full subgroup family.

Selected prespecified safety outcomes were explored descriptively. Event counts for several classic GLP-1 RA safety concerns, including pancreatitis, gallbladder disease, non-arteritic anterior ischemic optic neuropathy, and suicidal ideation or attempt, were low and estimates were imprecise; therefore, these analyses should be interpreted as descriptive rather than definitive safety reassurance.

Several limitations should temper interpretation. The all-cause mortality result is large, exceeds the benefit of most active cancer treatments, and has no established biological explanation. Channeling bias is the most likely explanation: Adding GLP-1 RA to ICI reflects better baseline health and more active clinical management, and key confounders including ECOG status, tumor stage, and prior treatment lines were unavailable in the propensity model. We computed E-values to gauge how strong an unmeasured confounder would need to be to explain the associations; because channeling here reflects several correlated prognostic factors acting together, E-values alone cannot exclude residual bias in the mortality estimate. Several falsification tests bound this concern further. Prespecified negative-control outcomes, chosen for having no plausible relationship to GLP-1 RA exposure, showed no association, arguing against a pervasive non-specific healthy-user artifact. Both groups lost a modest and comparable amount of body weight, and the exposed group did not lose more weight despite receiving a weight-lowering agent, so the lower wasting and acute-care burden is unlikely to be attributable to weight loss. GLP-1 RA recipients also had a higher baseline body mass index, and this measure together with HbA1c retained minor residual imbalance after matching even though categorical strata were balanced; that the heavier exposed group nonetheless lost less weight yet had fewer wasting diagnoses argues further against weight loss as the driver. As a further check on channeling, we compared GLP-1 RA against usual antidiabetic care (metformin or sulfonylurea) in a separate active-comparator cohort, in which the survival difference persisted, consistent with the primary analysis; this comparator nonetheless involves patients receiving antidiabetic therapy and therefore differs from the diabetes-free primary cohort, and residual confounding cannot be excluded. The analysis was observational and based on routinely collected electronic health records, so residual confounding, channeling bias, and misclassification are inevitable despite target-trial emulation, careful eligibility definitions, and propensity-score matching. Important clinical features that influence both GLP-1 RA prescribing and ICI outcomes, such as tumor stage, burden of metastatic disease, prior lines of therapy, performance status, patient reported symptoms, and detailed nutritional status, were incompletely captured or unavailable. Obesity was defined by diagnostic codes and baseline body mass index, without direct information on body composition or subsequent weight trajectories, and irAEs were identified using administrative codes rather than standardized adjudication or grading. Death was based on demographic flags and ICD codes, with limited detail on cause of death and potential under ascertainment. The TriNetX platform aggregates data from multiple healthcare organizations, each with its own coding practices and missingness patterns, and laboratory and vital sign data were handled using categorical variables with indicator levels for missing values rather than formal imputation.

Future work could extend these observations through prospective clinical cohorts and, where feasible, randomized studies comparing structured GLP-1 RA approaches with alternative weight management strategies or usual care in patients with obesity receiving ICIs.

Conclusion

In ICI patients with obesity and no baseline diabetes, add-on GLP-1 RA was associated with fewer wasting diagnoses, less acute care, and fewer irAEs over 36 months. All-cause mortality most likely reflects channeling bias and is exploratory; residual confounding cannot be excluded.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Yu-Jung Lin and Pei-Yun Li conceived the study and contributed to the target-trial emulation design. Pei-Yun Li and Shao-Chia Chen performed data curation, query implementation, and validation. Yu-Jung Lin led the statistical analyses, sensitivity analyses, and visualization. Gideon Meyerowitz-Katz provided methodological oversight and contributed to interpretation and critical revision of the manuscript. Yu-Nan Huang and Pen-Hua Su provided clinical and scientific supervision, contributed to study resources and interpretation, and co-led manuscript development and revision. All authors reviewed and approved the final manuscript and agree to be accountable for all aspects of the work. Yu-Nan Huang and Pen-Hua Su contributed equally as joint senior authors.

Funding

Research support was provided through the National Science and Technology Council, Taiwan (NSTC 113-2314-B-040-026-MY2 and NSTC 114-2622-B-040-001), Taipei Tzu Chi Hospital, Buddhist Tzu Chi Medical Foundation (TCRD-TPE-115-01) and the Chung Shan Medical University Hospital (CSH-2026-A-009, CSH-2026-C-030 and CSH-2026-F-003). Funders had no role in study design, data collection, data analysis, interpretation, or the writing of this report. The corresponding author (Y.N.H. and P.H.S) retained full access to the data and final responsibility for the decision to submit.

Data availability

Consistent access to comparable data is available to qualified investigators via the TriNetX portal (https://www.trinetx.com). Study protocol materials, variable definitions, and analysis code are available upon reasonable academic request to the corresponding authors, contingent on IRB authorization at Taipei Tzu Chi Hospital and adherence to privacy protections. Propensity-score specifications, outcome definitions, and sensitivity-analysis plans are provided in the Supplementary Materials to facilitate reproducibility while maintaining appropriate oversight.

Declarations

Conflict of interest

Yu-Jung Lin, Pei-Yun Li, Shao-Chia Chen, Gideon Meyerowitz-Katz, Yu-Nan Huang, Pen-Hua Su declare no competing interests. The funding bodies had no role in the study design, data collection, analysis, interpretation, or manuscript preparation. The authors affirm independence in the conduct of this research and in the decision to publish.

Ethical approval and consent to participate

This cohort study used the TriNetX US Collaborative Network, a federated research platform that integrates electronic health record (EHR) data from participating healthcare organizations. Records accessed through TriNetX are de-identified in accordance with the HIPAA Privacy Rule, with de-identification attested by expert determination. Direct patient identifiers are not available, and individual-level chart review is not supported within the platform. The protocol was approved by the Institutional Review Board of Taipei Tzu Chi Hospital (14-IRB154) and aligned with the principles of the Declaration of Helsinki. The analytic dataset was derived from the TriNetX US Collaborative Network, which comprises de-identified EHR data contributed by approximately 70 healthcare organizations in the USA.