Tirzepatide versus GLP-1 receptor agonists and stroke in patients with atrial fibrillation and type 2 diabetes

Abstract

Patients with atrial fibrillation and type 2 diabetes remain at high risk of stroke despite anticoagulation and glucose-lowering treatment. We examined whether tirzepatide was associated with lower cerebrovascular risk than established glucagon-like peptide-1 receptor agonists in a retrospective cohort from a large international federated electronic health-record network. Here we show that, among 14,118 matched pairs of adults initiating tirzepatide or another glucagon-like peptide-1 receptor agonist, tirzepatide use was associated with lower risks of any stroke, ischemic stroke, hemorrhagic stroke, and death from any cause. The absolute difference in stroke risk was 2.1 percentage points at one year, and the absolute difference in mortality was 6.2 percentage points. These observational findings support further assessment of tirzepatide as a potential option for people with both conditions, but randomized trials are required before treatment recommendations can change.

Among adults with atrial fibrillation and type 2 diabetes, tirzepatide use was associated with lower risks of stroke and death than other GLP-1 receptor agonists. The findings support trials of its cerebrovascular benefits in this high-risk group.

INTRODUCTION

Stroke remains one of the leading causes of death and long-term disability worldwide, with an estimated 12 million new cases and 6.5 million deaths annually^1,2. Atrial fibrillation (AF) is the strongest modifiable risk factor for stroke, conferring a five-fold higher risk compared with the general population, and accounting for up to 20–30% of ischemic strokes^3,4. Type 2 diabetes mellitus (T2D) independently increases stroke risk by 1.5- to 2-fold^5,6. When AF and T2D coexist, the combined effect is synergistic, leading to disproportionately high risks of stroke, cardiovascular death, and all-cause mortality, with annual stroke incidence exceeding 3–4% compared with about 1% in the general population^7–10. Both ischemic and hemorrhagic strokes contribute to this burden, resulting in long-term disability, diminished quality of life, and heavy dependence on healthcare and caregiving resources^7,11–13. Globally, more than 15–20 million patients live with concurrent AF and T2D, and their associated expenditures are estimated in the tens of billions of dollars annually^9,14,15. Despite the use of direct oral anticoagulants (DOACs) for stroke prevention in AF and guideline-based glycemic management for T2D, a substantial residual stroke risk—estimated at 2–3% annually despite optimal therapy—persists^16,17. This is particularly critical because these patients are already on glucose-lowering agents; thus, if a diabetes medication could simultaneously reduce stroke risk, it would provide dual benefits—metabolic control and cerebrovascular protection.

Glucagon-like peptide-1 receptor agonists (GLP-1 RAs) have become a cornerstone of T2D management, with robust evidence from large randomized controlled trials (LEADER, SUSTAIN-6, REWIND) demonstrating reductions in major adverse cardiovascular events (MACE), cardiovascular death, and ischemic stroke^18–21. These findings led to their endorsement in international guidelines by the American Diabetes Association (ADA) and European Association for the Study of Diabetes (EASD)²². Recent real-world studies have further suggested that GLP-1RA therapy may be associated with reduced risks of both ischemic and hemorrhagic cerebrovascular events, including lower risks of nontraumatic intracerebral hemorrhage and improved neurovascular outcomes, supporting a potential class-level cerebrovascular benefit^23–26. However, most prior trials compared GLP-1 RAs with placebo, leaving unanswered the clinically important question of whether the next-generation dual glucose-dependent insulinotropic polypeptide (GIP)/GLP-1 RA, tirzepatide, provides incremental benefits over traditional GLP-1 RAs^{3,20,21,27–32}. Recent real-world evidence has begun to address this: a 2024 study (Chuang et al.) showed that tirzepatide, compared with GLP-1 RAs, was associated with lower risks of all-cause mortality, MACE, and renal outcomes in a general T2D population³³, while Wu et al³⁴. reported lowerAF burden among patients with AF and T2D treated with tirzepatide³⁴. Nevertheless, these studies did not evaluate stroke subtypes, nor did they address potential heterogeneity among individual GLP-1RAs²⁴. Yet, neither study evaluated stroke subtypes, and no large-scale head-to-head analysis has specifically examined cerebrovascular outcomes in high-risk AF + T2D patients.

Tirzepatide, a dual glucose-dependent insulinotropic polypeptide and glucagon-like peptide-1 receptor agonist, has effects on glycemic control, body weight, and kidney-related outcomes, and may also influence inflammation, lipid metabolism, blood pressure, and vascular health^22,35–38. These properties provide a rationale for examining potential cerebrovascular differences relative to established glucagon-like peptide-1 receptor agonists. We therefore assessed whether tirzepatide use was associated with lower risks of overall stroke, ischemic stroke, hemorrhagic stroke, and all-cause mortality in adults with atrial fibrillation and type 2 diabetes.

RESULTS

Baseline characteristics

Before PSM, patients receiving tirzepatide (N = 14,126) were slightly younger (mean age 64.3 years vs. 66.7 years) and more often White (82.4% vs. 74.6%) compared with those receiving non-tirzepatide GLP-1 RAs (N = 105,239). They also had a higher prevalence of obesity (55.6% vs. 42.4%), sleep disorders (48.0% vs. 35.2%), and insulin use (23.3% vs. 37.1%), while showing lower rates of chronic kidney disease (18.3% vs. 23.5%), diabetic complications (neuropathy 10.1% vs. 16.0%; nephropathy 11.8% vs. 18.4%), and concomitant metformin therapy (22.8% vs. 32.6%).

After 1:1 PSM, 14,118 pairs of tirzepatide and GLP-1 users were successfully balanced across demographic, clinical, and pharmacologic characteristics (all ASMD < 0.1). Post-matching, the mean age was nearly identical between groups (64.3 vs. 64.0 years), sex distribution was balanced (female 43.8% vs. 44.6%), and comorbidity burdens such as hypertension (71.5% vs. 72.9%), chronic kidney disease (18.4% vs. 19.0%), and ischemic heart disease (32.8% vs. 33.4%) were comparable. Similarly, concomitant use of cardiovascular drugs (statins 51.5% vs. 52.6%; beta-blockers 58.7% vs. 59.4%) and anticoagulants (apixaban 32.8% vs. 33.5%) was well balanced. Key laboratory indices, including body mass index (37.9 vs. 38.1 kg/m²), glycated hemoglobin (6.7% vs. 6.8%), low-density lipoprotein cholesterol (86.4 vs. 86.0 mg/dL), and estimated glomerular filtration rate (71.2 vs. 70.4 mL/min/1.73 m²), were also comparable between groups after matching. In the propensity score–matched cohort, median follow-up was 0.5 years in the tirzepatide group and 0.9 years in the non–tirzepatide GLP-1 receptor agonist group (p < 0.001), reflecting the more recent clinical adoption of tirzepatide.

After matching, measured baseline characteristics were balanced between exposure groups, supporting comparability for the subsequent observational analyses.

Primary outcome: stroke

In the propensity-score-matched cohort, the cumulative incidence of all stroke was lower in the tirzepatide group than in the non-tirzepatide glucagon-like peptide-1 receptor agonist group (Table 2 and Fig. 2). The adjusted hazard ratio for any stroke was 0.71 (95% confidence interval, 0.58–0.87).

Table 2.

Clinical Outcomes after Propensity-Score Matching among Patients with AF and T2D Receiving Tirzepatide versus Non–Tirzepatide GLP-1 RAs

Outcome	Patients with Outcome, n (%)	Risk (%)	Incidence Rate per 10,000 PY	Hazard Ratio (95% CI)	Risk Difference (95% CI)	Risk Ratio (95% CI)	Odds Ratio (95% CI)	E-value (95% LCL)
All Stroke	Tirzepatide: 136/14,118 (1.0) GLP-1: 472/14,118 (3.3)	0.96 vs 3.34	129.0 vs 163.8	0.71 (0.58–0.87)	–2.38% (–2.72 to –2.04)	0.29 (0.24–0.35)	0.28 (0.23–0.34)	2.17 (1.56)
Hemorrhagic Stroke	Tirzepatide: 12/14,118 (0.1) GLP-1: 65/14,118 (0.5)	0.08 vs 0.46	11.4 vs 22.6	0.39 (0.21–0.74)	–0.38% (–0.50 to –0.25)	0.18 (0.10–0.34)	0.18 (0.10–0.34)	4.57 (2.04)
Ischemic Stroke	Tirzepatide: 133/14,118 (0.9) GLP-1: 434/14,118 (3.1)	0.94 vs 3.07	126.2 vs 150.6	0.77 (0.63–0.95)	–2.13% (–2.46 to –1.81)	0.31 (0.25–0.37)	0.30 (0.25–0.36)	1.92 (1.29)
Mortality	Tirzepatide: 204/14,118 (1.4) GLP-1: 1079/14,118 (7.6)	1.44 vs 7.64	193.5 vs 374.5	0.53 (0.45–0.62)	–6.20% (–6.68 to –5.72)	0.19 (0.16–0.22)	0.18 (0.15–0.21)	3.18 (2.61)

Risk denotes the proportion of patients with the outcome during follow-up. Incidence rates are per 10,000 person-years. Hazard ratios were estimated with Cox proportional-hazards models after propensity-score matching. Risk differences are expressed as absolute percentage differences between groups. The E-value represents the minimum strength of association that an unmeasured confounder would need to explain the observed effect.

AF atrial fibrillation, T2D type 2 diabetes, GLP-1 RAs glucagon-like peptide-1 receptor agonist, PY person-years, CI confidence interval, LCL lower confidence limit.

Fig. 2. Cumulative Incidence of All Stroke among Patients with Atrial Fibrillation and Type 2 Diabetes Receiving Tirzepatide versus Non-Tirzepatide GLP-1 RAs after Propensity-Score Matching.

Kaplan–Meier curves for all stroke in the 1:1 propensity-score-matched cohort of 14,118 patients per group. Curves were estimated using cause-specific hazards with death treated as a censoring event; death-restricted sensitivity analyses addressed competing mortality. Source data are provided as a Source Data file.

Secondary outcomes: mortality

All-cause mortality was lower among tirzepatide users than among non-tirzepatide glucagon-like peptide-1 receptor agonist users (Table 2 and Supplementary Fig. 1). Mortality occurred in 1.4% (204/14,118) of tirzepatide users and 7.6% (1,079/14,118) of comparator users, corresponding to an adjusted hazard ratio of 0.53 (95% confidence interval, 0.45–0.62) and an absolute risk difference of –6.20% (95% confidence interval, –6.68 to –5.72).

Specific stroke subtypes

The incidence of ischemic stroke was lower in the tirzepatide group than in the comparator group (Table 2 and Fig. 4; adjusted hazard ratio, 0.77; 95% confidence interval, 0.63–0.95). Hemorrhagic stroke was also less frequent in the tirzepatide group (Table 2 and Fig. 3; adjusted hazard ratio, 0.39; 95% confidence interval, 0.21–0.74).

Fig. 4. Cumulative Incidence of Ischemic Stroke among Patients with Atrial Fibrillation and Type 2 Diabetes Receiving Tirzepatide versus Non-Tirzepatide GLP-1 RAs.

Kaplan–Meier curves for ischemic stroke in the 1:1 propensity-score-matched cohort of 14,118 patients per group. Curves were estimated using cause-specific hazards with death treated as a censoring event; death-restricted sensitivity analyses addressed competing mortality. Source data are provided as a Source Data file.

Fig. 3. Cumulative Incidence of Hemorrhagic Stroke among Patients with Atrial Fibrillation and Type 2 Diabetes Receiving Tirzepatide versus Non-Tirzepatide GLP-1 RAs.

Kaplan–Meier curves for hemorrhagic stroke in the 1:1 propensity-score-matched cohort of 14,118 patients per group. Curves were estimated using cause-specific hazards with death treated as a censoring event; death-restricted sensitivity analyses addressed competing mortality. Source data are provided as a Source Data file.

E-values quantify the strength of unmeasured confounding that would be required to move the observed associations to the null (Table 2).

Associations across primary, secondary, and control outcomes

In the propensity-score-matched cohort, tirzepatide use was associated with lower risks of stroke and all-cause mortality than non-tirzepatide glucagon-like peptide-1 receptor agonist use. Associations were also observed for ischemic and hemorrhagic stroke. Positive outcome controls, including major adverse cardiovascular events and major adverse kidney events, showed estimates below the null, whereas negative outcome controls (burn injury, traumatic brain injury, and suicide attempt or ideation) were not associated with treatment. The positive exposure control (metformin) was associated with lower stroke risk, whereas the negative exposure control (other antiseptics and disinfectants) was not associated with stroke (Fig. 5).

Fig. 5. Adjusted Associations of Tirzepatide versus Non–Tirzepatide GLP-1 RAs across Primary, Secondary, Specificity, and Control Outcomes in Patients with Atrial Fibrillation and Type 2 Diabetes.

Cox proportional-hazards models estimated adjusted hazard ratios (aHRs) and 95% confidence intervals (CIs) in the 1:1 propensity-score-matched cohort of 14,118 patients per group (28,236 distinct patients). Squares denote aHR point estimates and horizontal lines denote 95% CIs; the vertical reference line denotes an aHR of 1. Two-sided confidence intervals are shown. Outcomes were classified as primary (stroke), secondary (all-cause mortality), specificity outcomes (hemorrhagic and ischemic stroke), positive outcome controls (major adverse cardiovascular events and major adverse kidney events), negative outcome controls (burn injury, traumatic brain injury, suicide attempt or ideation), positive exposure control (metformin), and negative exposure control (other antiseptics or disinfectants). Source data are provided as a Source Data file.

Subgroup analyses

Across subgroup analyses defined by age, sex, obesity, renal function, and lipid status, most point estimates for stroke were below the null. Confidence intervals were wider in smaller strata; these estimates should therefore be interpreted with appropriate caution (Supplementary Tables 2–5).

Sensitivity analyses

Sensitivity analyses across prespecified intervals from atrial fibrillation diagnosis to treatment initiation showed directionally similar associations between tirzepatide use and reduced stroke risk (Supplementary Table 6). Analyses restricted to survivors across successive follow-up intervals showed estimates that remained below the null (Supplementary Table 7).

Additional analyses using a strict new-user design, an on-treatment design with a 90-day grace period, stricter stroke definitions, and a restricted cohort of baseline direct oral anticoagulant users produced estimates broadly consistent with the primary analysis (Supplementary Tables 10–12).

For clinical interpretation, absolute risk differences corresponded to numbers needed to treat of 48 at one year and 22 at three years for any stroke; 63 and 30 for ischemic stroke; 200 and 95 for hemorrhagic stroke; and 16 and 8 for all-cause mortality, respectively (Supplementary Table 13). E-values for the point estimates ranged from 1.92 to 4.57, with values from 1.29 to 2.61 for the confidence interval limit nearest the null.

DISCUSSION

In this real-world cohort of patients with atrial fibrillation and type 2 diabetes, tirzepatide use was associated with lower risks of stroke and mortality than use of non-tirzepatide glucagon-like peptide-1 receptor agonists. An active-comparator, new-user design, matching on demographic, clinical, medication, and laboratory variables, and sensitivity analyses were used to address measured confounding. Residual confounding, including channeling of newer therapies to patients with different clinical profiles, cannot be excluded. Associations for hemorrhagic stroke should be interpreted cautiously because of the limited number of events and the resulting uncertainty.

Prior randomized trials have generally compared glucagon-like peptide-1 receptor agonists with placebo and commonly reported composite cardiovascular outcomes or overall stroke.3,20,21,27–32 Existing real-world studies have suggested heterogeneity between agents, but cerebrovascular outcomes by stroke subtype have been less well characterized.23–26,33 The present analysis provides comparative observational estimates for tirzepatide and established glucagon-like peptide-1 receptor agonists in a high-risk population.

Tirzepatide, as a dual GIP/GLP-1 RA, may lower stroke risk in patients with AF and T2D through complementary pathways. Improved glycemic control, weight reduction, mitigation of microvascular injury, enhancement of endothelial function, and lowering of blood pressure are particularly relevant for hemorrhagic stroke^30,38–40, in which chronic hyperglycemia and vascular fragility amplify bleeding risk; indeed, every 5 mmHg reduction in systolic blood pressure is associated with a 20–25% lower risk of hemorrhagic stroke⁴¹. In parallel, tirzepatide exerts favorable effects on blood pressure, lipid metabolism, and inflammation, lowering systolic blood pressure by 4–6 mmHg, reducing LDL and triglycerides, raising HDL, and attenuating markers such as hsCRP and ICAM-1^30,42–46. These changes mitigate atherosclerosis and thrombogenesis, thereby reducing ischemic stroke incidence^30,42–46. Tirzepatide also improves renal outcomes by lowering albuminuria and slowing eGFR decline, which may indirectly decrease cerebrovascular risk through better hemodynamic and vascular regulation^{33,39,45,47–49}. Beyond GLP-1 effects, GIP receptor agonism confers additional benefits, including enhanced insulin secretion, improved adipocyte function, promotion of lipid clearance, and stabilization of vascular endothelium^{30,38,39,40,42–46}. These mechanisms not only augment metabolic control but may also contribute directly to vascular integrity. Collectively, tirzepatide provides biologic plausibility for protection against both ischemic and hemorrhagic stroke, extending its benefits beyond traditional GLP-1 RAs. While AF is increasingly characterized as an atrial cardiomyopathy driving cardioembolic risk, and T2D primarily accelerates atherosclerosis and small-vessel disease, patients with both conditions face synergistic cerebrovascular risks^12,15,50. Our findings suggest that tirzepatide’s broad pleiotropic effects—ranging from metabolic and anti-inflammatory actions to potential hemodynamic benefits—may mitigate stroke risk across these distinct but overlapping pathogenic pathways. Future research in populations with broader indications (e.g., obesity or heart failure) or longer follow-up may further elucidate the agent’s impact on isolated embolic versus atherosclerotic mechanisms. To address potential heterogeneity within the GLP-1RA class, we additionally conducted exploratory head-to-head comparisons between tirzepatide and each individual GLP-1RA. Across these analyses, point estimates consistently favored tirzepatide and were directionally concordant with the primary pooled analysis. For less frequently prescribed agents, confidence intervals were wider due to limited event counts, and these findings should therefore be interpreted as exploratory and hypothesis-generating rather than definitive evidence of drug-specific superiority.

An important methodological consideration is the shorter follow-up duration observed among tirzepatide users, reflecting its more recent clinical adoption. In the propensity score–matched population, the median follow-up was 0.5 years in the tirzepatide group compared with 0.9 years in the non–tirzepatide GLP-1 receptor agonist group. While shorter follow-up could theoretically limit event accrual, the present analyses were designed to estimate hazard-based associations rather than cumulative incidence. Accordingly, Cox proportional-hazards models were employed, which inherently account for right censoring and variable follow-up durations. Importantly, the observed protective association of tirzepatide was evident early after treatment initiation: in a sensitivity analysis restricted to the first 0–6 months of follow-up, tirzepatide use remained significantly associated with a lower risk of stroke (adjusted HR 0.91; 95% CI, 0.66–0.99). In addition, death-restricted and latency-stratified analyses yielded directionally consistent results, arguing against spurious effects driven by differential follow-up duration.

Beyond methodological considerations, weight reduction is a biologically plausible contributor to the observed cerebrovascular benefit of tirzepatide. In the present study, baseline body mass index was well balanced after propensity-score matching (37.9 vs. 38.1 kg/m²; ASMD < 0.1), and subgroup analyses stratified by baseline BMI demonstrated consistent associations between tirzepatide use and lower stroke risk (Supplementary Table 5). Importantly, our study was designed to estimate the total cerebrovascular effect of tirzepatide rather than to decompose individual mediating pathways. Because weight change occurs after treatment initiation and may lie on the causal pathway between exposure and stroke risk, formal adjustment for longitudinal weight change in this real-world dataset could introduce bias rather than clarify the observed association.

Associations were generally consistent across subgroups and sensitivity analyses, including intention-to-treat, latency-window, death-restricted, new-user, on-treatment, and stricter outcome-definition analyses (Supplementary Tables 2–13). These analyses do not eliminate the possibility of residual confounding or outcome misclassification, but they provide complementary assessments of the stability of the primary estimates.

This study used a large multicenter electronic health record network and a head-to-head active-comparator design. Several limitations warrant emphasis. The target population was restricted to adults with both atrial fibrillation and type 2 diabetes to improve event accrual; the findings should not be broadly extrapolated to lower-risk populations with type 2 diabetes receiving routine diabetes care. In a descriptive feasibility audit, the broader population without atrial fibrillation had sparse stroke events and shorter follow-up among tirzepatide initiators (Supplementary Table 14). Differential calendar time, shorter available follow-up for tirzepatide, residual confounding, diagnostic-code misclassification, and limited hemorrhagic stroke events remain important considerations.

In this observational, head-to-head comparison of adults with atrial fibrillation and type 2 diabetes, tirzepatide use was associated with lower risks of ischemic stroke, hemorrhagic stroke, and all-cause mortality than use of non-tirzepatide glucagon-like peptide-1 receptor agonists. Randomized cardiovascular outcomes trials are needed before these associations can inform treatment recommendations.

METHODS

Data source

This study complied with all relevant ethical regulations. The TriNetX Global Collaborative Network is a federated research platform that provides only de-identified, aggregate electronic health record outputs from participating healthcare organizations. The investigators did not access identifiable or linkable individual-level records, did not recruit or contact participants, and did not obtain data from their own institutions. Accordingly, this secondary analysis of fully de-identified, non-linkable aggregate data did not require local Institutional Review Board review or informed consent under the institutional policies applicable to research that does not involve identifiable human-participant data. No participant compensation was provided. The database captures longitudinal information on demographics, diagnoses, procedures, medication exposures, laboratory results, and healthcare utilization. At the time of analysis, the Global Collaborative Network encompassed 152 healthcare organizations and more than 150 million patients worldwide.

Study population

Eligible participants were adults aged 18 years or older with atrial fibrillation (International Classification of Diseases, Tenth Revision, Clinical Modification code I48) and type 2 diabetes (code E11). We selected this high-risk phenotype to ensure adequate cerebrovascular event accrual within the currently available follow-up for tirzepatide users. All eligible patients meeting the prespecified cohort criteria between 1 January 2020 and 20 November 2024 were included; no formal sample-size calculation was performed. Patients initiated tirzepatide (Anatomical Therapeutic Chemical code A10BX16; RxNorm 2601723) or a non-tirzepatide glucagon-like peptide-1 receptor agonist (Anatomical Therapeutic Chemical code A10BJ, excluding fixed insulin/glucagon-like peptide-1 receptor agonist combinations). The comparator cohort included semaglutide, dulaglutide, liraglutide, exenatide, and lixisenatide. Fixed-dose insulin/glucagon-like peptide-1 receptor agonist combinations were excluded. In both cohorts, type 2 diabetes had to precede the index date. We excluded patients with stroke or outcome events before atrial fibrillation diagnosis or treatment initiation, those exposed to study drugs before atrial fibrillation diagnosis, and those receiving both study exposure categories. Recorded sex was treated as a biological variable and included in matching and subgroup analyses; gender identity was not consistently available in the contributing electronic health record systems.

The index date for the tirzepatide group was the first prescription date of tirzepatide following AF diagnosis; for the comparator group, it was the first prescription date of a non-tirzepatide GLP-1 RA after AF diagnosis. In both cohorts, treatment initiation occurred after diagnoses of both AF and T2D, thereby ensuring a uniform temporal anchor for index date definition. Events occurring between AF diagnosis and treatment initiation were uncommon relative to the analytic cohort size and were excluded symmetrically by design in both treatment groups, thereby minimizing the potential for differential immortal time bias. A 12-month baseline period before the index date was used to ascertain comorbidities, laboratory results, and concomitant medications (Table 1). Patients were followed until the earliest of the outcome event, death, loss to follow-up, or the end of the study period, with the last follow-up set at August 20, 2025. Because AF diagnosis dates in real-world EHR data may be heterogeneous and repeatedly recorded, latency between AF diagnosis and treatment initiation was evaluated using prespecified clinically interpretable time windows rather than a single summary statistic. Follow-up duration was defined from the index date to the earliest occurrence of the outcome of interest, death, loss to follow-up, or end of the study period. The detailed process of cohort identification, application of exclusion criteria, and derivation of the final analytic sample before and after propensity-score matching is illustrated in Fig. 1. Differences in follow-up time between treatment groups were explicitly quantified and addressed using time-to-event methods.

Table 1.

Baseline Characteristics of Patients with Atrial Fibrillation and Type 2 Diabetes Receiving Tirzepatide or Non–Tirzepatide GLP‑1 RAs, Before and After Propensity‑Score Matching

Characteristic	Before Matching				After Matching
	Tirzepatide (N = 14,126)	GLP-1 RAs* (N = 105,239)	ASMD	P	Tirzepatide (N = 14,118)	GLP-1 RAs* (N  = 14,118)	ASMD	P
Demographics
Age at Index
Age at index — mean (SD), yr	64.3 (11.3)	66.7 (11.2)	0.211	<0.001	64.3 (11.3)	64.0 (11.6)	0.032	0.027
Sex
Male	7719 (54.6)	60,636 (57.6)	0.060	<0.001	7715 (54.6)	7563 (53.6)	0.022	0.082
Female	6184 (43.8)	41,852 (39.8)	0.081	<0.001	6180 (43.8)	6297 (44.6)	0.017	0.177
Unknown Gender	223 (1.6)	2751 (2.6)	0.072	<0.001	223 (1.6)	258 (1.8)	0.019	0.109
Race
White	11,644 (82.4)	78,507 (74.6)	0.192	<0.001	11,636 (82.4)	11,533 (81.7)	0.019	0.130
Black or African American	1226 (8.7)	11,759 (11.2)	0.083	<0.001	1226 (8.7)	1272 (9.0)	0.011	0.354
Asian	181 (1.3)	2705 (2.6)	0.094	<0.001	181 (1.3)	194 (1.4)	0.008	0.505
Native Hawaiian or Other Pacific Islander	68 (0.5)	794 (0.8)	0.035	<0.001	68 (0.5)	67 (0.5)	0.001	0.931
American Indian or Alaska Native	33 (0.2)	371 (0.4)	0.022	0.005	33 (0.2)	26 (0.2)	0.011	0.358
Other Race	271 (1.9)	2,913 (2.8)	0.056	<0.001	271 (1.9)	258 (1.8)	0.007	0.570
Unknown Race	703 (5.0)	8,190 (7.8)	0.115	<0.001	703 (5.0)	768 (5.4)	0.021	0.091
Comorbidities
Hypertension	10,104 (71.5)	72,273 (68.7)	0.062	<0.001	10,097 (71.5)	10,292 (72.9)	0.031	0.009
Dyslipidemia	9025 (63.9)	62,528 (59.4)	0.092	<0.001	9017 (63.9)	9127 (64.6)	0.016	0.191
Obesity	7859 (55.6)	44,605 (42.4)	0.267	<0.001	7852 (55.6)	7947 (56.3)	0.014	0.244
Systemic connective tissue disorders	370 (2.6)	2006 (1.9)	0.048	<0.001	368 (2.6)	385 (2.7)	0.007	0.531
Acute kidney failure	1486 (10.5)	14,918 (14.2)	0.111	<0.001	1486 (10.5)	1530 (10.8)	0.010	0.408
Chronic kidney disease (CKD)	2592 (18.3)	24,742 (23.5)	0.127	<0.001	2591 (18.4)	2687 (19.0)	0.017	0.160
Any cancer	3908 (27.7)	24,298 (23.1)	0.105	<0.001	3902 (27.6)	3947 (28.0)	0.007	0.534
Type 2 Diabetes with Nephropathy	1664 (11.8)	19,410 (18.4)	0.187	<0.001	1664 (11.8)	1772 (12.6)	0.023	0.049
Type 2 Diabetes with Retinopathy	523 (3.7)	6587 (6.3)	0.118	<0.001	523 (3.7)	561 (4.0)	0.014	0.249
Type 2 Diabetes with Neuropathy	1420 (10.1)	16,792 (16.0)	0.176	<0.001	1,420 (10.1)	1475 (10.4)	0.013	0.284
Type 2 Diabetes with Circulatory Complications	880 (6.2)	9097 (8.6)	0.092	<0.001	879 (6.2)	959 (6.8)	0.023	0.052
Chronic Lower Respiratory Disease	3452 (24.4)	24,014 (22.8)	0.038	<0.001	3448 (24.4)	3470 (24.6)	0.004	0.732
Nicotine dependence	1087 (7.7)	8794 (8.4)	0.024	0.008	1087 (7.7)	1164 (8.2)	0.020	0.098
Tobacco use	358 (2.5)	2970 (2.8)	0.018	0.057	358 (2.5)	383 (2.7)	0.011	0.358
Alcohol-Related Disorder	454 (3.2)	2718 (2.6)	0.038	<0.001	453 (3.2)	427 (3.0)	0.011	0.378
Heart failure	3759 (26.6)	31,211 (29.7)	0.068	<0.001	3755 (26.6)	3869 (27.4)	0.018	0.134
Peripheral Vascular Disease	851 (6.0)	7785 (7.4)	0.055	<0.001	851 (6.0)	890 (6.3)	0.011	0.347
Cerebrovascular Disease	824 (5.8)	6871 (6.5)	0.029	0.001	823 (5.8)	828 (5.9)	0.002	0.901
Ischemic Heart Disease	4639 (32.8)	37,891 (36.0)	0.067	<0.001	4637 (32.8)	4719 (33.4)	0.012	0.288
Sleep disorders	6783 (48.0)	37,045 (35.2)	0.262	<0.001	6775 (48.0)	6737 (47.7)	0.005	0.634
Liver Disease	1,518 (10.7)	9,991 (9.5)	0.042	<0.001	1516 (10.7)	1517 (10.7)	0.000	0.985
Medications
Insulins and Analogues	3293 (23.3)	39,013 (37.1)	0.303	<0.001	3293 (23.3)	3333 (23.6)	0.007	0.575
Metformin	3222 (22.8)	34,274 (32.6)	0.219	<0.001	3221 (22.8)	3324 (23.5)	0.017	0.165
Sulfonylureas	1008 (7.1)	16,895 (16.1)	0.281	<0.001	1008 (7.1)	992 (7.0)	0.004	0.718
Alpha glucosidase inhibitors	11 (0.1)	289 (0.3)	0.047	<0.001	11 (0.1)	11 (0.1)	0.000	1.000
Thiazolidinediones	193 (1.4)	2997 (2.8)	0.103	<0.001	193 (1.4)	183 (1.3)	0.006	0.603
DPP-4 Inhibitors	596 (4.2)	10,687 (10.2)	0.231	<0.001	596 (4.2)	602 (4.3)	0.002	0.864
SGLT2 Inhibitors	2407 (17.0)	17,621 (16.7)	0.008	0.354	2402 (17.0)	2481 (17.6)	0.015	0.203
Other Glucose-Lowering Drugs, Excluding Insulin	63 (0.4)	1087 (1.0)	0.069	<0.001	63 (0.4)	71 (0.5)	0.008	0.488
Apixaban	4,641 (32.9)	27,249 (25.9)	0.153	<0.001	4,635 (32.8)	4723 (33.5)	0.013	0.264
Rivaroxaban	1514 (10.7)	11,091 (10.5)	0.006	0.505	1513 (10.7)	1610 (11.4)	0.022	0.068
Edoxaban	11 (0.1)	840 (0.8)	0.109	<0.001	11 (0.1)	15 (0.1)	0.009	0.432
Dabigatran etexilate	114 (0.8)	1476 (1.4)	0.057	<0.001	114 (0.8)	94 (0.7)	0.017	0.165
Warfarin	753 (5.3)	10,410 (9.9)	0.173	<0.001	753 (5.3)	770 (5.5)	0.005	0.667
Enoxaparin	1641 (11.6)	14,340 (13.6)	0.061	<0.001	1641 (11.6)	1676 (11.9)	0.008	0.508
Amiodarone	1385 (9.8)	10,450 (9.9)	0.004	0.767	1,382 (9.8)	1422 (10.1)	0.009	0.449
Mexiletine	43 (0.3)	300 (0.3)	0.004	0.817	43 (0.3)	41 (0.3)	0.003	0.826
Lidocaine	5990 (42.4)	38,055 (36.2)	0.128	<0.001	5982 (42.4)	6071 (43.0)	0.013	0.297
Flecainide	734 (5.2)	3,028 (2.9)	0.118	<0.001	731 (5.2)	722 (5.1)	0.003	0.812
Dronedarone	180 (1.3)	1,010 (1.0)	0.030	0.001	180 (1.3)	188 (1.3)	0.005	0.676
Statins	7277 (51.5)	57,074 (54.2)	0.054	<0.001	7274 (51.5)	7428 (52.6)	0.022	0.067
Ezetimibe	812 (5.7)	4,881 (4.6)	0.050	<0.001	806 (5.7)	821 (5.8)	0.005	0.707
Evolocumab	218 (1.5)	859 (0.8)	0.067	<0.001	213 (1.5)	222 (1.6)	0.005	0.665
Alirocumab	56 (0.4)	346 (0.3)	0.011	0.222	56 (0.4)	56 (0.4)	0.000	1.000
Beta-Blockers	8293 (58.7)	60,535 (57.5)	0.024	0.007	8286 (58.7)	8383 (59.4)	0.014	0.236
Renin–Angiotensin System Agents	6758 (47.8)	51,936 (49.4)	0.030	<0.001	6750 (47.8)	6832 (48.4)	0.012	0.320
Diuretics	6662 (47.2)	52,195 (49.6)	0.049	<0.001	6,657 (47.2)	6,782 (48.0)	0.018	0.134
Calcium-Channel Blockers	4959 (35.1)	35,954 (34.2)	0.020	0.027	4,951 (35.1)	5,057 (35.8)	0.016	0.190
Alpha-Adrenergic Antagonists	251 (1.8)	2230 (2.1)	0.025	0.008	251 (1.8)	251 (1.8)	0.000	1.000
Organic nitrates	2,029 (14.4)	18,963 (18.0)	0.099	<0.001	2029 (14.4)	2068 (14.6)	0.008	0.501
Ranolazine	123 (0.9)	1,323 (1.3)	0.038	<0.001	123 (0.9)	136 (1.0)	0.010	0.418
Spironolactone	1710 (12.1)	12,685 (12.1)	0.002	0.825	1708 (12.1)	1743 (12.3)	0.008	0.515
Eplerenone	125 (0.9)	846 (0.8)	0.009	0.364	125 (0.9)	126 (0.9)	0.001	0.949
Vericiguat	10 (0.1)	47 (0.0)	0.011	0.285	10 (0.1)	11 (0.1)	0.003	0.827
Digitalis glycosides	430 (3.0)	5766 (5.5)	0.121	<0.001	430 (3.0)	446 (3.2)	0.007	0.551
Ivabradine	45 (0.3)	1590 (1.5)	0.125	<0.001	45 (0.3)	50 (0.4)	0.006	0.608
Antiplatelet Agents (non–heparin)
Aspirin	3414 (24.2)	30,298 (28.8)	0.105	<0.001	3412 (24.2)	3492 (24.7)	0.013	0.291
Clopidogrel	939 (6.6)	9710 (9.2)	0.096	<0.001	939 (6.7)	949 (6.7)	0.003	0.814
Ticagrelor	89 (0.6)	1327 (1.3)	0.065	<0.001	89 (0.6)	107 (0.8)	0.015	0.198
Prasugrel	50 (0.4)	554 (0.5)	0.026	0.007	50 (0.4)	45 (0.3)	0.006	0.605
Cangrelor	10 (0.1)	58 (0.1)	0.006	0.548	10 (0.1)	10 (0.1)	0.000	1.000
Laboratory Values
Body-mass index-mean (SD)	37.9 (7.8)	36.4 (8.2)	0.190	<0.001	37.9 (7.8)	38.1 (8.0)	0.028	0.019
HbA1c-mean (SD), %	6.7 (1.6)	7.7 (1.9)	0.565	<0.001	6.7 (1.6)	6.8 (1.7)	0.072	<0.001
LDL cholesterol-mean (SD)	86.3 (35.5)	81.3 (35.6)	0.140	<0.001	86.4 (35.5)	86.0 (36.0)	0.011	0.352
eGFR-mean (SD)	71.2 (23.8)	66.7 (26.3)	0.180	<0.001	71.2 (23.8)	70.4 (24.7)	0.035	0.004
Albumin-ean (SD)	4.0 (0.5)	3.9 (0.5)	0.200	<0.001	4.0 (0.5)	4.0 (0.5)	0.027	0.023

GLP-1 RA glucagon-like peptide-1 receptor agonists, ASMD absolute standardized mean difference, SD standard deviation, CKD chronic kidney disease, DOAC direct oral anticoagulant, DPP-4 dipeptidyl peptidase-4, SGLT2 sodium–glucose cotransporter-2, LDL low-density lipoprotein, HbA1c hemoglobin A1c, eGFR estimated glomerular filtration rate.

P values were generated directly by the TriNetX analytics platform using two-sided chi-square tests for categorical variables and two-sided Student’s t tests for continuous variables; they are provided for descriptive purposes only. Exact P values are reported where available; values below the platform reporting threshold are shown as P < 0.001. Covariate balance after matching was primarily assessed using absolute standardized mean differences, with values < 0.1 indicating adequate balance.

Fig. 1. STROBE-compliant patient flow diagram.

The diagram illustrates cohort identification, exclusion criteria, propensity-score matching, and the final analytic cohorts of 14,118 patients in each treatment group. Source data are provided as a Source Data file. Abbreviations: AF, atrial fibrillation; GLP-1 RA, glucagon-like peptide-1 receptor agonist; ICD-10, International Classification of Diseases, 10th Revision; PSM, propensity score matching; SD, standard deviation; T2D, type 2 diabetes.

Outcome measures

The primary outcome was the occurrence of any stroke (ischemic or hemorrhagic). Secondary outcomes included ischemic stroke, hemorrhagic stroke, and all-cause mortality (Table 2; Figs. 2–4 and Supplementary Fig. 1). Stroke outcomes were defined using International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) diagnosis codes. Ischemic stroke was defined using I63-series codes (I63.0–I63.9), whereas hemorrhagic stroke was defined as nontraumatic intracranial hemorrhage, including subarachnoid hemorrhage (I60.0–I60.9), intracerebral hemorrhage (I61.0–I61.9), and other nontraumatic intracranial hemorrhage (I62.0–I62.9). To avoid outcome misclassification, all traumatic intracranial injury codes (S06-series), including traumatic subdural hemorrhage (S06.5), were explicitly excluded from hemorrhagic stroke definitions. A complete list of ICD-10-CM codes used for stroke outcome ascertainment is provided in Supplementary Table 1.

To assess specificity, we examined outcomes not biologically related to GLP-1 RAs or tirzepatide. Multiple negative control outcomes were prespecified, including burn injury, traumatic brain injury, and suicide attempt or ideation (Fig. 5), each selected to probe different dimensions of potential residual confounding. Burn injury was included as a representative negative control because it lacks any plausible biological or pharmacologic link to incretin-based therapies and is primarily determined by accidental or environmental exposures, while being reliably captured in electronic health records. Major adverse cardiovascular events (MACE) and major adverse kidney events (MAKE) were prespecified as positive outcome controls, with traumatic brain injury and suicide attempt or ideation serving as additional negative control outcomes (Fig. 5).

Exposure definitions

The exposure of interest was treatment with tirzepatide compared with non-tirzepatide GLP-1 RAs. Exposure was defined as initiation of the study drug after AF diagnosis within the study period. Although ATC code A10BJ was used for initial cohort identification, final exposure classification was based on agent-level medication records mapped using RxNorm within the TriNetX platform. To avoid treatment contamination, patients in the tirzepatide group were not allowed to receive non-tirzepatide GLP-1 RAs during follow-up, and vice versa. This crossover prohibition was enforced in the primary analysis.

Covariates

Baseline covariates, measured during the 12-month pre-index period, included demographics (age, sex, race); comorbidities (hypertension, dyslipidemia, obesity, chronic kidney disease, ischemic heart disease, heart failure, tobacco use, and alcohol use); concomitant medications (metformin, sodium–glucose cotransporter-2 [SGLT2] inhibitors, statins, diuretics, beta-adrenergic blocking agents, calcium channel blocking agents, agents acting on the renin–angiotensin system, antiplatelet agents, and anticoagulants); and laboratory values (body mass index [BMI], glycated hemoglobin [HbA1c], low-density lipoprotein [LDL] cholesterol, and estimated glomerular filtration rate [eGFR]). Longitudinal weight change during follow-up was not modeled because repeated weight measurements were not systematically captured across the TriNetX network, and formal adjustment for post-baseline weight could introduce selection bias or inappropriate control for mediators. These are presented in Table 1 and expanded in Supplementary Tables 2–5 (subgroup analyses).

TriNetX provides aggregate analytic outputs rather than downloadable individual-level records. Variable-specific counts of missing laboratory observations were not available to the authors in the aggregate output, and no author-implemented imputation was performed. Matching and outcome analyses were executed within the platform using the available data fields.

Sensitivity analyses

We conducted several prespecified sensitivity analyses to assess the robustness of our findings (Supplementary Tables 6–7). In addition, we performed exploratory head-to-head sensitivity analyses comparing tirzepatide with each individual GLP-1 receptor agonist (semaglutide, dulaglutide, liraglutide, exenatide, and lixisenatide) using the same propensity-score–matched framework as the primary analysis (Supplementary Table 8). Agent-level distributions of non-tirzepatide GLP-1RAs before and after propensity-score matching are reported in Supplementary Table 9 to address GLP-1RA heterogeneity. First, in an intention-to-treat (ITT) framework, patients were permitted to crossover between tirzepatide and non-tirzepatide GLP-1 RAs during follow-up to evaluate whether treatment switching influenced effect estimates. We examined latency between AF diagnosis and treatment initiation using prespecified strata (≤6, ≤12, ≤18, ≤24, and >24 months) to assess potential confounding by AF disease duration and treatment timing (Supplementary Table 6).

Because death represents a major competing risk for stroke in this high-risk population, and differential mortality could potentially attenuate observed stroke incidence, we prespecified strategies to mitigate survival bias. Although Fine–Gray subdistribution hazard models would ideally address competing risks, the TriNetX platform does not currently support such estimation. To approximate this, we performed death-restricted analyses that sequentially limited follow-up to patients who survived through successive intervals (0–6, 6–12, 12–18, 18–24, and ≥24 months). In parallel, we calculated cause-specific hazard ratios using Cox proportional-hazards models, treating death as a censoring event, to estimate the instantaneous risk of stroke among patients still alive. The consistency of stroke risk estimates across these death-restricted analyses supports the conclusion that the observed association is unlikely to be an artifact of competing mortality.

Additional analyses included a strict new-user design requiring no tirzepatide or GLP-1 RA prescriptions during the 12-month baseline, with an extended 18-month washout to address residual exposure carry-over (Supplementary Table 10). An on-treatment analysis incorporated a 90-day grace period and censored patients at discontinuation or crossover to evaluate the effect of persistence with therapy (Supplementary Table 11). To address outcome misclassification, we applied a stricter stroke definition, restricting events to inpatient primary diagnoses and, secondarily, to those accompanied by neuroimaging procedure codes (Supplementary Table 12). We also performed a restricted cohort analysis among baseline DOAC users to account for anticoagulation effects (Supplementary Table 10, secondary analysis). Finally, to enhance clinical interpretability, we calculated absolute risk reductions and numbers needed to treat at one and three years for major outcomes, complementing hazard-ratio estimates (Supplementary Table 13).

Statistical analysis

To minimize measured confounding, we performed 1:1 propensity-score matching using demographic, clinical, medication, and laboratory covariates, with a caliper of 0.2 standard deviations of the logit of the propensity score. Covariate balance was assessed using absolute standardized mean differences, with values below 0.1 considered adequately balanced (Table 1). For descriptive baseline comparisons, the TriNetX Analytics Platform used two-sided chi-square tests for categorical variables and two-sided Student’s t tests for continuous variables; exact P values were reported where available and values below the platform reporting threshold were reported as P < 0.001. Cox proportional-hazards models estimated adjusted hazard ratios and 95% confidence intervals for time-to-event outcomes. Kaplan–Meier curves and cumulative incidence plots treated death as a censoring event; death-restricted sensitivity analyses addressed competing mortality. Prespecified subgroup and sensitivity analyses are described above. No multiplicity adjustment was applied to subgroup or control analyses, which should be interpreted as supportive and exploratory. E-values were calculated according to VanderWeele’s method using reciprocal hazard ratios for estimates below 1.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

C.L.L., K.C.L., and J.M.Y. contributed to the conception and design of the study, clinical interpretation of the findings, and drafting of the manuscript. C.L.C. contributed to the literature review, clinical interpretation, and critical revision of the manuscript. Y.Y.L. contributed to data collection and statistical analysis. C.Y.F. contributed to methodology and data collection. W.M.C. contributed to data curation, formal analysis, visualization, and interpretation of the results. S.Y.W. conceived and supervised the study, developed the methodology, acquired funding, and critically revised the manuscript. All authors read and approved the final manuscript and agree to be accountable for all aspects of the work.

Peer review

Peer review information

Nature Communications thanks Sean O’Leary, Shuai Yuan and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

This work was supported by the Lo-Hsu Medical Foundation, Lotung Poh-Ai Hospital (grants 11403 and 11404), which supported S.-Y.W.

Data availability

The aggregated, de-identified electronic health record data analyzed in this study were obtained through the TriNetX Global Collaborative Network. Individual-level data are not available to the authors or the public under the TriNetX data use agreement and applicable privacy requirements. Eligible researchers may apply directly to TriNetX for access through an authorized institution and for a defined research purpose; access is governed by TriNetX contractual, data-governance, and privacy requirements, and the authors cannot independently grant access. Source data are provided with this paper.

Code availability

No custom code was used. Cohort identification, propensity-score matching, and statistical analyses were conducted with the built-in tools of the TriNetX Analytics Platform. The proprietary platform code is not publicly available; authorized users may reproduce the analytic workflow using the cohort definitions and methods reported here.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77047-5.