Cardiovascular outcome of glucagon-like peptide-1 receptor agonists vs dipeptidyl peptidase-4 inhibitor on end-stage kidney disease patients with heart failure: an emulated target trial in patients with diabetes Authors: Po-Yi Liu, Chao-Kuei Shih, Miyuki Hsing-Chun Hsieh, Ching-Chun Lin, Chieh-Yen Liu, Edward Chia-Cheng Lai, Chung-Yi Li, Chih-Hen Yu, Junne-Ming Sung Published: 2026 ABSTRACT AbstractPatients with end-stage kidney disease (ESKD) receiving maintenance dialysis have a high burden of heart failure (HF), with cardiovascular disease the leading cause of death, yet are largely excluded from randomized trials. Using the TriNetX federated electronic health record network, we emulate a target trial comparing initiation of glucagon-like peptide-1 receptor agonists (GLP-1RAs) versus dipeptidyl peptidase-4 inhibitors (DPP-4is) in diabetic ESKD patients with HF, undergoing maintenance dialysis. We identify a primary ESKD-HF population and a confirmatory dialysis-dependence-coded HF subset, and apply a new-user, active-comparator design and use propensity score matching to balance baseline characteristics. In the primary ESKD-HF population, GLP-1RAs use is associated with a lower risk of the primary composite ischemic cardiovascular events plus HF exacerbations, than DPP-4i initiation (31.7% vs. 41.4%; HR 0.72[0.64–0.82], P < 0.0001), with concordance reductions in ischemic events (HR 0.74), HF exacerbations (HR 0.76), all-cause mortality (HR 0.68), and a death-inclusive composite (HR 0.72). Results are consistent in multivariable models, across prespecified subgroups, and in extensive sensitivity analyses, and are corroborated in the confirmatory dialysis-dependence–coded HF subset HR (0.71[0.60–0.85]). These findings provide real-world evidence suggesting GLP-1RAs may improve cardiovascular outcomes in dialysis-dependent diabetic ESKD with HF and support prospective randomized evaluation.Subject terms: Hypertension, End-stage renal disease, Diabetes Introduction End-stage kidney disease (ESKD) refers to kidney failure within the spectrum of chronic kidney diseases (CKD) requiring kidney replacement therapy (KRT), whereas ‘maintenance dialysis’ denotes dialysis treatment status1. Globally, approximately 4.6 million individuals are living with ESKD2. More than one-third of ESKD patients undergoing maintenance dialysis develop heart failure (HF)3–5. The presence of HF in ESKD dramatically worsens outcomes6,7, with cardiovascular events, including HF exacerbations, accounting for nearly half of all deaths in this population8,9. Despite this substantial burden and poor prognosis, few therapeutic interventions have convincingly reduced cardiovascular events in this vulnerable group10–12. Evidence remains scarce because randomized clinical trials of foundational guideline-directed medical therapies (GDMT) of HF, including trials testing β-blockers, renin–angiotensin system inhibitors, angiotensin receptor–neprilysin inhibitors, mineralocorticoid receptor antagonists, and sodium–glucose cotransporter-2 inhibitors, have systematically excluded patients with ESKD13. As a result, current guideline recommendations offer little guidance on disease-modifying pharmacotherapies for HF in this population13–15. Moreover, observational data consistently show that patients with HF in dialysis are less likely to receive GDMT and experience worse outcomes than their non-ESKD counterparts16, underscoring the urgent need to identify effective and broadly applicable therapies in this high-risk setting17. Diabetes mellitus (DM) is the underlying cause in nearly half of ESKD cases, and it is associated with higher all-cause and cardiovascular mortality among patients requiring dialysis therapy compared with non-diabetic counterparts18,19. In this context, glucagon-like peptide-1 receptor agonists (GLP-1RAs) have emerged as promising agents with cardiovascular benefits beyond glycemic control. While earlier cardiovascular outcome trials (CVOTs) established their efficacy in reducing major adverse cardiovascular events (MACE) in high-risk patients with type 2 diabetes (T2D)20, more recent evidence suggests that their clinical relevance may extend to heart failure as well21. GLP-1RAs significantly reduced HF worsening events and the composite of cardiovascular death or HF hospitalization22. The FLOW trial, specifically in the chronic kidney disease population, demonstrated that semaglutide significantly lowered the risk of heart failure (HF) events and cardiovascular death23. Given the shared risk factors and pathophysiological continuum between CKD and ESKD, these findings raise the possibility that GLP-1RAs could provide meaningful cardiovascular benefit in patients with HF in diabetic ESKD. Prior observational evidence, such as a nationwide retrospective cohort study in Taiwan, has suggested benefit for reducing mortality of GLP-1RAs in diabetic advanced CKD stage 5 or ESKD compared with cardiovascular-neutral agents like Dipeptidyl Peptidase-4 Inhibitor (DPP-4is)24. Whether GLP-1RA agents can also reduce cardiovascular events, including ischemic events and HF exacerbations, that ultimately drive mortality in diabetics dialysis-dependent ESKD with HF is largely unknown. In this work, we emulate a target trial using the global TriNetX research network to evaluate the comparative effectiveness of GLP-1RAs versus DPP-4is in patients with HF in diabetic ESKD24. Our target population is the high-risk clinical setting of HF in ESKD, a setting that in routine practice predominantly occurs in patients treated with maintenance dialysis4,25,26. ICD-10-CM N18.6 is a widely used and broadly available code of clinically documented ESKD across health systems, whereas dialysis-dependence coding (Z99.2) may be variably captured across encounters and datasets in real-world electronic health record (EHR)27–29. Accordingly, to balance representativeness with dialysis-dependent specificity, we use N18.6 (excluding kidney transplant status [Z94.0]) to define the ESKD population and corroborate findings in a confirmatory subset requiring both N18.6 and Z99.2 (the confirmatory dialysis-dependence–coded subset). This design enables us, in a real-world setting, to broaden the representativeness of the study population and directly test the hypothesis that GLP-1RAs may confer cardiovascular benefit in this uniquely high-risk and underserved group. Together, this approach provides hypothesis-generating real-world evidence that GLP-1RA initiation is associated with improved cardiovascular outcomes in diabetic dialysis-dependent ESKD with HF. Results Patient selection and demographic characteristics Between August 1, 2016, and Feb 1, 2026, 195,436,691 individuals were identified in TriNetX. We identified 389,265 adults with T2D and ESKD; among these, 193,921 also had a heart failure diagnosis. Using ICD-10-CM N18.6 (excluding kidney transplant status [Z94.0]) to define ESKD for the primary analysis, and after applying inclusion and exclusion criteria (new-user design and timing restrictions), 5360 patients comprised the eligible ESKD–HF new-user population, including 1766 GLP-1RAs initiators and 3594 DPP-4i initiators. This eligible population corresponded to the primary ESKD–HF population (I50 and N18.6, excluding transplant status [Z94.0]) for the primary analysis. Following 1:1 propensity score matching (PSM), 1388 pairs (n = 2776) comprised the analytic cohort for the target emulated trial (Fig. 1, Supplementary Table 1). Flowchart of study population selection. After identifying patients with both type 2 diabetes mellitus (T2DM) and end-stage kidney disease (ESKD), we enrolled those who also had a diagnosis of heart failure (HF), including prevalent HF (HF diagnosed before ESKD) and incident HF (HF diagnosed after ESKD) (Supplementary Fig. 1). We then applied a prespecified new-user, target trial–eligible design: patients were required to initiate a study medication after having both (i) a recorded HF diagnosis and (ii) a recorded ESKD diagnosis, with initiation occurring within 2 years of each diagnosis, yielding the eligible ESKD–HF new-user population. The primary analysis population (Primary ESKD–HF population) was defined by ICD-10-CM I50 and N18.6 (excluding kidney transplant status [Z94.0]). We further evaluated a prespecified higher-specificity nested confirmatory dialysis-dependence–coded subset requiring Z99.2 in addition to I50 and N18.6 (I50 and N18.6 + Z99.2); 1:1 propensity score matching was repeated within this subset. ESKD end-stage kidney disease, HF heart failure, HCOs healthcare organizations, ICD-10-CM International Classification of Diseases, Tenth Revision, Clinical Modification, T2D type 2 diabetes. In the eligible ESKD–HF new-user population (n = 5360), patients initiated the study medications after receiving both a HF diagnosis within the previous 2 years and an ESKD diagnosis within the previous 2 years (Supplementary Fig. 1). This population was confirmed by total coverage of first incident HF (ICD10CM:I50) and first incident ESKD diagnosis (ICD10CM:N18.6) (Table 1, Supplementary Fig. 1 and Supplementary Table 2). We quantified dialysis certainty in these ESKD-HF patients using a prespecified composite definition (vascular access procedures, dialysis procedure codes, dialysis dependence [Z99.2], or peritoneal dialysis). Evidence consistent with maintenance dialysis was observed in 81.3% of GLP-1RA users and 80.0% of DPP-4i users in the cohort (Supplementary Table 3). Among those without composite evidence, persistent ESKD coding (≥2 N18.6 diagnoses ≥90 days apart or ≥4 distinct encounters) was present in an additional 10.4% and 13.7%, respectively, yielding >90% with either composite dialysis evidence or persistent ESKD coding (Supplementary Table 3). Table 1Baseline characteristics of GLP-1 RAs and DPP4i groups before and after propensity score matchingVariablesaBefore matchingAfter matchingGLP-1 RA groupb(n = 1766)DPP4i group(n = 3594)StandardizeddifferenceGLP-1 RA group(n = 1388)DPP4i group(n = 1388)StandardizeddifferenceAge at index, years Mean (SD)60.9 (11.8)66.3 (12.3)0.44962.3 (11.3)62.4 (12.2)0.004Sex, n (%) Female781 (44.2)1550 (43.1)0.022616 (44.4)621 (44.7)0.007 Male983 (55.7)2043 (56.8)0.024772 (55.6)766 (55.2)0.009Race, n (%) White878 (49.7)1270 (35.3)0.294655 (47.2)687 (49.5)0.046 Black or African American533 (30.2)817 (22.7)0.169409 (29.5)406 (29.3)0.005 Asian78 (4.4)439 (12.2)0.28576 (5.5)66 (4.8)0.033 Other race120 (6.8)176 (4.9)0.081102 (7.3)99 (7.1)0.008 Unknown race100 (5.7)798 (22.2)0.49298 (7.1)85 (6.1)0.038Diagnosis of heart failure, n (%)c Heart failure1766 (100)3594 (100)1388 (100)1388 (100) Diastolic heart failure743 (42.1)1143 (31.8)0.214541 (39)545 (39.3)0.006 Systolic heart failure515 (29.2)1038 (28.9)0.006384 (27.7)390 (28.1)0.01 Combined systolic and diastolic heart failure277 (15.7)478 (13.3)0.068201 (14.5)196 (14.1)0.01 Heart failure, unspecified1034 (58.6)1995 (55.5)0.061772 (55.6)752 (54.2)0.029 Other heart failure121 (6.9)137 (3.8)0.13673 (5.3)72 (5.2)0.003 Left ventricular failure42 (2.4)108 (3)0.03935 (2.5)31 (2.2)0.019Dialysis profile, n (%)d End-stage kidney disease1766 (100)3594 (100)1388 (100)1388 (100) Peritoneal dialysis19 (1.1)39 (1.1)0.00113 (0.9)13 (0.9)<0.001 Dialysis procedure other than hemodialysis81 (4.6)118 (3.3)0.06756 (4)61 (4.4)0.018Types of hemodialysis access, n (%)d Arteriovenous fistula170 (9.6)239 (6.7)0.109119 (8.6)122 (8.8)0.008 Arteriovenous graft46 (2.6)91 (2.5)0.00538 (2.7)35 (2.5)0.014 Tunneled dialysis catheter276 (15.6)460 (12.8)0.081202 (14.6)207 (14.9)0.01Dialysis-related complications, n (%) Intradialytic hypotension58 (3.3)99 (2.8)0.03137 (2.7)38 (2.7)0.004Cardiovascular disease, n (%) Ischemic heart diseases966 (54.7)1878 (52.3)0.049722 (52)718 (51.7)0.006 Hypertensive heart disease395 (22.4)559 (15.6)0.174266 (19.2)250 (18)0.03 Cardiomyopathy304 (17.2)516 (14.4)0.078227 (16.4)219 (15.8)0.016 Chronic rheumatic heart diseases230 (13)495 (13.8)0.022180 (13)184 (13.3)0.009 Atrial fibrillation339 (19.2)611 (17)0.057249 (17.9)255 (18.4)0.011 Cerebrovascular diseases308 (17.4)583 (16.2)0.033237 (17.1)232 (16.7)0.01 Peripheral vascular disease301 (17)574 (16)0.029233 (16.8)238 (17.1)0.01 Alcohol-related disorders71 (4)77 (2.1)0.10943 (3.1)47 (3.4)0.016 Nicotine dependence130 (7.4)193 (5.4)0.08290 (6.5)98 (7.1)0.023 Hypertension1296 (73.4)2225 (61.9)0.247956 (68.9)954 (68.7)0.003 Dyslipidemia1082 (61.3)1771 (49.3)0.243802 (57.8)780 (56.2)0.032 Overweight and obesity914 (51.8)920 (25.6)0.558601 (43.3)604 (43.5)0.004Comorbidities, n (%) Chronic obstructive pulmonary diseases311 (17.6)548 (15.2)0.064244 (17.6)243 (17.5)0.002 Neoplasmse414 (23.4)646 (18)0.135299 (21.5)291 (21)0.014 Liver cirrhosis90 (5.1)166 (4.6)0.02270 (5)75 (5.4)0.016 Systemic lupus erythematosus14 (0.8)25 (0.7)0.01110 (0.7)10 (0.7)<0.001 Glomerular diseases144 (8.2)254 (7.1)0.041105 (7.6)98 (7.1)0.019Socioeconomic status indicators, n (%) Problems related to housing and economic circumstances68 (3.9)78 (2.2)0.09846 (3.3)40 (2.9)0.025 Problems related to the primary support group28 (1.6)32 (0.9)0.06316 (1.2)14 (1)0.014 Problems related to the social environment20 (1.1)41 (1.1)0.00113 (0.9)15 (1.1)0.014Systolic blood pressure, mmHg, Mean (SD)134.8 (22.6)136.5 (25.7)0.072135.2 (22.8)136.7 (26)0.059BMI, kg/m², mean (SD)35.1 (8.5)29.3 (7.1)33.6 (7.8)32.5 (7.7) ≤18.5, n (%)82 (4.6)177 (4.9)0.01360 (4.3)58 (4.2)0.007 18.5–25, n (%)258 (14.6)887 (24.7)0.256231 (16.6)224 (16.1)0.014 25–30, n (%)486 (27.5)1169 (32.5)0.109418 (30.1)416 (30)0.003 30–35, n (%)642 (36.4)912 (25.4)0.239490 (35.3)486 (35)0.006 35–40, n (%)578 (32.7)549 (15.3)0.417387 (27.9)371 (26.7)0.026 ≥40, n (%)464 (26.3)358 (10)0.433256 (18.4)279 (20.1)0.042Laboratory examination Sodium, mmol/L, mean (SD)137.5 (3.7)136.5 (4.1)0.253137.4 (3.8)136.8 (3.9)0.169 Potassium, mmol/L, mean (SD)4.4 (0.6)4.3 (0.7)0.0574.4 (0.6)4.4 (0.7)0.014 Calcium, mmol/L, mean (SD)9 (0.8)8.7 (0.9)0.3128.9 (0.9)8.8 (0.9)0.186 Phosphate, mmol/L, mean (SD)4.5 (1.5)4.6 (1.6)0.074.4 (1.5)4.6 (1.6)0.09 BUN, mg/dL, mean (SD)43.7 (22.5)49.6 (27.9)0.23544.4 (22.3)47.4 (25)0.125 Hemoglobin, g/dL, mean (SD)10.9 (2)10 (1.8)0.47310.7 (2)10.2 (1.9)0.277 Iron, µg/dL, mean (SD)58.5 (33.2)56.2 (35.5)0.06759.1 (34.8)56.6 (35.9)0.072 Transferrin, mg/dL, mean (SD)193.6 (55.1)178.1 (56.2)0.279190.5 (56.1)184.1 (57.7)0.113 Iron binding capacity, μg/dL, mean (SD)633.5 (3381.7)740.5 (3607.7)0.031490.9 (2575.9)737.9 (3700.3)0.077 Parathyroid hormone, pg/mL, mean (SD)278.1 (338.3)283.4 (279.3)0.017277.5 (323.2)274.1 (257.5)0.012 Urine output, mL, mean (SD)167 (214)154.3 (296.5)0.049160.1 (206.4)110.6 (151.5)0.274Albumin, g/dL, mean (SD)3.6 (0.6)3.3 (0.7)3.5 (0.7)3.4 (0.7) 0–2.5, n (%)394 (22.3)800 (22.3)0.001306 (22)300 (21.6)0.01 2.5–3, n (%)700 (39.6)1402 (39)0.013537 (38.7)547 (39.4)0.015 3–3.5, n (%)949 (53.7)1777 (49.4)0.086703 (50.6)707 (50.9)0.006 3.5–4, n (%)1051 (59.5)1749 (48.7)0.219758 (54.6)748 (53.9)0.014 ≥4, n (%)815 (46.1)1163 (32.4)0.285556 (40.1)554 (39.9)0.003Ferritin, μg/dL, mean (SD)491 (744.8)796.3 (5628)490.5 (674.6)601.2 (1563.3) ≤100, n (%)280 (15.9)427 (11.9)0.115206 (14.8)209 (15.1)0.006 100–200, n (%)289 (16.4)451 (12.5)0.109207 (14.9)214 (15.4)0.014 ≥200, n (%)547 (31)1014 (28.2)0.06399 (28.7)404 (29.1)0.008Hemoglobin A1c, %, mean (SD)7.8 (2.1)7.3 (1.8)7.8 (2.1)7.5 (1.9) 5–6.5, n (%)606 (34.3)1151 (32)0.049448 (32.3)449 (32.3)0.002 6.5–7.5, n (%)618 (35)1110 (30.9)0.088467 (33.6)486 (35)0.029 7.5–12, n (%)782 (44.3)1129 (31.4)0.268555 (40)549 (39.6)0.009 ≥12, n (%)151 (8.6)153 (4.3)0.17692 (6.6)93 (6.7)0.003Cholesterol in LDL, mg/dL, mean (SD)79.7 (43)80.3 (44.1)78.4 (41.7)79.3 (45.1) ≤55, n (%)412 (23.3)641 (17.8)0.136304 (21.9)291 (21)0.023 55–70, n (%)318 (18)532 (14.8)0.087234 (16.9)230 (16.6)0.008 70–100, n (%)455 (25.8)712 (19.8)0.142304 (21.9)324 (23.3)0.034 100–130, n (%)277 (15.7)459 (12.8)0.083191 (13.8)190 (13.7)0.002 130–160, n (%)139 (7.9)207 (5.8)0.08496 (6.9)86 (6.2)0.029 ≥160, n (%)116 (6.6)177 (4.9)0.07183 (6)79 (5.7)0.012Natriuretic peptide B, pg/mL, mean (SD)1042.3 (3564.8)1924.6 (4793.6)1253.1 (4089.7)1193.6 (2065.3) ≤100, n (%)185 (10.5)186 (5.2)0.198119 (8.6)117 (8.4)0.005 100–300, n (%)256 (14.5)310 (8.6)0.184172 (12.4)173 (12.5)0.002 300–600, n (%)194 (11)312 (8.7)0.077137 (9.9)141 (10.2)0.01 ≥600, n (%)281 (15.9)619 (17.2)0.035227 (16.4)239 (17.2)0.023Natriuretic peptide B prohormone N-Terminal, pg/mL, mean (SD)8069.2 (12,684.4)15,610.8 (17,576.4)9549.6 (13,896.5)11,283.8 (13,519.2) ≤400, n (%)97 (5.5)70 (1.9)0.18853 (3.8)53 (3.8)<0.001 400–800, n (%)111 (6.3)69 (1.9)0.22153 (3.8)58 (4.2)0.018 800–1200, n (%)72 (4.1)67 (1.9)0.13142 (3)46 (3.3)0.016 1200–70,000, n (%)339 (19.2)689 (19.2)0.001262 (18.9)256 (18.4)0.011 ≥70,000, n (%)10 (0.6)24 (0.7)0.01310 (0.7)10 (0.7)<0.001LVEF, %, mean (SD)f51.7 (15.2)52.6 (14.7)52.1 (15.2)51.8 (15.2) ≤40, n (%)59 (3.3)85 (2.4)0.05937 (2.7)41 (3)0.017 41–49, n (%)53 (3)92 (2.6)0.02738 (2.7)39 (2.8)0.004 ≥50, n (%)161 (9.1)289 (8)0.038116 (8.4)122 (8.8)0.015Heart failure guideline-directed medical therapy, n (%) Beta blockers1277 (72.3)2322 (64.6)0.166959 (69.1)964 (69.5)0.008 ACEi464 (26.3)659 (18.3)0.192337 (24.3)334 (24.1)0.005 ARB/ARNI567 (32.1)1028 (28.6)0.076401 (28.9)372 (26.8)0.047 MRAg,h251 (14.2)341 (9.5)0.147158 (11.4)170 (12.2)0.027 SGLT2i214 (12.1)149 (4.1)0.295107 (7.7)99 (7.1)0.022Diuretics, n (%)h Loop diureticsh1162 (65.8)1936 (53.9)0.245847 (61)842 (60.7)0.007 Thiazideh545 (30.9)688 (19.1)0.273368 (26.5)361 (26)0.011Cardiovascular agents, n (%) Aspirin978 (55.4)1717 (47.8)0.153739 (53.2)715 (51.5)0.035 Heparin1018 (57.6)1892 (52.6)0.101775 (55.8)754 (54.3)0.03 Clopidogrel348 (19.7)717 (20)0.006260 (18.7)261 (18.8)0.002 Ticagrelor33 (1.9)72 (2)0.0125 (1.8)24 (1.7)0.005 Prasugrel12 (0.7)24 (0.7)0.00110 (0.7)10 (0.7)<0.001 Organic nitrates650 (36.8)1347 (37.5)0.014492 (35.4)491 (35.4)0.002 Digoxin48 (2.7)99 (2.8)0.00233 (2.4)39 (2.8)0.027 Amiodarone179 (10.1)354 (9.9)0.01128 (9.2)129 (9.3)0.002Hypertensive drugs, n (%) Calcium channel blockers1025 (58)1927 (53.6)0.089771 (55.5)763 (55)0.012 Hydralazine855 (48.4)1435 (39.9)0.172643 (46.3)638 (46)0.007Anti-diabetic drugs, n (%) Sulfonylurea255 (14.4)470 (13.1)0.04197 (14.2)212 (15.3)0.03 Insulin and analogues1238 (70.1)2255 (62.7)0.156936 (67.4)915 (65.9)0.032Lipid-lowering agents, n (%) Antilipemic agents1207 (68.3)1965 (54.7)0.284897 (64.6)886 (63.8)0.017 HMG CoA reductase inhibitors1175 (66.5)1927 (53.6)0.266879 (63.3)869 (62.6)0.015 Fibrates66 (3.7)96 (2.7)0.06148 (3.5)48 (3.5)<0.001 Ezetimibe80 (4.5)123 (3.4)0.05758 (4.2)53 (3.8)0.018 Nicotinic acid and derivatives35 (2)77 (2.1)0.01129 (2.1)33 (2.4)0.02NSAID, n (%)486 (27.5)701 (19.5)0.19324 (23.3)318 (22.9)0.01ACEi angiotensin-converting enzyme inhibitor, ARB angiotensin receptor blocker, ARNI angiotensin receptor–neprilysin inhibitor, BMI body mass index, BUN blood urea nitrogen, CKD chronic kidney disease, DPP-4i dipeptidyl peptidase-4 inhibitor, ESKD end-stage kidney disease, GLP-1 RA glucagon-like peptide-1 receptor agonist, HFpEF heart failure with preserved ejection fraction, HFrEF heart failure with reduced ejection fraction, HFmrEF heart failure with mildly reduced ejection fraction, LDL low-density lipoprotein, MRA mineralocorticoid receptor antagonist, NSAID nonsteroidal anti-inflammatory drug, SD standard deviation.aBaseline covariates definitions are provided in Supplementary Table 4.bDistribution of GLP-1 receptor agonist use among study participants is provided in Supplementary Table 5.cDiagnosis of heart failure was based on ICD-10-CM codes recorded before the index date. Baseline HF diagnosis categories are presented according to their original ICD-10-CM code categories. For subgroup analyses by code-based EF phenotype proxy in Fig. 4, I50.3, diastolic HF, was used as a proxy for HFpEF with EF > 50%, and I50.2, systolic HF, was used as a proxy for reduced-EF HF, including HFrEF/HFmrEF with EF ≤ 50%. Other I50.x HF diagnosis categories were retained for baseline characterization but not used for EF-phenotype proxy subgrouping. Systolic HF codes (I50.2) identify patients with EF ≤ 50% with a positive predictive value (PPV) of ~90%, and diastolic HF codes (I50.3) identify patients with EF > 50% with a PPV of ~92%30.dHemodialysis access and peritoneal dialysis in Table 1 were defined using representative individual codes to reduce redundancy and prevent overparameterization in propensity-score matching. A more comprehensive composite definition of composite dialysis evidence, including multiple code categories, is presented in Supplementary Table 3 for descriptive validation.eNeoplasms were defined using ICD-10-CM C00–D49 codes recorded during the 3-year baseline period before the index date, including malignant, in situ, benign, uncertain-behavior, and unspecified-behavior neoplasms. This variable represents documented baseline neoplasm history rather than active malignancy at treatment initiation. A sensitivity analysis excluding baseline malignant neoplasm codes (C00–C96) is presented in Supplementary Table 17.fDocumented LVEF values were incorporated as a continuous variable in propensity-score matching, and LVEF-stratified subgroup analyses in the documented subset are presented in Fig. 4.gMineralocorticoid receptor antagonist (MRA) use was defined based on spironolactone, which accounted for the majority of MRA prescriptions in this cohort; eplerenone and finerenone were infrequently prescribed and were not analyzed separately.hDiuretic use was assessed during the 3-year pre-index window; thus, prescriptions may reflect use during advanced CKD prior to ESKD diagnosis. Baseline characteristics of GLP-1 RAs and DPP4i groups before and after propensity score matching ACEi angiotensin-converting enzyme inhibitor, ARB angiotensin receptor blocker, ARNI angiotensin receptor–neprilysin inhibitor, BMI body mass index, BUN blood urea nitrogen, CKD chronic kidney disease, DPP-4i dipeptidyl peptidase-4 inhibitor, ESKD end-stage kidney disease, GLP-1 RA glucagon-like peptide-1 receptor agonist, HFpEF heart failure with preserved ejection fraction, HFrEF heart failure with reduced ejection fraction, HFmrEF heart failure with mildly reduced ejection fraction, LDL low-density lipoprotein, MRA mineralocorticoid receptor antagonist, NSAID nonsteroidal anti-inflammatory drug, SD standard deviation. aBaseline covariates definitions are provided in Supplementary Table 4. bDistribution of GLP-1 receptor agonist use among study participants is provided in Supplementary Table 5. cDiagnosis of heart failure was based on ICD-10-CM codes recorded before the index date. Baseline HF diagnosis categories are presented according to their original ICD-10-CM code categories. For subgroup analyses by code-based EF phenotype proxy in Fig. 4, I50.3, diastolic HF, was used as a proxy for HFpEF with EF > 50%, and I50.2, systolic HF, was used as a proxy for reduced-EF HF, including HFrEF/HFmrEF with EF ≤ 50%. Other I50.x HF diagnosis categories were retained for baseline characterization but not used for EF-phenotype proxy subgrouping. Systolic HF codes (I50.2) identify patients with EF ≤ 50% with a positive predictive value (PPV) of ~90%, and diastolic HF codes (I50.3) identify patients with EF > 50% with a PPV of ~92%30. dHemodialysis access and peritoneal dialysis in Table 1 were defined using representative individual codes to reduce redundancy and prevent overparameterization in propensity-score matching. A more comprehensive composite definition of composite dialysis evidence, including multiple code categories, is presented in Supplementary Table 3 for descriptive validation. eNeoplasms were defined using ICD-10-CM C00–D49 codes recorded during the 3-year baseline period before the index date, including malignant, in situ, benign, uncertain-behavior, and unspecified-behavior neoplasms. This variable represents documented baseline neoplasm history rather than active malignancy at treatment initiation. A sensitivity analysis excluding baseline malignant neoplasm codes (C00–C96) is presented in Supplementary Table 17. fDocumented LVEF values were incorporated as a continuous variable in propensity-score matching, and LVEF-stratified subgroup analyses in the documented subset are presented in Fig. 4. gMineralocorticoid receptor antagonist (MRA) use was defined based on spironolactone, which accounted for the majority of MRA prescriptions in this cohort; eplerenone and finerenone were infrequently prescribed and were not analyzed separately. hDiuretic use was assessed during the 3-year pre-index window; thus, prescriptions may reflect use during advanced CKD prior to ESKD diagnosis. Before PSM, patients receiving GLP-1RAs were younger and more often had overweight/obesity. They more frequently had histories of ischemic heart disease, hypertension and hypertensive heart disease, and hyperlipidemia and exhibited higher hemoglobin A1c (HbA1c) values. Left ventricular ejection fraction (LVEF) was comparable between the GLP-1RAs and DPP-4is groups (51.7 ± 15.2 vs 52.6 ± 14.7, n = 273 and 466), among the subset of patients with available measurements. HF GDMTs—including SGLT2 inhibitors, β-blockers, renin-angiotensin system inhibitors, and mineralocorticoid receptor antagonists—were more frequently used in the GLP-1RAs group (Table 1, Supplementary Tables 4, 5). After 1:1 propensity score matching, 1388 pairs were retained with well-balanced baseline covariates (most standardized mean differences <0.10; Fig. 2). Propensity score density plots before and after matching for GLP-1 RA and DPP-4i groups. The figure shows the distribution of propensity scores for the GLP-1 RA and DPP-4i groups before (left panel) and after (right panel) 1:1 matching, demonstrating well-balanced distributions between the two groups post-matching. DPP-4i dipeptidyl peptidase-4 inhibitor, GLP-1 RA glucagon-like peptide-1 receptor agonist. Outcome analyses for the ESKD-HF population The prespecified primary outcome was a composite of ischemic cardiovascular events (acute myocardial infarction, ischemic stroke, or cardiac arrest, operationalized as an EHR-based 3-point MACE) and heart failure exacerbation (acute-on-chronic heart failure and pulmonary edema) (Supplementary Table 6). For the primary outcome, the event rate was significantly lower in GLP-1RAs than DPP-4is group after PSM (31.7% vs 41.4%; hazard ratio [HR], 0.72; 95% CI, 0.64–0.82; P < 0.0001; Table 2). Kaplan–Meier analysis showed separation of the curves within the follow-up period, with consistently fewer events in the GLP-1RAs group (log-rank P < 0.001; Fig. 3). The E-value was 2.1 (95% upper confidence limit, 1.7), suggesting that only relatively strong unmeasured confounding (2-fold with both exposure and outcome) could fully explain away the observed association. Ischemic cardiovascular events occurred at rates of 18.7% vs 25.1% (HR, 0.74; 95% CI, 0.63–0.87; P = 0.0002; E-value, 2.0 [1.6]) Rates of acute myocardial infarction, stroke, and cardiac arrest were also significantly lower in the GLP-1 RAs group, with hazard ratios of 0.78 (P = 0.0125), 0.76 (P = 0.0459), and 0.61 (P = 0.0097), respectively. Heart failure exacerbation occurred at rates of 21.8% vs 27.9% (HR, 0.76; 95% CI, 0.65–0.88; P = 0.0004; E-value, 2.0 [1.5]); analysis of both subcomponents was also significant, including acute-on-chronic heart failure (coded with I50.23/I50.33; P = 0.0053) and pulmonary edema (coded with J81; P = 0.0042), which reflect acute cardiopulmonary congestion events. Consistent with these results, multivariable Cox regression analyses after adjusting for the same covariates used in the propensity-score model yielded a similar effect estimate (HR, 0.75; 95% CI, 0.67–0.84; P < 0.0001, Supplementary Table 7). The mean follow-up time was 488 ± 279 days in the GLP-1RAs group and 498 ± 279 days in the DPP-4is group, and the median follow-up time was 634 days in the GLP-1RAs group and 721 days in the DPP-4is group. To account for the high competing risk of death, we also evaluated a death-inclusive composite endpoint (primary composite outcome plus all-cause mortality) (528/1388 [38.0%] vs 693/1388 [49.9%]; HR 0.72, 95% CI 0.64–0.80; P < 0.0001) (Table 2). Secondary outcomes, both all-cause mortality (event rates 12.3% vs 18.4%, HR, 0.68; 95% CI, 0.56–0.83; P < 0.0001) and hospitalization (P < 0.0001), were significantly reduced in the GLP-1RAs group (P < 0.001). The mean levels of HbA1c declined 0.4% in the GLP-1RAs group, whereas body mass index (BMI) showed decreased changes in the DPP-4is group and albumin showed slight increases in both groups (Supplementary Table 8). Table 2Incidence of outcomes of interest among the glucagon-like peptide-1 receptor agonist users compared to dipeptidyl peptidase-4 inhibitor users after propensity score matchingGLP-1 receptor agonistDPP-4 inhibitorHazard ratio (95% CI)P-valueE-value (95% UCL)OutcomesEvents, n (%)Events, n (%)ESKD-HF population (I50 and N18.6)aPrimary outcomeb,cComposite ischemic cardiovascular and heart failure exacerbation eventsb,c440/1388 (31.7%)575/1388 (41.4%)0.724 (0.64–0.82)<0.00012.1 (1.7) Heart failure exacerbationb302/1388 (21.8%)387/1388 (27.9%)0.76 (0.654–0.884)0.00042.0 (1.5) Acute on chronic heart failureb231/1388 (16.6%)291/1388 (21.0%)0.782 (0.658–0.93)0.00531.9 (1.4) Pulmonary edemab144/1388 (10.4%)198/1388 (14.3%)0.731 (0.59–0.906)0.00422.0 (1.4) Ischemic cardiovascular eventb,c259/1388 (18.7%)349/1388 (25.1%)0.737 (0.627–0.865)0.00022.0 (1.6) Acute myocardial infarctionb,c172/1388 (12.4%)223/1388 (16.1%)0.776 (0.636–0.947)0.01251.9 (1.3) Strokeb,c88/1388 (6.3%)118/1388 (8.5%)0.755 (0.573–0.995)0.04592.0 (1.1) Cardiac arrestb,c45/1388 (3.2%)75/1388 (5.4%)0.614 (0.424–0.888)0.00972.6 (1.5)Composite ischemic cardiovascular, heart failure exacerbation events and all-cause mortalityd528/1388 (38.0%)693/1388 (49.9%)0.716 (0.639–0.802)<0.00012.1 (1.8)Secondary outcomeAll-cause mortalitye171/1388 (12.3%)256/1388 (18.4%)0.680 (0.561–0.826)<0.00012.3 (1.7)Hospitalizationf802/1388 (57.8%)933/1388 (67.2%)0.743 (0.676–0.816)<0.00012.0 (1.8)Confirmatory dialysis-dependence–coded HF subset (I50 and N18.6 + Z99.2)a,dPrimary outcomeb,cComposite ischemic cardiovascular and heart failure exacerbation eventsb,c231/773 (29.9%)307/773 (39.7%)0.716 (0.604–0.85)0.00042.1 (1.7) Heart failure exacerbationb155/773 (20.1%)206/773 (26.6%)0.732 (0.595–0.902)0.00392.1 (1.5) Ischemic cardiovascular eventb,c140/773 (18.1%)186/773 (24.1%)0.753 (0.605–0.937)0.01102.0 (1.3)Composite ischemic cardiovascular, heart failure exacerbation events and all-cause mortalityd291/773 (37.6%)377/773 (48.8%)0.726 (0.623–0.846)0.00012.1 (1.7)Secondary outcomeAll-cause mortalitye105/773 (13.6%)149/773 (19.3%)0.713 (0.555–0.915)0.00801.9 (1.5)Hospitalizationf440/773 (56.9%)489/773 (63.3%)0.782 (0.687–0.889)0.00021.9 (1.5)Hazard ratios (HRs) and corresponding 95% confidence intervals (CIs) were estimated using Cox proportional hazards models. All statistical tests were two-sided. Exact P-values are reported when applicable; P-values < 0.0001 are presented as such. No adjustment for multiple comparisons was performed.UCL upper confidence limit, CI confidence interval, DPP-4i dipeptidyl peptidase-4 inhibitor, GLP-1 RA glucagon-like peptide-1 receptor agonist.aTwo complementary analytic populations were examined: a primary ESKD-coded cohort (N18.6), which maximizes sensitivity for ESKD identification, and a more restrictive dialysis-dependent-coded subset additionally requiring Z99.2 coding to enhance specificity.bOutcome definitions: Composite of ischemic cardiovascular and heart failure (HF) exacerbation events: acute on chronic systolic (congestive) HF (ICD-10-CM: I50.23), acute on chronic diastolic (congestive) HF (I50.33), pulmonary edema (J81), cerebral infarction (I63), cerebral ischemia (I67.82), cardiac arrest (I46), acute myocardial infarction (I21), subsequent ST elevation (STEMI) and non-ST elevation (NSTEMI) myocardial infarction (I22), and acute ischemic heart disease, unspecified (I24.9), Heart failure exacerbation (HFE): Defined as acute on chronic systolic (congestive) HF (I50.23), acute on chronic diastolic (congestive) HF (I50.33), or pulmonary edema (J81). Ischemic cardiovascular outcome: Defined as cerebral infarction (I63), cerebral ischemia (I67.82), cardiac arrest (I46), acute myocardial infarction (I21), subsequent STEMI and NSTEMI myocardial infarction (I22), or acute ischemic heart disease, unspecified (I24.9). Acute myocardial infarction (AMI): Defined as acute myocardial infarction (I21), subsequent STEMI and NSTEMI myocardial infarction (I22), or acute ischemic heart disease, unspecified (I24.9). Stroke: Defined as cerebral infarction (I63) or cerebral ischemia (I67.82). Cardiac arrest: Defined as cardiac arrest (I46) (Supplementary Table 6).cThese definitions were not restricted to non-fatal myocardial infarction or non-fatal stroke; the diagnosis codes used for myocardial infarction and stroke may also encompass cardiovascular deaths. Accordingly, the composite outcome in this study more closely reflects a 3-point major adverse cardiovascular event (MACE), consistent with the operational definitions adopted in prior database studies46,47.dA composite outcome including the primary outcome and all-cause mortality was evaluated to account for competing risks.eAll-cause mortality: Defined by demographic records indicating death (Deceased) or diagnosis of ill-defined and unknown cause of mortality (ICD-10-CM: R99).fHospitalization: Defined by any of the following visit types: inpatient acute (HL7V3.0:Visit ACUTE), inpatient encounter (IMP), inpatient non-acute (NONAC), emergency (EMER), short stay (SS), or observation encounter (OBSENC) (Supplementary Table 6). Incidence of outcomes of interest among the glucagon-like peptide-1 receptor agonist users compared to dipeptidyl peptidase-4 inhibitor users after propensity score matching Hazard ratios (HRs) and corresponding 95% confidence intervals (CIs) were estimated using Cox proportional hazards models. All statistical tests were two-sided. Exact P-values are reported when applicable; P-values < 0.0001 are presented as such. No adjustment for multiple comparisons was performed. UCL upper confidence limit, CI confidence interval, DPP-4i dipeptidyl peptidase-4 inhibitor, GLP-1 RA glucagon-like peptide-1 receptor agonist. aTwo complementary analytic populations were examined: a primary ESKD-coded cohort (N18.6), which maximizes sensitivity for ESKD identification, and a more restrictive dialysis-dependent-coded subset additionally requiring Z99.2 coding to enhance specificity. bOutcome definitions: Composite of ischemic cardiovascular and heart failure (HF) exacerbation events: acute on chronic systolic (congestive) HF (ICD-10-CM: I50.23), acute on chronic diastolic (congestive) HF (I50.33), pulmonary edema (J81), cerebral infarction (I63), cerebral ischemia (I67.82), cardiac arrest (I46), acute myocardial infarction (I21), subsequent ST elevation (STEMI) and non-ST elevation (NSTEMI) myocardial infarction (I22), and acute ischemic heart disease, unspecified (I24.9), Heart failure exacerbation (HFE): Defined as acute on chronic systolic (congestive) HF (I50.23), acute on chronic diastolic (congestive) HF (I50.33), or pulmonary edema (J81). Ischemic cardiovascular outcome: Defined as cerebral infarction (I63), cerebral ischemia (I67.82), cardiac arrest (I46), acute myocardial infarction (I21), subsequent STEMI and NSTEMI myocardial infarction (I22), or acute ischemic heart disease, unspecified (I24.9). Acute myocardial infarction (AMI): Defined as acute myocardial infarction (I21), subsequent STEMI and NSTEMI myocardial infarction (I22), or acute ischemic heart disease, unspecified (I24.9). Stroke: Defined as cerebral infarction (I63) or cerebral ischemia (I67.82). Cardiac arrest: Defined as cardiac arrest (I46) (Supplementary Table 6). cThese definitions were not restricted to non-fatal myocardial infarction or non-fatal stroke; the diagnosis codes used for myocardial infarction and stroke may also encompass cardiovascular deaths. Accordingly, the composite outcome in this study more closely reflects a 3-point major adverse cardiovascular event (MACE), consistent with the operational definitions adopted in prior database studies46,47. dA composite outcome including the primary outcome and all-cause mortality was evaluated to account for competing risks. eAll-cause mortality: Defined by demographic records indicating death (Deceased) or diagnosis of ill-defined and unknown cause of mortality (ICD-10-CM: R99). fHospitalization: Defined by any of the following visit types: inpatient acute (HL7V3.0:Visit ACUTE), inpatient encounter (IMP), inpatient non-acute (NONAC), emergency (EMER), short stay (SS), or observation encounter (OBSENC) (Supplementary Table 6). Kaplan–Meier time-to-event curves comparing GLP-1 RA and DPP-4i groups. Kaplan–Meier curves depict cumulative event-free or survival probabilities over time in the GLP-1 RA (purple) and DPP-4i (teal) cohorts after propensity score matching for a composite ischemic cardiovascular and heart failure exacerbation events; b heart failure exacerbation events; c ischemic cardiovascular events; d all-cause mortality; e composite ischemic cardiovascular, heart failure exacerbation events, and all-cause mortality; and f composite ischemic cardiovascular and heart failure exacerbation events in the dialysis-dependence–coded subset. Solid lines represent Kaplan–Meier estimates (center), and shaded bands indicate 95% confidence intervals (error bands). P-values were calculated using two-sided log-rank tests. Exact P-values are reported when applicable; P-values < 0.0001 are presented as such. No adjustment for multiple comparisons was performed. Source data are provided as a Source data file. DPP-4i dipeptidyl peptidase-4 inhibitor, GLP-1 RA glucagon-like peptide-1 receptor agonist. Analyses in confirmatory dialysis-dependence–coded HF subset We further evaluated a prespecified higher-specificity confirmatory dialysis-dependence–coded HF subset, requiring both ICD-10-CM N18.6 and Z99.2 coding (Fig. 1). After repeating 1:1 propensity score matching within this nested subset (n = 2927), baseline covariates remained well-balanced (n = 773 pairs; Table 2; Supplementary Table 9 for baseline characteristics). In this dialysis-dependence–coded HF subset, GLP-1RAs use was associated with a lower risk of the primary composite outcome (29.9% vs 39.7%; HR 0.716, 95% CI 0.604–0.850; P = 0.0004), with concordant reductions in both heart failure exacerbation (HR 0.732, 95% CI 0.595–0.902; P = 0.0039) and ischemic cardiovascular events (HR 0.753, 95% CI 0.605–0.937; P = 0.0110). Secondary outcomes were directionally consistent, including the death-inclusive composite endpoint (HR 0.726, 95% CI 0.623–0.846; P = 0.0001), all-cause mortality (HR 0.713, 95% CI 0.555–0.915; P = 0.0080), and hospitalization (HR 0.782, 95% CI 0.687–0.889; P = 0.0002) (Table 2). Subgroup analyses in the primary ESKD-HF population Subgroup analyses showed no modification effects for primary outcome by sex, age, types, coded-based HF subtype, and prevalent or incident HF in GLP-1RAs and DPP-4is groups (interaction test P > 0.05 for each). Stratification by baseline use of GDMTs for heart failure (β-blockers, renin–angiotensin system inhibitors, mineralocorticoid receptor antagonists, and SGLT2 inhibitors) demonstrated consistent protective associations. When stratified by ischemic cardiovascular history, the hazard ratio was 0.76 (95% CI, 0.59–0.97; P = 0.026) among patients with prior events, and 0.71 (95% CI, 0.61–0.82; P < 0.001) among those without prior events, without mediation effect (P for interaction 0.663) (Fig. 4). Using validated ICD-10-CM code-based ejection fraction (EF) phenotype proxies30, GLP-1RA initiation was associated with lower risk in both the heart failure with preserved ejection fraction (HFpEF) proxy subgroup (I50.3, diastolic HF; proxy for EF > 50%) and the heart failure with reduced ejection fraction or heart failure with mildly reduced ejection fraction (HFrEF/HFmrEF) proxy subgroup (I50.2, systolic HF; proxy for EF ≤ 50%), with no evidence of interaction (P for interaction = 0.491). Other I50.x HF diagnosis categories were used for baseline characterization only, because combined or unspecified HF codes poorly predicted HFrEF or HFpEF in prior validation data. In the subset with documented LVEF, the association remained significant (HR 0.64, 95% CI 0.49–0.84; P = 0.001); EF-based reclassification yielded significant associations for LVEF ≥ 50% (HR 0.66, 95% CI 0.48–0.9; P = 0.008) and LVEF < 50% (HR 0.64, 95% CI 0.42–0.98; P = 0.036), with no interaction by EF category (P = 0.910)(Fig. 4). GLP-1RAs initiation was associated with lower risk in both females (HR 0.63, 95% CI 0.52–0.77) and males (HR 0.77, 95% CI 0.65–0.91), with no sex interaction (P = 0.138). Subgroup analysis of all-cause mortality and composite of primary outcome and all-cause mortality is demonstrated in Supplementary Fig. 2 and Fig. 3. Subgroup analyses of primary outcome, composite ischemic cardiovascular and heart failure exacerbation events. Forest plots display hazard ratios for GLP-1 RA versus DPP-4i across prespecified subgroups after propensity score matching. Squares represent hazard ratio estimates (center), and horizontal lines indicate 95% confidence intervals (error bars). Hazard ratios and corresponding 95% confidence intervals were estimated using Cox proportional hazards models. Interaction P-values were derived from likelihood-ratio tests comparing Cox proportional hazards models with and without treatment-by-subgroup interaction terms. All statistical tests were two-sided. Exact P-values are reported when applicable; P-values < 0.001 are presented as such. No adjustment for multiple comparisons was performed. Source data are provided as a Source data file. BMI body mass index, DPP-4i dipeptidyl peptidase-4 inhibitor, GLP-1 RA glucagon-like peptide-1 receptor agonist, HbA1c glycated hemoglobin, HFmrEF heart failure with mildly reduced ejection fraction, HFpEF heart failure with preserved ejection fraction, HFrEF heart failure with reduced ejection fraction, MRA mineralocorticoid receptor antagonist, RASi renin–angiotensin system inhibitor, SGLT2i sodium–glucose cotransporter 2 inhibitor, y/o years old. Sensitivity analysis in the primary ESKD-HF population Excluding patients who underwent kidney transplantation after index day (HR, 0.71; 95% CI, 0.62–0.81; P < 0.0001; Supplementary Table 10) or those who switched to the alternative drug class produced results consistent with the primary analysis (HR, 0.72; 95% CI, 0.62–0.83; P < 0.0001; Supplementary Table 11), which remained robust under sustained-exposure restrictions (Supplementary Table 11A), mitigating concern about transient prescriptions, and in analyses restricted to semaglutide initiators versus DPP-4i users (Supplementary Table 11B). Findings were also similar in analyses restricted to patients with extended population with prevalent heart failure (Supplementary Table 12) and an alternative enrollment specification without a 3-month washout (including patients with any prior MACE; Supplementary Table 12A). In supportive durability analyses evaluating potential dependence on follow-up horizon and censoring, right-censored Cox models demonstrated early separation by 6 months, and comparable effect estimates over the primary 2-year window and supportive 3- and 5-year horizons (Supplementary Table 13). Varying the permissible window for treatment initiation at 6, 9, 12, 15, and 18 months yielded hazard ratios consistent with the main analysis (Supplementary Table 14). To assess heterogeneity by dialysis vintage, we summarized time from first recorded ESKD diagnosis to treatment initiation (Supplementary Table 14A) and conducted stratified analyses by initiation within <6 versus ≥6 months; effect estimates were similar across strata with no evidence of interaction by dialysis vintage (P = 0.7) (Supplementary Table 14B). Finally, negative-control outcomes (skin cancer and traumatic brain injury) showed no significant associations, whereas the positive-control outcome (nausea) showed an increased risk among GLP-1RA users as expected (Supplementary Table 15). Notably, mechanical complication of a vascular dialysis catheter, used as a proxy for dialysis care intensity, was also not significantly different between groups (HR 0.784, 95% CI 0.521–1.178; P = 0.2402), suggesting comparable dialysis-related healthcare engagement (Supplementary Table 15). Consistently, the odds of receiving seasonal influenza and/or pneumococcal vaccination, a proxy for health-seeking behavior, were similar between groups (OR 1.099, 95% CI 0.927–1.302). Baseline healthcare utilization (≥2 emergency department visits) was comparable between groups, mitigating concerns about differential surveillance (Supplementary Table 15A). Results were directionally consistent in the full unmatched cohort (Supplementary Table 16). Given the potential prognostic impact of malignancy, an additional analysis excluding patients with baseline malignant neoplasm codes (C00–C96) recorded within 3 years before the index date yielded concordant results (Supplementary Table 17). Discussion In this large-scale target trial emulation, we provide the first real-world evidence that GLP-1RAs initiation was associated with a significantly lower risk of cardiovascular events, including ischemic cardiovascular complications and HF exacerbations, in patients with HF in diabetic ESKD receiving maintenance dialysis. This ESKD population was historically excluded from randomized controlled trials and carries an extraordinarily high cardiovascular risk. Importantly, the cardiovascular benefit signals were consistent across clinically relevant subgroups, including those with or without prior MACE, and remained evident in patients receiving baseline GDMT. These findings suggest that GLP-1RAs may address a longstanding therapeutic gap in ESKD by favorably modifying cardiovascular outcomes in a population where reducing MACE has traditionally been extremely challenging and where conventional pharmacotherapies have proven ineffective or lack robust supporting evidence, potentially translating to survival benefit. Our study leverages the global TriNetX network to address the challenges of studying HF in ESKD. Our comprehensive baseline characterization of this population—including HF status, cardiovascular comorbidities, dialysis modality, vascular access, anemia, iron status, body weight, lipid profile, use of cardiovascular medications, compatible coverage of GDMT use, socioeconomic status indicators, and healthcare engagement—allowed us to recognize disease characteristics and potential channeling patterns before matching and strengthening comprehensive confounding control. To reduce channeling at treatment initiation and differences in healthcare engagement inherent, we designed the analysis as an active-comparator, new-user emulation with extensive PSM, including proxies of care intensity and comorbidity burden. Prior to matching, the GLP-1RAs group had a higher burden of cardiovascular risk factors and advanced disease, including ischemic heart disease, hypertension, and hyperlipidemia, but also received more comprehensive GDMT. Greater GDMT use among GLP-1RAs initiators may reflect more complete implementation of guideline-recommended HF therapy, but it may also reflect higher care intensity and healthcare engagement that could confound observational associations. Because GLP-1RAs initiators also had more cardiovascular comorbidity and advanced disease, confounding by indication (channeling) may operate in the opposite direction, making the net direction of residual confounding uncertain. After matching, both GLP-1RAs and DPP-4is groups had similar disease characteristics, and the beneficial effects of GLP-1RAs persisted across multiple outcomes. The robustness checks across multiple bias probes (PSM with care-intensity proxies, negative-control analyses, and stricter sustained-exposure definitions) were directionally concordant, arguing against a single explanation based solely on prescribing channeling or a healthy-user artifact and supporting the internal consistency of the observed associations. HF in ESKD remains exceptionally challenging, with therapeutic options limited and GDMT use hampered by scarce supporting evidence. Since early trials such as those by Cice et al. (carvedilol, telmisartan)10,31, subsequent interventions have rarely demonstrated clear prognostic benefit. Trials of spironolactone (ALCHEMIST)11 and ramipril (ACRDIA)12 also failed to demonstrate MACE or survival improvement in this high cardiovascular risk ESKD population. Even statins, cornerstones of lipid management with proven efficacy in earlier stages of CKD, failed to significantly reduce MACE in large dialysis trials such as 4D and AURORA, with benefits largely confined to non-dialysis CKD patients, as seen in SHARP32. Other approaches, including PCSK9 inhibitors and anti-inflammatory agents such as IL-6 blockade, remain largely investigational with only preliminary evidence33,34. Against this backdrop, our finding that GLP-1RAs may confer cardiovascular protection in diabetic dialysis patients is unexpected and potentially paradigm-shifting. Our findings appear striking and provide clinically relevant real-world evidence, although confirmation in prospective randomized studies remains warranted. Because our analyses rely on EHR-based phenotyping in dialysis-dependent ESKD, several aspects of HF classification and outcome coding warrant clinically informed interpretation. HF phenotypes were code-based in the primary analyses, and incomplete echocardiography capture—particularly the high prevalence of unspecified HF codes—may introduce phenotype misclassification; thus, EF-stratified findings are supportive and exploratory, especially for LVEF < 50%21,35. For outcomes, HF-exacerbation codes (pulmonary edema [J81] and acute-on-chronic HF [I50.23/I50.33]) were interpreted as clinically consequential cardiopulmonary congestion prompting urgent management rather than etiologic distinction of between HF and pure fluid overload36,37. Because all enrolled patients in the current study had clinically established HF history, these acute congestion events are unlikely to represent volume overload alone. In dialysis-dependent ESKD, weight change is bidirectional: intentional weight loss may be beneficial in obesity, whereas unintentional loss can signal protein–energy wasting38. This differs from obesity-related HFpEF trials, where large pharmacologic weight loss was accompanied by symptomatic and functional improvement22,39. In our two PSM groups, body size metrics showed no major separation over follow-up, and serum albumin increased slightly without evidence of decline (Supplementary Table 8), providing reassurance against an overt nutritional-wasting signal. HbA1c declined by approximately 0.4 percentage points (~0.4%), but HbA1c and BMI are imperfect metabolic markers in dialysis because of anemia and volume shifts. Cardiovascular protection was directionally consistent across BMI strata (P for interaction = 0.604), supporting that the observed association is not confined to higher-BMI strata and is unlikely to be explained solely by weight or glycemic indices. ESKD is a uremic, pro-oxidant, and pro-inflammatory state with endothelial dysfunction and heightened thrombogenicity. In this context, proposed non-metabolic actions of GLP-1RAs include attenuation of systemic inflammatory mediators and improvements in endothelial or microvascular function40–42, with additional signals suggesting modulation of platelet activation and reduced ischemic injury in proof-of-concept myocardial infarction studies43,44. Importantly, beyond preventing overt MACE, because hemodialysis can provoke recurrent subclinical myocardial ischemia (myocardial stunning) that accumulates over time and contributes to ventricular dysfunction45, these mechanisms could plausibly reduce the cumulative burden of silent ischemic injury that fuels HF progression in ESKD. Sex differences merit attention in ESKD and HF, given known variation in cardiovascular risk profiles, treatment patterns, and outcomes. In sex-stratified analyses, associations were directionally consistent, with numerically stronger estimates in females than males and no evidence of interaction. This pattern aligns with meta-analyses reporting greater risk reductions for MACE or renal endpoints among women46. Key strengths of this study include the large sample size and multinational coverage of TriNetX24,25,47, enabling to enroll a large ESKD–HF population. Robustness was supported by extensive prespecified analyses, including multivariable models and sensitivity analyses varying key design assumptions and cohort specifications (e.g., enrollment definitions, eligibility windows, follow-up horizons, and exposure persistence), with consistent benefit was also observed in the full unmatched cohort (Supplementary Table 16). Second, we incorporated available socioeconomic measures and proxies of care intensity and healthcare engagement, together with dialysis-related indicators, to strengthen confounding control in federated EHR data. Third, we focused on cardiovascular- and HF-specific endpoints rather than all-cause mortality or mortality-dominated composites, providing greater specificity than most prior real-world studies. Finally, dialysis status was addressed using complementary ascertainment strategies, including an ESKD-coded HF population with dialysis-ascertainment analyses and corroboration in a confirmatory dialysis-dependence–coded HF subset, strengthening the relevance to maintenance dialysis. This study also has limitations. First, competing risks by death may preclude the occurrence of non-fatal events such as MACE. Although the median follow-up was modestly shorter in the GLP-1RAs group, all-cause mortality was lower among GLP-1RAs users, and death-inclusive composite analyses yielded consistent results, suggesting that the difference in follow-up duration was unlikely to be driven by earlier deaths. Relatedly, a prespecified 2-year primary follow-up window may not fully capture longer-term effectiveness; supportive 3- and 5-year durability estimates require cautious interpretation given increasing crossover, secular trends, and competing risks over extended follow-up (Supplementary Table 13). Second, cause-specific cardiovascular death is not available in TriNetX. We therefore used an EHR-based 3-point MACE defined by ischemic-event codes, with all-cause mortality analyzed separately to preserve endpoint specificity. Third, residual confounding cannot be fully excluded, particularly from incompletely captured frailty and functional status, provider-level prescribing channeling, and dialysis prescription parameters (e.g., ultrafiltration targets, modality details, adequacy, and intradialytic hemodynamics). We partially addressed these concerns by adjusting for socioeconomic indicators and proxies of healthcare utilization, and by leveraging available dialysis-related profiles (laboratory measures, intradialytic hypotension, and dialysis vintage). Nevertheless, unmeasured differences in patient engagement and disease severity cannot be fully excluded, and the observed magnitude may therefore reflect residual confounding or channeling in addition to any pharmacologic effect. The magnitude should be interpreted cautiously in dialysis-dependent ESKD, where several cardioprotective trials have shown modest or neutral effects, although this does not preclude benefit from drug classes with distinct mechanisms and established cardiovascular effects in diabetes and CKD. Fourth, medication exposure was based on baseline prescription orders, and time-varying use during follow-up (e.g., crossover and persistence) could not be fully characterized; therefore, exposure misclassification remains possible. However, per-protocol and sustained-exposure sensitivity analyses were directionally consistent (Supplementary Tables 11, 11A). Last, the active-comparator design required diabetes, limiting extrapolation to non-diabetic ESKD. Notably, recent trials such as SELECT and STEP-HFpEF have demonstrated benefits of GLP-1RAs in HF patients without diabetes22,35. Future studies are needed to clarify whether similar benefits extend to non-diabetic ESKD patients with HF. In conclusion, this large-scale target trial emulation provides the real-world evidence that, compared with DPP-4is, GLP-1RAs reduce ischemic complications and heart failure exacerbations in patients with diabetic dialysis-dependent ESKD with HF, a population with extreme cardiovascular vulnerability and limited therapeutic options. These findings suggest GLP-1RAs may confer cardiovascular protection and survival benefits and could be considered a treatment option in this high-risk group, awaiting confirmation in randomized trials. Methods Ethics approval and reporting This study complied with all relevant ethical regulations and was approved by the Institutional Review Board of National Cheng Kung University Hospital, Taiwan (approval numbers B-ER-113-460 and B-ER-114-456). The requirement for informed consent was waived by the Institutional Review Board because the study used de-identified, aggregated electronic health record data without direct participant contact or access to identifiable individual-level data, in accordance with Taiwan’s Human Subjects Research Act and Personal Data Protection Act. The study complied with the Declaration of Helsinki and followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guidelines48. The Guidelines for Accurate and Transparent Health Estimates Reporting (GATHER)49 recommendations were considered not applicable because this study used a single federated EHR platform for a cohort-based target trial emulation and did not generate population-level health estimates across geographic regions, calendar periods, or subnational populations. Data source We conducted a cohort study using the TriNetX Global Health Research Network, which aggregates de-identified electronic health records from ~150 healthcare organizations encompassing >180 million patients worldwide. Available data include diagnoses, medications, procedures, laboratory results, demographics, and healthcare utilization. Cohorts and outcomes were prespecified within the platform. This platform has been widely used to conduct real-world studies addressing treatment effectiveness and safety in specific disease cohorts and under-represented populations50,51. Study design and population We applied a target trial emulation framework to observational data to emulate the design of a randomized controlled trial, following the TARGET reporting checklist52 (Supplementary Table 1). Participants aged ≥18 years with documented HF and diabetic ESKD during August 2016 and Feb 2026 were assigned to two exposure groups. Sex was obtained from structured demographic fields in TriNetX and used as recorded; the method used to determine sex or gender at the point of care was not uniformly available. Gender identity was not available in the aggregated TriNetX output. The GLP-1RAs group included patients who initiated a GLP-1RAs within 2 years after the first diagnosis of ESKD and, concomitantly, within 2 years following the first diagnosis of HF. The comparator group consisted of patients who initiated a DPP-4is (Supplementary Fig. 1). To ensure a new-user, active-comparator design, we excluded individuals with any prior use of GLP-1RAs or DPP-4is, those who had undergone kidney transplantation before the index date, and those with a diagnosis of cardiovascular events within 3 months preceding the index date. End-stage kidney disease (ESKD) is operationally defined as chronic kidney failure requiring kidney replacement therapy (KRT), including maintenance dialysis or kidney transplantation, to sustain life, consistent with the definition applied in the US Renal Data System (USRDS). HF and ESKD were identified using ICD-10-CM codes I50 and N18.6, respectively. In the SCREAM cohort, 99.1% of dialysis patients had a registry diagnosis of renal failure, and the coding scheme identified eGFR <30 with high accuracy (positive predictive value (PPV) 93.5%, negative predictive value (NPV) 99.2%)53. Moreover, multiple studies have applied ICD-10-CM N18.6 as the definition of ESKD for enrollment or outcome54–56. To address potential misclassification and under-recording of dialysis status, we prespecified two complementary analytic populations: a sensitivity-maximizing target ESKD-coded population (ICD-10-CM N18.6 excluding kidney transplant status) and a PPV-enriched, higher-specificity subset additionally requiring Z99.2 (dialysis-dependence-coded subset). To characterize maintenance dialysis in the N18.6-defined ESKD-coded HF population enrolled in our study, we used a composite dialysis-evidence definition incorporating (i) vascular access–related procedures, (ii) Z99.2, (iii) peritoneal dialysis–related codes, or (iv) any dialysis procedure records during ESKD to support sustained rather than temporary dialysis (Supplementary Table 3) to ascertain patient undergoing dialysis before beginning use of testing drugs. ICD-10-CM codes (I50.x) in the first diagnostic position, which have shown a PPV of 98% for acute HF hospitalizations30. HF phenotypes were classified using ICD-10-CM codes: systolic HF (I50.2) was used as a proxy for HFrEF or HFmrEF with LVEF ≤ 50%, and diastolic HF (I50.3) as a proxy for HFpEF with LVEF > 50%. These codes have been validated against echocardiography, with positive predictive values of ~90% for EF ≤ 50% and ~92% for EF > 50%. We therefore refer to these as ‘code-based’ HFrEF/HFmrEF and HFpEF subgroups throughout the study30. The index date was the date of the first prescription for the exposures; baseline covariates were assessed using the most recent data during the 3 years preceding the index date. Primary analyses used an intention-to-treat framework, preserving baseline comparability and yielding policy-relevant effectiveness estimates despite post-index treatment changes. Follow-up began 1 day after the index date and continued until the first occurrence of an outcome, loss to follow-up, or 2 years, whichever occurred first. The 2-year administrative censoring horizon was prespecified as the primary follow-up window to prioritize internal validity; longer horizons (3 and 5 years) were evaluated as supportive durability sensitivity analyses57,58. Follow-up time was calculated by the last record or death in TriNetX. Detailed cohort definitions and code lists are provided in Supplementary Table 2. Prespecified outcomes The primary outcome was a composite of ischemic cardiovascular events and heart failure exacerbation (HFE). Ischemic cardiovascular events (EHR-based 3-point major adverse cardiovascular event [MACE]) included acute myocardial infarction (MI), ischemic stroke, and cardiac arrest. These definitions were not restricted to non-fatal events, and some coded MI or stroke episodes may represent events culminating in death; however, cause-specific cardiovascular death is not uniformly available in TriNetX56. Accordingly, all-cause mortality was analyzed as a separate endpoint and was not included in the 3-point MACE definition38. All-cause mortality was analyzed separately. Restriction to ischemic (rather than hemorrhagic) stroke was adopted to minimize potential confounding from trauma-related hemorrhage, consistent with prior real-world data studies. Heart failure exacerbation was identified using ICD-10-CM codes I50.23, I50.33 (acute on chronic systolic or diastolic heart failure), and J81 (pulmonary edema) to capture clinically meaningful acute cardiopulmonary congestion/decompensation events that typically prompt urgent medical attention in dialysis-dependent ESKD rather than to adjudicate renal versus cardiac etiology. Risk of individual components within the primary outcome was also analyzed. Secondary outcomes included all-cause mortality and hospitalization. In addition to propensity score–matched analyses, we applied a multivariable Cox proportional hazards regression model in the overall cohort, adjusting for the same baseline covariates, to estimate hazard ratios for the study outcomes (Supplementary Table 7). Temporal changes in HbA1c levels and BMI were evaluated between the matched groups to determine whether the observed cardiovascular protective effect of GLP-1RAs could be explained by improvements in glycemic control or weight reduction, or instead may reflect alternative underlying mechanisms. Complete outcome code lists are provided in Supplementary Table 6. Covariates Baseline covariates were assessed using the most recent data available within 3 years before the index date and were selected. Because they provide relevant clinical information, were potential confounders or important risk factors. Covariates included demographics (age and sex), lifestyle factors, comorbidities (eg, obesity, heart failure subtype, cardiovascular and cerebrovascular disease, peripheral vascular disease), and medication use (GDMT, hypoglycemic, cardiovascular, and lipid-lowering agents, nonsteroidal anti-inflammatory drugs). Dialysis-related variables comprised vascular access type and peritoneal dialysis. Clinical and laboratory measures included body mass index, blood pressure, hemoglobin A1c, albumin, hemoglobin, renal and lipid profiles, and iron indices. Cardiac parameters included LVEF, B-type natriuretic peptide, and N-terminal pro–B-type natriuretic peptide. Detailed definitions and coding are provided in Supplementary Table 4. Statistical analysis Baseline characteristics were summarized, with categorical variables reported as numbers and percentages and continuous variables as means with standard deviations. We performed 1:1 propensity score (PS) matching before the primary, subgroup, and sensitivity analyses to reduce confounding. Relevant covariates were used to estimate a propensity score for each subject using logistic regression to model the probability of receiving GLP-1RAs therapy. Patients in the treatment group were then matched to comparators using a greedy nearest-neighbor algorithm, with an absolute standardized mean difference (SMD) of less than 0.1 considered indicative of adequate balance. Kaplan–Meier methods with log-rank tests were used to depict and compare survival probabilities, while other time-to-event outcomes were analyzed using Cox proportional hazards models to estimate hazard ratios (HRs) with 95% confidence intervals (CIs). The E-value represents the minimum strength of association, on the risk ratio scale, that an unmeasured confounder would need with both treatment and outcome to fully explain away the observed effect estimate, analogous to a sensitivity measure derived from the observed hazard ratio. All analyses were conducted within the TriNetX platform on 2026/02/10. Primary statistical analyses, including propensity-score matching, Kaplan–Meier analyses, log-rank tests, Cox proportional hazards models, and standardized mean difference calculations, were conducted using the integrated analytic tools within the TriNetX research platform. R software (version 4.4.1; R Foundation for Statistical Computing, Vienna, Austria) was used for data organization, figure preparation, and supplementary statistical calculations where applicable. Prespecified subgroup analysis To evaluate heterogeneity of treatment effect, we prespecified subgroup analyses by sex, age (18–64 vs ≥65 years), heart failure phenotype (code-based HFpEF vs HFrEF or HFmrEF), history of cardiovascular events, use of guideline-directed medical therapy, HbA1c (≤7% vs >7%), and BMI. Within each subgroup, propensity score matching was repeated and hazard ratios were estimated using Cox proportional hazards models, thereby addressing subgroup-specific confounding. Sensitivity analyses To assess robustness of the primary estimates, we conducted prespecified sensitivity analyses varying key design assumptions on eligibility, exposure, follow-up, and residual confounding. We excluded patients who underwent kidney transplantation after the index date to minimize bias from altered cardiovascular risk after transplantation59, and we performed per-protocol analyses excluding/censoring treatment switchers with sustained-exposure restrictions, and evaluated agent-level consistency in semaglutide initiators versus DPP-4 inhibitor users. We tested generalizability/eligibility by including an extended cohort with prevalent HF before ESKD and by removing the 3-month washout (allowing prior MACE). We examined dependence on follow-up horizon using right-censored Cox models over 2-, 3-, and 5-year windows. We varied the allowable initiation window after the first recorded ESKD/HF diagnoses (6–18 months) and assessed dialysis vintage via time from ESKD diagnosis to initiation, including stratified analyses by initiation timing. To probe residual confounding and differential healthcare engagement, we evaluated negative controls (skin cancer, traumatic brain injury)60, a positive control (nausea), a dialysis care-intensity proxy (mechanical complication of vascular dialysis catheter), preventive-care behavior (seasonal influenza and/or pneumococcal vaccination), and baseline surveillance (≥2 emergency department visits). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data