Heterogeneity of Cardiovascular Effects of Second‐Line Glucose‐Lowering Therapies in Adults With Type 2 Diabetes Across the Range of Moderate Baseline Cardiovascular Risk

Abstract

Background

Glucagon‐like peptide‐1 receptor agonists (GLP‐1RAs) and sodium–glucose cotransporter‐2 inhibitors (SGLT2is) have favorable cardiovascular outcomes compared with dipeptidyl peptidase‐4 inhibitors (DPP4is) and sulfonylureas in adults with type 2 diabetes and high cardiovascular risk. How these benefits vary across lower levels of cardiovascular risk is unknown.

Methods

We used nationwide claims data to emulate a comparative effectiveness trial and examine the heterogeneity of treatment effects of GLP‐1RAs, SGLT2is, DPP4is, and sulfonylureas on major adverse cardiovascular events (MACEs) among adults with type 2 diabetes and moderate cardiovascular risk (annualized MACE risk 1%–5%, estimated using the annualized claims‐based MACE estimator).

Results

Among 386 276 included adults with type 2 diabetes, 25.2% had baseline ACME–predicted MACE risk >1% to ≤2% (lower‐risk patients) and 13.3% had ACME–predicted risk >4% to ≤5% (higher‐risk patients). By year 3 of treatment, higher‐risk patients derived greater absolute benefit than lower‐risk patients when treated with GLP‐1RAs versus sulfonylureas (absolute reduction in the estimated rate of MACE of 3.1% in higher‐risk patients and 1.6% in lower‐risk patients), SGLT2is versus sulfonylureas (absolute reduction, 3.9% in higher‐risk patients and 1.3% in lower‐risk patients), and GLP‐1RAs versus DPP4is (absolute reduction, 1.6% in higher‐risk patients and 0.5% in lower‐risk patients). The relative benefits for MACE were also greater in higher‐risk than lower‐risk patients with SGLT2is versus DPP4is (hazard ratio [HR], 0.78 [95% CI, 0.70–0.87] in higher‐risk patients; HR, 0.99 [95% CI, 0.88–1.12] in lower‐risk patients). Conversely, the relative benefits of DPP4is and GLP‐1RAs versus sulfonylureas were greater in lower‐risk patients: HR 0.76 (95% CI, 0.71–0.81) in lower‐risk and HR 0.91 (95% CI, 0.97–0.96) in higher‐risk patients for DPP4is versus sulfonylureas; HR 0.67 (95% CI, 0.58–0.78) in lower‐risk and HR 0.80 (95% CI, 0.70–0.93) in higher‐risk patients for GLP‐1RAs versus sulfonylurea. Benefits of SGLT2is and GLP‐1RAs were comparable across all risk levels.

Conclusions

Cardiovascular benefits of SGLT2is and GLP‐1RAs exist across all levels of moderate cardiovascular risk, reinforcing the importance of choosing glucose‐lowering therapies that can prevent MACE in all people with type 2 diabetes.

Nonstandard Abbreviations and Acronyms

ACME

annualized claims‐based major adverse cardiovascular event estimator

DPP4i

dipeptidyl peptidase‐4 inhibitor

GLP‐1RA

glucagon‐like peptide‐1 receptor agonist

GRADE

Glycemia Reduction Approaches in Type 2 Diabetes: A Comparative Effectiveness

HHF

hospitalization for heart failure

HTE

heterogeneous treatment effect

MACE

major adverse cardiovascular event

OLDW

OptumLabs Data Warehouse

SGLT2i

sodium–glucose cotransporter 2 inhibitor

T2D

type 2 diabetes

Research Perspective.

What Is New?

• Among adults with type 2 diabetes at moderate risk for cardiovascular disease (baseline predicted risk of major adverse cardiovascular events, 1%–5%) initiating glucagon‐like peptide‐1 receptor agonist, sodium–glucose cotransporter 2 inhibitor, dipeptidyl peptidase‐4 inhibitor, or sulfonylurea therapy, we observed heterogeneous treatment effects as a function of patients' baseline cardiovascular disease risk, with higher‐risk patients deriving greater benefit with respect to preventing major adverse cardiovascular events with preferential use of glucagon‐like peptide‐1 receptor agonists and sodium–glucose cotransporter 2 inhibitors than lower‐risk patients; higher‐risk patients also experienced greater harm when treated with sulfonylureas.

• Glucagon‐like peptide‐1 receptor agonist and sodium–glucose cotransporter 2 inhibitor medications were associated with greater reductions in major adverse cardiovascular events, heart failure hospitalizations, and death compared with dipeptidyl peptidase‐4 inhibitors and sulfonylureas across all levels of baseline cardiovascular disease risk.

What Question Should Be Addressed Next?

• While cardiovascular benefits of sodium–glucose cotransporter 2 inhibitors and glucagon‐like peptide‐1 receptor agonists exist across all levels of moderate cardiovascular risk, further research is needed to establish the cost‐effectiveness thresholds for their preferential use and to develop and implement sustainable and equitable care delivery and payment models to ensure access to pharmacotherapies best equipped to reduce cardiovascular morbidity and death in all people with type 2 diabetes.

The primary objective of type 2 diabetes (T2D) treatment is to prevent the morbidity, disability, and death associated with its complications. Cardiovascular disease (CVD) is the leading cause of death among people with T2D, and 23% of all CVD deaths and 28% of health care expenditures for CVD are attributable to diabetes. ¹ As a result, clinical practice guidelines recommend choosing glucose‐lowering medications specifically to lower the risk of CVD events in individuals with established CVD or with indicators of high CVD risk. ² , ³ These include glucagon‐like peptide‐1 receptor agonists (GLP‐1RAs) and sodium–glucose cotransporter 2 inhibitors (SGLT2is), which have been shown in multiple clinical trials to reduce risks of major adverse cardiovascular events (MACEs) as well as cardiovascular and all‐cause death in patients with indicators of high CVD risk. ² Individuals with lower levels of CVD risk are advised to use other considerations to select T2D pharmacotherapy, such as reduction of kidney disease progression and impacts on weight and glycemia. ³ The recently completed GRADE (Glycemia Reduction Approaches in Type 2 Diabetes: A Comparative Effectiveness) trial showed that individuals with low CVD risk may not derive cardiovascular benefits from preferential use of GLP‐1RAs as compared with dipeptidyl peptidase‐4 inhibitors, sulfonylurea, or basal insulin, ⁴ but GRADE was neither designed nor powered to assess CVD outcomes and focused primarily on glucose lowering. In contrast, patients with moderate levels of CVD risk (ie, higher than included in GRADE but lower than considered in cardiovascular outcome trials) may experience reduction in CVD events with use of GLP‐1RA and SGLT2i classes as compared with dipeptidyl peptidase‐4 inhibitors (DPP4is) or sulfonylureas. ⁵

With current guidelines recommending preferred use of GLP‐1RAs and SGLT2is in patients with T2D and established or at high risk of CVD, ³ , ⁶ and our recent work ⁵ extending this recommendation to patients with more moderate levels of CVD risk, >95% of adults with T2D ⁷ may be advised to be treated with these drugs. To support individualized T2D management, particularly in an era of drug shortages and rising costs of care, it is essential to understand whether specific patient subgroups benefit more from GLP‐1RAs and SGLT2is than others. Such evidence of heterogeneous treatment effects ⁸ would allow clinicians to focus therapy on those most likely to derive benefit from these drugs.

Prior studies assessing heterogeneity of treatment effects of glucose‐lowering therapies in adults with T2D primarily compared patients with, versus without, established CVD and did not focus on those at moderate CVD risk nor directly (or indirectly) compare across the 4 commonly used second‐line glucose‐lowering medications. For example, SGLT2is were associated with lower risk of myocardial infarction (MI) or stroke compared with GLP‐1RAs in patients with established CVD, but not without; however, this study did not include semaglutide in the GLP‐1RA arm. ⁹ Similarly, separate comparisons of empagliflozin (an SGLT2i) with either sitagliptin (a DPP4i) or liraglutide (a GLP‐1RA) found greater benefit of SGLT2i therapy in those with a history of CVD than in those without. ¹⁰ In a study comparing DPP4is with sulfonylureas, DPP4is were associated with greater improvements in CVD outcomes in those with prior ischemic cerebrovascular disease than those without. ¹¹ These results were consistent with findings from cardiovascular outcome trials (which included participants with either established CVD or indicators of high CVD risk), which found that cardiovascular benefits of GLP‐1RAs and SGLT2is, when compared with placebo, were most apparent in individuals with established CVD. ¹² While these studies highlighted the importance of assessing heterogeneous treatment effects (HTEs) to individualize therapy, they did not focus explicitly on the most prevalent subgroup of adults with T2D: those with moderate levels of CVD risk. ⁷

We therefore used real‐world data from a large and diverse population of adults with commercial, Medicare Advantage, and Medicare fee‐for‐service health plans across the United States to conduct prespecified HTE analyses of an emulated comparative effectiveness trial of GLP‐1RAs, SGLT2is, DPP4is, and sulfonylureas among adults with T2D at moderate risk for CVD, examining cardiovascular outcomes as a function of baseline CVD risk. This emulated trial, the design and results of which have been reported previously, ⁵ focused specifically on patients at moderate CVD risk, defined as having a baseline (ie, before treatment initiation) annualized risk of experiencing a MACE of 1% to 5% estimated using the annualized claims‐based MACE estimator (ACME). ⁷

Methods

Study Design

This is a prespecified secondary analysis of a previously published ⁵ retrospective emulation of a target trial examining the comparative effectiveness of GLP‐1RAs, SGLT2is, DPP4is, and sulfonylureas on MACEs in adults with T2D who are at moderate CVD risk. ⁵ The study was exempt from review by the Mayo Clinic Institutional Review Board, as it involved analysis of deidentified data. The protocol was preregistered on ClinicalTrials.gov (NCT05214573) and the statistical analysis plan was shared previously. ⁵ Results are reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology guideline. ¹³

Data Sharing Statement

This study was conducted using deidentified data from OptumLabs Data Warehouse (OLDW) linked to a 100% sample of Medicare fee‐for‐service claims. These data are third‐party data owned by OptumLabs and contain sensitive patient information; therefore, the data are available only upon request. Interested researchers engaged in Health Insurance Portability and Accountability Act–compliant research may contact connected@optum.com for data access requests. The data use requires researchers to pay for rights to use and access the data. These data are subject to restrictions on sharing as a condition of access.

Data Source

We used administrative claims data from OLDW linked to a 100% sample of Medicare fee‐for‐service claims. OLDW includes deidentified medical and pharmacy claims and enrollment records for enrollees in commercial and Medicare Advantage health plans across the United States. ¹⁴ , ¹⁵ The study cohort is therefore composed of individuals with commercial, Medicare Advantage (ie, Medicare Part C), and traditional Medicare (ie, Medicare Parts A/B/D) plans. OLDW and Medicare fee‐for‐service claims were linked on personal identifiers and then deidentified by OptumLabs before being made available to researchers, allowing for uninterrupted observation of included individuals across health plans.

Study Cohort

We identified adults aged ≥21 years with T2D and valid demographic information (age, sex, and US Census region) who filled a new prescription of a GLP‐1RA, SGLT2i, DPP4i, or sulfonylurea between January 1, 2014, and December 31, 2021 (index date), and had a second fill of the same medication class (to confirm use) with a ≤30‐day gap between fills. Included individuals were required to have 12 months of baseline enrollment with medical and pharmacy coverage and we excluded those with a fill of any other study medications, glinide, or insulin during the baseline period or fill of any other study medication class during the 30 days after and including the index date (ie, those treated with combination therapy). We further excluded individuals with type 1 diabetes, pregnancy, or metastatic cancer diagnoses during the baseline period. Then, we restricted the study cohort to individuals with moderate CVD risk, defined as >1% to ≤5% annualized risk of experiencing a MACE as calculated using the ACME model. ⁷ ACME is a claims‐based prediction model for 1‐year risk of MACEs in T2D that was developed and internally validated using linked OLDW and Medicare fee‐for‐service (100% sample) claims data for >6.6 million adults aged ≥21 years with T2D between 2014 and 2021. The cross‐validated concordance index was 0.74 and exhibited strong calibration across the full range of risk values.

Outcomes

The primary outcome was MACE, defined as the first occurrence of hospitalization for acute MI, hospitalization for stroke, or all‐cause death. Secondary outcomes were expanded MACE (the composite of MACE, hospitalization for heart failure [HHF], and arterial revascularization procedure) and the individual components of expanded MACE. Mortality data are sourced from the Social Security Administration Death Master File, deceased status from OLDW‐linked electronic health records, death as a reason for disenrollment from an included health plan, death indicated by inpatient discharge status, obituary information, and Medicare Advantage beneficiary report information. Cause of death is not available in the data set. All other outcomes were ascertained using International Classification of Diseases, Ninth Revision (ICD‐9) and Tenth Revision (ICD‐10) codes in the medical claims (Table S1). Consistent with the intention‐to‐treat analytic approach, outcome ascertainment began the day after the index date, and patients were censored upon disenrollment from an included health plan, death (when not considered as an outcome), or the end of the study period (December 31, 2021), whichever came first (Figure S1).

Preexposure Patient Characteristics

Our primary independent variables were estimated CVD risk, based on ACME‐predicted annualized risk of MACE ⁷ between 1% and 5%, and the treatment group (GLP‐1RA, SGLT2i, DPP4i, or sulfonylurea). ACME‐predicted CVD risk was used as a continuous variable in the model; for ease of interpretation of the HTEs, we reported the effects on the outcomes stratified by the ACME‐predicted CVD risk strata. Baseline demographics (age, sex, race and ethnicity, and US region), comorbidities (Table S2), and medications (Table S3) were ascertained during the baseline 12‐month period.

Statistical Analysis

We first estimated propensity score models using baseline covariates (provided in Tables S2 and S3), ¹⁶ including a diverse set of binomial prediction algorithms in a super learner ensemble to estimate a separate prediction model for each treatment (versus a pool of the other 3). ¹⁷ Each model was used to predict the corresponding probability of treatment and construct stabilized inverse probability of treatment weights; stabilized weights ¹⁸ were used because propensity scores had extreme values when examining their distributions. ⁵ To examine the balance of study covariates across treatment groups, we computed inverse probability of treatment weighted standardized mean differences (SMDs) across all pairwise comparisons, with maximum SMD across pairwise comparisons <0.1 indicating good balance for a given covariate. ¹⁹ We also assessed the balance of baseline covariates within each CVD risk quartile to make sure the treatment arms were balanced at baseline within the CVD risk subgroups.

To examine the association between treatment exposure and outcome, we used inverse probability of treatment weighted Cox proportional hazard regression models to estimate treatment effects, stratified on CVD risk levels. ²⁰ We used the intention‐to‐treat analytic approach, in which patients were censored on the earliest of disenrollment from an included health plan, death, or the end of the study period. CVD risk strata, calculated using ACME, ⁷ were set to >1% to ≤2% (lower risk), >2% to ≤3%, >3% to ≤4%, and >4% to ≤5% (higher risk). To assess for heterogeneity of hazard ratios (HRs) across levels of ACME‐predicted CVD risk, we estimated a series of models ²¹ , ²² incorporating different parameterizations for the functional form of the interaction effect between treatment and CVD risk on the hazard scale. ²³ , ²⁴ A separate model was estimated for each outcome. We tested 3 nested models: (1) a base model that did not include interaction parameters between treatment and CVD risk; (2) a model adding CVD risk as a continuous variable with an interaction term for each treatment group; and (3) a model further adding an interaction term of CVD risk and risk squared for each treatment group. Model 3 (quadratic term) would allow the detection of nonlinearity effect (see Data S1). We used the likelihood ratio test to compare the nested models to select the best model for each outcome, and this final model was used to assess interaction. Interactions between ACME‐predicted CVD risk and treatments were tested with the likelihood ratio test, with a significance level of 5% used to identify significant HTEs (ie, variation in the treatment effects as assessed by HR across different strata of CVD risk). Our a priori specified plan was to use, for each outcome, the simplest model out of the 3; that is, model 3 was to be used if the quadratic term was statistically significant, model 2 was to be used if the main interaction term was statistically significant but the quadratic term was not, and model 1 was to be used if there was no statistically significant heterogeneity.

HRs and 95% CIs were calculated for all pairwise comparisons between the treatment arms at different levels of ACME‐predicted CVD risk. Cumulative incidence rates of each outcome by treatment arm were estimated using the inverse probability of treatment weighted Kaplan–Meier method. Clinical interpretation relied on estimates of absolute risk differences at 1, 2, and 3 years after the index date; these were calculated for each pairwise comparison using the estimated event rates.

Sensitivity Analyses

We performed several sensitivity analyses to test the robustness of our findings. First, we repeated all analyses excluding patients who had experienced a nonfatal MACE (ie, MI or stroke) at baseline. Second, we replicated this analysis excluding patients who had experienced a nonfatal expanded MACE (ie, MI, stroke, HHF, or revascularization procedure) at baseline. Third, we adjusted our model for the index year of drug initiation to account for any secular effects of drug use patterns.

Data management was conducted using SAS 9.04 (SAS Institute Inc., Cary, NC), and statistical analyses were conducted using R version 4.1 (R Foundation for Statistical Computing, Vienna, Austria).

Results

We identified 386 276 adults with T2D at moderate CVD risk initiating 1 of the study drugs. Their preweighting characteristics are shown in Table S4; additional detail regarding the overall cohort has been published previously. ⁵ After inverse probability weighting, 25.2% of the cohort was at lower risk (>1% to ≤2%) for CVD and 13.3% was at higher risk (>4% to ≤5%). The study population was well balanced across the 4 treatment arms (maximum SMD <0.10) on all baseline characteristics (Table). Median (interquartile range [IQR]) weighted follow‐up and reasons for censoring were similar across the 4 treatment groups: 1186 (IQR, 611–1873) days for DPP4is, 1149 (IQR, 574–1854) days for GLP‐1RAs, 1153 (IQR, 582–1810) days for SGLT2is, and 1155 (IQR, 571–1843) days for sulfonylureas (Table S5). In accordance with the intention‐to‐treat principle, our primary analysis included all patients who initiated treatment with the study drug to assess drug long‐term treatment effects. Their median times on treatment were 252 (IQR, 137–389) days for DPP4is, 166 (IQR, 95–263) days for GLP‐1RAs, 215 (IQR, 121–330) days for SGLT2is, and 315 (IQR, 167–483) days for sulfonylureas (Table S5).

Table 1.

Baseline Characteristics of the Study Cohort

	DPP4i	GLP‐1RA	SGLT2i	Sulfonylurea	Largest SMD
N	82 438	44 255	46 473	207 187
CVD risk level, n (%)
>1% to ≤2%	20 383 (24.7)	11 371 (25.7)	12 167 (26.2)	51 871 (25.0)	0.03
>2% to ≤3%	30 762 (37.3)	16 614 (37.5)	17 604 (37.9)	76 389 (36.9)	0.02
>3% to ≤4%	20 156 (24.4)	10 827 (24.5)	10 966 (23.6)	50 799 (24.5)	0.02
>4% to ≤5%	111 37 (13.5)	5443 (12.3)	5737 (12.3)	28 128 (13.6)	0.04
Age, y, mean±SD	65.76±8.38	65.47±8.25	65.46±8.35	65.69±8.38	0.04
Age group, y, n (%)					0.06
<45	696.4 (0.8)	356.6 (0.8)	400.4 (0.9)	1902.6 (0.9)
45–49	3239.8 (3.9)	1782.3 (4.0)	1848.7 (4.0)	8014.5 (3.9)
50–54	6618.1 (8.0)	3661.8 (8.3)	3946.1 (8.5)	16 841.0 (8.1)
55–59	8679.1 (10.5)	4740.4 (10.7)	5094.5 (11.0)	21 847.4 (10.5)
60–64	8422.4 (10.2)	4674.1 (10.6)	4919.8 (10.6)	21 377.4 (10.3)
65–69	23 394.6 (28.4)	13 185.8 (29.8)	13 412.9 (28.9)	59 435.0 (28.7)
70–74	22 186.8 (26.9)	11 645.9 (26.3)	12 161.2 (26.2)	55 035.6 (26.6)
≥75	9200.4 (11.2)	4207.9 (9.5)	4689.4 (10.1)	22 732.8 (11.0)
Sex, n (%)					0.01
Female	40 506.38 (49.1)	21 653.53 (48.9)	22 971.49 (49.4)	101 537.38 (49.0)
Male	41 931.2 (50.9)	22 601.3 (51.1)	23 501.5 (50.6)	105 648.8 (51.0)
Race and ethnicity, n (%)					0.04
Asian	2486.9 (3.0)	1029.9 (2.3)	1289.4 (2.8)	6176.2 (3.0)
Black	8567.0 (10.4)	4525.1 (10.2)	4698.8 (10.1)	21 662.7 (10.5)
Hispanic	7016.2 (8.5)	3693.6 (8.3)	3901.3 (8.4)	17 839.4 (8.6)
White	61 815.0 (75.0)	33 602.0 (75.9)	35 125.2 (75.6)	155 032.2 (74.8)
Unknown	2552.4 (3.1)	1404.3 (3.2)	1458.4 (3.1)	6475.6 (3.1)
Region, n (%)					0.04
Midwest	21 806.2 (26.5)	11 701.7 (26.4)	12 128.8 (26.1)	54 935.4 (26.5)
Northeast	10 976.3 (13.3)	5360.0 (12.1)	6061.3 (13.0)	27 349.0 (13.2)
South	38 555.7 (46.8)	20 999.2 (47.5)	21 918.5 (47.2)	96 976.0 (46.8)
West	10 949.5 (13.3)	6130.7 (13.9)	6299.6 (13.6)	27 531.1 (13.3)
Unknown	149.9 (0.2)	63.2 (0.1)	64.8 (0.1)	394.7 (0.2)
Prescriber specialty, n (%)					0.04
Cardiology	84.6 (0.1)	54.4 (0.1)	93.7 (0.2)	175.5 (0.1)
Endocrinology	1086.7 (1.3)	672.3 (1.5)	706.0 (1.5)	2704.0 (1.3)
Nephrology	51.5 (0.1)	17.4 (0.0)	21.7 (0.0)	115.9 (0.1)
Primary care	49 245.0 (59.7)	26 431.5 (59.7)	27 839.5 (59.9)	123 700.2 (59.7)
Other	12 492.0 (15.2)	6779.2 (15.3)	7070.2 (15.2)	31 597.7 (15.3)
Unknown	19 477.7 (23.6)	10 300.0 (23.3)	10 742.0 (23.1)	48 892.9 (23.6)
Index year, n (%)					0.07
2014	10 756.0 (13.0)	5480.4 (12.4)	5069.1 (10.9)	26 956.8 (13.0)
2015	11 369.2 (13.8)	5724.1 (12.9)	6269.9 (13.5)	28 492.6 (13.8)
2016	11 779.3 (14.3)	6086.5 (13.8)	6657.3 (14.3)	29 466.5 (14.2)
2017	12 720.7 (15.4)	6720.9 (15.2)	7354.6 (15.8)	31 863.5 (15.4)
2018	12 468.2 (15.1)	6867.1 (15.5)	7203.3 (15.5)	31 221.4 (15.1)
2019	12 164.9 (14.8)	6819.3 (15.4)	7158.7 (15.4)	30 438.3 (14.7)
2020	5389.9 (6.5)	3096.8 (7.0)	3216.0 (6.9)	13 786.5 (6.7)
2021	5789.4 (7.0)	3459.7 (7.8)	3544.1 (7.6)	14 960.5 (7.2)
Source data, n (%)					0.02
OLDW	37 317 (45.3)	20 318 (45.9)	21 584 (46.4)	94 401 (45.6)
Medicare fee‐for‐service	45 120.5 (54.7)	23 936.5 (54.1)	24 888.6 (53.6)	112 785.2 (54.4)
Comorbidities, n (%)
Acute MI, mo 1–9	144.2 (0.2)	72.7 (0.2)	83.5 (0.2)	393.8 (0.2)	0.01
Acute MI, mo 10–12	35.0 (0.0)	16.6 (0.0)	18.0 (0.0)	108.5 (0.1)	0.01
Coronary artery disease (other than MI)	12 446.4 (15.1)	6953.4 (15.7)	6919.1 (14.9)	31 340.9 (15.1)	0.02
Acute stroke, mo 1–9	92.0 (0.1)	50.3 (0.1)	47.2 (0.1)	238.8 (0.1)	0.0
Acute stroke, mo 10–12	23.8 (0.0)	10.4 (0.0)	10.5 (0.0)	60.4 (0.0)	0.0
Cerebrovascular disease (other than stroke)	4755.8 (5.8)	2515.6 (5.7)	2594.5 (5.6)	11 979.1 (5.8)	0.01
HHF, mo 1–9	82.9 (0.1)	50.9 (0.1)	45.0 (0.1)	225.0 (0.1)	0.01
HHF, mo 10–12	68.8 (0.1)	27.1 (0.1)	30.7 (0.1)	181.0 (0.1)	0.01
Heart failure (other than HHF)	1511.2 (1.8)	836.4 (1.9)	889.3 (1.9)	3834.0 (1.9)	0.01
Revascularization, mo 1–9	884.3 (1.1)	447.2 (1.0)	461.8 (1.0)	2269.2 (1.1)	0.01
Revascularization, mo 10–12	423.1 (0.5)	212.9 (0.5)	223.7 (0.5)	1128.2 (0.5)	0.01
Atrial fibrillation/flutter	3521.5 (4.3)	1895.4 (4.3)	1965.6 (4.2)	8820.9 (4.3)	0.0
Hypertension	67 381.5 (81.7)	36 506.5 (82.5)	37 741.6 (81.2)	169 182.4 (81.7)	0.03
Nephropathy with CKD <stage 3	6366.4 (7.7)	3475.0 (7.9)	3402.8 (7.3)	16 052.4 (7.7)	0.02
CKD, stages 3–4	4298.9 (5.2)	2317.7 (5.2)	2160.3 (4.6)	10 746.1 (5.2)	0.03
CKD stage 5 or ESKD	86.7 (0.1)	44.3 (0.1)	38.1 (0.1)	224.3 (0.1)	0.01
Renal replacement therapy	278.3 (0.3)	140.1 (0.3)	140.5 (0.3)	714.5 (0.3)	0.01
Acute kidney injury	2010.9 (2.4)	972.9 (2.2)	998.8 (2.1)	5143.4 (2.5)	0.02
Peripheral vascular disease	5443.6 (6.6)	2918.8 (6.6)	3037.5 (6.5)	13 801.5 (6.7)	0.01
Neuropathy	14 960.1 (18.1)	8210.1 (18.6)	8389.7 (18.1)	37 657.2 (18.2)	0.01
Amputation	280.8 (0.3)	144.4 (0.3)	141.2 (0.3)	723.7 (0.3)	0.01
Other lower extremity complications	1269.2 (1.5)	659.7 (1.5)	687.5 (1.5)	3274.9 (1.6)	0.01
Retinopathy	7392.3 (9.0)	3975.4 (9.0)	4025.4 (8.7)	18 634.9 (9.0)	0.01
Retinopathy treatment	353.1 (0.4)	195.1 (0.4)	189.6 (0.4)	899.3 (0.4)	0.01
Blindness	414.3 (0.5)	188.2 (0.4)	198.3 (0.4)	1092.7 (0.5)	0.01
Hyperglycemic crisis	74.3 (0.1)	48.8 (0.1)	38.3 (0.1)	228.5 (0.1)	0.01
Hypoglycemic crisis	70.2 (0.1)	43.1 (0.1)	45.2 (0.1)	199.7 (0.1)	0.0
Obesity	27 772.6 (33.7)	15 568.2 (35.2)	16 199.1 (34.9)	70 042.1 (33.8)	0.03
Metabolic/bariatric surgery	1158.9 (1.4)	596.5 (1.3)	654.0 (1.4)	2902.6 (1.4)	0.01
Pancreatitis	387.8 (0.5)	184.8 (0.4)	215.6 (0.5)	1060.5 (0.5)	0.01
Pancreatic cancer	104.3 (0.1)	64.6 (0.1)	56.3 (0.1)	280.1 (0.1)	0.01
Cancer	7297.8 (8.9)	3765.4 (8.5)	4039.5 (8.7)	18 263.5 (8.8)	0.01
Smoking	7159.4 (8.7)	3821.7 (8.6)	3973.2 (8.5)	18 129.2 (8.8)	0.01
Thyroid cancer	307.9 (0.4)	186.2 (0.4)	170.5 (0.4)	757.5 (0.4)	0.01
Genitourinary tract infection	3931.9 (4.8)	2011.5 (4.5)	2082.7 (4.5)	9943.4 (4.8)	0.02
Cirrhosis	734.9 (0.9)	384.0 (0.9)	396.2 (0.9)	1838.5 (0.9)	0.0
Dementia	938.8 (1.1)	462.3 (1.0)	441.5 (0.9)	2369.7 (1.1)	0.02
Falls	3406.1 (4.1)	1785.1 (4.0)	2001.0 (4.3)	8495.7 (4.1)	0.01
Urinary incontinence	3197.6 (3.9)	1676.3 (3.8)	1765.1 (3.8)	8033.0 (3.9)	0.0
Unplanned hospitalizations	6607.9 (8.0)	3308.6 (7.5)	3548.5 (7.6)	16 785.2 (8.1)	0.02
Medications, N (%)
Antiplatelet drug	3679.0 (4.5)	2096.9 (4.7)	2017.8 (4.3)	9386.5 (4.5)	0.02
Diuretics	32 495.1 (39.4)	17 580.5 (39.7)	18 155.6 (39.1)	81 575.0 (39.4)	0.01
RAAS inhibitors	55 638.6 (67.5)	30 107.8 (68.0)	31 059.4 (66.8)	139 745.3 (67.4)	0.03
Sacubitril/valsartan	73.1 (0.1)	55.5 (0.1)	64.6 (0.1)	185.5 (0.1)	0.01
MRA	1882.8 (2.3)	1043.9 (2.4)	1058.2 (2.3)	4722.0 (2.3)	0.01
β Blockers	27 425.3 (33.3)	14 987.5 (33.9)	15 215.6 (32.7)	68 909.0 (33.3)	0.02
Calcium channel blocker	21 075.9 (25.6)	11 306.7 (25.5)	11 648.7 (25.1)	53 036 (25.6)	0.01
Other antihypertensives	3928.9 (4.8)	2142.0 (4.8)	2131.8 (4.6)	9946.7 (4.8)	0.01
Lipid‐lowering meds	56 316.6 (68.3)	30 481.3 (68.9)	31 588.2 (68.0)	141 341.0 (68.2)	0.02
Anticoagulants	4060.4 (4.9)	2176.2 (4.9)	2262.4 (4.9)	10 190.5 (4.9)	0.0
Metformin	65 126.6 (79.0)	35 102.3 (79.3)	36 691.3 (79.0)	163 472.0 (78.9)	0.01
Thiazolidinedione	3720.5 (4.5)	2213.3 (5.0)	2235.0 (4.8)	9488.3 (4.6)	0.02

All numbers in this table are weighted. Maximum SMD <0.10 is indicative of excellent balance among groups. CKD indicates chronic kidney disease; CVD, cardiovascular disease; DPP4i, dipeptidyl peptidase‐4 inhibitor; ESKD, end‐stage kidney disease; GLP‐1RA, glucagon‐like peptide‐1 receptor agonist; HHF, hospitalization for heart failure; MI, myocardial infarction; MRA, mineralocorticoid receptor antagonist; OLDW, OptumLabs Data Warehouse; RAAS, renin–angiotensin–aldosterone system; SGLT2i, sodium–glucose cotransporter 2 inhibitor; and SMD, standardized mean difference.

Patients in the higher‐risk group were older, more likely to be men, and more likely to be non‐Hispanic White individuals than those in the lower‐risk group (Table 1). Baseline characteristics of the cohort subset by level of CVD risk are detailed in Tables S6 through S9. Within each subcohort, the treatment arms were balanced (SMD, <0.10) on all baseline covariates except the year of cohort entry (maximum SMD, 0.17) in the higher‐risk group, where SGLT2i‐treated patients were overrepresented in the later years of the study (Table S9).

During model selection, model 3 (quadratic term) was not selected for any of the outcomes (all P>0.05, indicating that there is no evidence for rejecting the assumption of a linear trend). Model 2 (linear interaction) was statistically significant for MACE (P=0.001), expanded MACE (P<0.001), MI (P=0.02), and arterial revascularization (P=0.03) outcomes. No statistically significant interaction between CVD risk and treatments was detected for all‐cause death, stroke, or HHF, indicating consistent relative treatment effects with increasing baseline levels of CVD risk. The HRs, reflecting treatment effects as assessed on the relative hazard scale (Figure 1), and the absolute risk differences in event rates seen with use of each pair of study medications (Figure 2) for each study outcome are discussed in detail below.

Figure 1. Relative hazards of experiencing each of the MACEs by baseline CVD risk.

Hazard ratios were calculated using model 1 (no interaction term) for all‐cause death, acute stroke, and HHF, and using model 2 (linear interaction) for outcomes of MACE, expanded MACE, myocardial infarction, and revascularization procedure. Though the pairwise comparisons are presented separately, results were derived from a single statistical model including all treatment groups. CVD indicates cardiovascular disease; DPP4, dipeptidyl peptidase‐4 inhibitor; GLP1, glucagon‐like peptide‐1 receptor agonist; HHF, hospitalization for heart failure; MACE, major adverse cardiovascular event; MI, myocardial infarction; SGLT2, sodium–glucose cotransporter 2 inhibitor; and SU, sulfonylurea.

Figure 2. Absolute risk differences of experiencing each of the MACEs at 1, 2, and 3 years after treatment initiation.

These absolute risk differences were calculated on the basis of the inverse probability of treatment weighted Kaplan–Meier estimation of event rates in each treatment group. Though the pairwise comparisons are presented separately, results were derived from a single statistical model including all treatment groups. CVD indicates cardiovascular disease; DPP4, dipeptidyl peptidase‐4 inhibitor; GLP1, glucagon‐like peptide‐1 receptor agonist; HHF, hospitalization for heart failure; MACE, major adverse cardiovascular event; MI, myocardial infarction; SGLT2, sodium–glucose cotransporter 2 inhibitor; and SU, sulfonylurea.

Major Adverse Cardiovascular Events

The magnitude of the relative hazard for experiencing MACEs with DPP4i versus sulfonylurea therapy diminished with higher baseline levels of CVD risk, from HR 0.76 (95% CI, 0.71–0.81) among lower‐risk patients to HR 0.91 (95% CI, 0.87–0.96) among higher‐risk patients, meaning that the relative benefit of DPP4is compared with sulfonylurea was greater in patients with lower (ie, >1% to ≤2%) baseline CVD risk. Still, at all CVD risk levels, DPP4i remained associated with lower risk of MACEs compared with sulfonylurea (Figure 1; Table S10). The absolute risk differences in predicted probabilities of MACEs between DPP4i‐ and sulfonylurea‐treated patients, however, increased with higher baseline CVD risk level and widened over time (Figure 2; Table S11; Figure S2). At 3 years of treatment, use of DPP4i instead of sulfonylurea resulted in absolute reductions in the estimated rate of MACE of 1.1% in lower‐risk patients and 1.5% in higher‐risk patients. Predicted probabilities of MACEs during each year of treatment are shown in Table S12 and Figure S3.

Conversely, the magnitude of the relative hazard for experiencing MACE with SGLT2i versus DPP4i therapy increased with higher baseline levels of CVD risk, from HR 0.99 (95% CI, 0.88–1.12) among lower‐risk patients (indicative of no difference between the treatment groups) to HR 0.78 (95% CI, 0.70–0.87) among higher‐risk patients (indicative of SGLT2i superiority). The absolute risk differences between SGLT2is and DPP4is widened with both increasing baseline CVD risk level and over time, from no difference in predicted event rates among lower‐risk patients to 2.4% fewer MACEs in higher‐risk patients treated with SGLT2is compared with DPP4is.

There was no difference across CVD risk levels in the relative hazard for experiencing MACEs between GLP‐1RAs and DPP4is (GLP‐1RA use was associated with lower risk of MACEs than DPP4is at each level of CVD risk), GLP‐1RAs and sulfonylureas (GLP‐1RA were associated with lower risk of MACEs than sulfonylureas), SGLT2is and GLP‐1RAs (no difference in the risk of MACEs), and SGLT2is and sulfonylureas (SGLT2is were associated with lower risk of MACEs). In absolute terms, the differences in treatment effects between GLP‐1RAs and DPP4is, GLP‐1RAs and sulfonylureas, and SGLT2is and sulfonylureas all widened with both rising baseline CVD risk level and longer follow‐up time. At 3 years of follow‐up, GLP‐1RAs were associated with 1.6% and 3.1% absolute risk reductions in MACEs compared with sulfonylureas among lower‐risk and higher‐risk patients, respectively; while SGLT2is were associated with absolute risk reductions in MACEs of 1.3% and 3.9% compared with sulfonylureas, respectively. The absolute differences in treatment effects between SGLT2is and GLP‐1RAs were small across CVD risk levels for up to 3 years of therapy, with the greatest difference being 0.8% lower predicted rate of MACEs in SGLT2i‐treated higher‐risk patients compared with GLP‐1RA‐treated patients.

Results for expanded MACEs were similar to those seen for MACEs. A detailed discussion of the results for individual MACE components is included in Data S2.

All‐Cause Death

There was no significant HTEs as assessed by HR in the death outcome as a function of baseline CVD risk, with DPP4is, GLP‐1RAs, and SGLT2is all associated with lower risks of death compared to sulfonylurea across CVD risk categories. SGLT2is and GLP‐1RAs were associated with lower risks of death compared with DPP4is, while no difference was seen between SGLT2is and GLP‐1RAs (Figure 1; Table S10). The absolute differences in treatment effects increased with higher CVD risk, reaching—by 3 years of treatment of higher‐risk patients—1.4% lower absolute rate of death for DPP4is versus sulfonylureas, 2.4% lower absolute rate of death for GLP‐1RAs versus sulfonylureas, 2.8% lower absolute rate of death for SGLT2is versus sulfonylureas, 1.4% lower absolute rate of death for SGLT2is versus DPP4is, and 1% lower absolute rate of death for GLP‐1RAs versus DPP4is (Figure 2; Table S11; Figure S2). Predicted probabilities of death during each year of treatment are shown in Table S12 and Figure S3.

Sensitivity Analyses

When replicating our analyses excluding patients who had experienced nonfatal MACEs (Tables S13 and S14) or nonfatal expanded MACEs (Tables S15 and S16), the results were not meaningfully changed from the primary analysis. However, in the second sensitivity analysis, the linear interaction term for HHF was significant, indicating the presence of HTEs, with greater magnitude of HHF risk reduction seen for DPP4is versus sulfonylureas at lower baseline levels of CVD risk, and for SGLT2i versus DPP4i for higher baseline levels of CVD risk. Results were similarly unchanged in the sensitivity analysis adjusted for the index year of drug initiation (Tables S17 and S18).

Discussion

Among adults with T2D at moderate CVD risk, for whom evidence for cardiovascular risk reduction with preferential selection of glucose‐lowering pharmacotherapies had been lacking, we found that when comparing GLP‐1RA, SGLT2i, DPP4i, and sulfonylurea therapy, SGLT2is and GLP‐1RAs were associated with lower risks of MACEs, expanded MACEs, and their component outcomes than DPP4is and sulfonylureas across all levels of baseline CVD risk. The beneficial effects of SGLT2is and GLP‐1Ras, particularly as compared with sulfonylureas, were apparent across the full range of moderate CVD risk, including among lower‐risk patients with 1% ACME‐predicted annualized risk of MACEs. Those with highest baseline CVD risk generally derived the largest absolute risk reduction with SGLT2i therapy—as compared with sulfonylureas, DPP4is, and even GLP‐1RAs—for MACEs, expanded MACEs, MI, HHF, and all‐cause death outcomes. Higher‐risk patients also derived greater absolute benefit with use of nonsulfonylurea agents compared with sulfonylureas for all outcomes except revascularization procedures.

Our study builds on a robust evidence base supporting GLP‐1RA and SGLT2i use for adults with T2D in the setting of high CVD risk, ²⁵ filling the knowledge gap for the treatment of individuals with moderate levels of CVD risk who comprise the majority of adults with T2D ⁷ yet were not specifically considered in prior clinical trials or observational studies. Indeed, randomized controlled trials firmly established the efficacy of GLP‐1RAs and SGLT2is to improve MACE and expanded MACE outcomes, compared with other glucose‐lowering therapies while maintaining comparable hemoglobin A_1c levels, in patients with T2D with established CVD or at high risk for CVD. ² , ²⁵ For individuals without established CVD, and presumably at low risk for CVD events due to their relatively recent diagnosis of T2D (and the low observed rate of CVD events), secondary analysis of the GRADE trial did not find a consistent difference between GLP‐1RAs, DPP4is, sulfonylureas, and basal insulin (studying 1 drug from each class). ⁴ However, GRADE was not powered to detect differences in CVD end points, particularly in a low‐risk patient population, and it did not include an SGLT2i treatment arm. Prior studies that sought to assess HTEs by CVD risk focused on those with versus without established CVD; however, the CVD risk profile of the latter group varied among studies, from high‐risk patients included in the cardiovascular outcomes trials ¹² , ²⁵ to unspecified (and likely heterogeneous) levels of risk in observational analyses. ⁹ , ¹⁰ , ¹¹ While none of the prior studies compared across all 4 T2D medication classes, their results suggested that SGLT2is are superior to GLP‐1RAs among patients with established CVD, but the 2 drugs are likely comparable among patients without such history.

A meta‐analysis of GLP‐1RA and SGLT2i randomized controlled trials that (separately) examined their HTEs by baseline CVD risk (measured as the observed cardiovascular death in the control group), found that the relative benefit of these medications is unchanged, and their absolute benefit increases with greater level of baseline CVD risk. ²⁵ The only notable exception was seen for the outcome of HHF, where SGLT2i therapy also had increasing relative benefits with rising baseline levels of CVD risk. Thus, our results are consistent with this meta‐analysis and further support preferential GLP‐1RA and SGLT2i use for patients across the spectrum of baseline CVD risk. Moreover, our analyses found even greater benefit with SGLT2i use than with GLP‐1RA use in patients at the higher spectrum of CVD risk. By comparing SGLT2is and GLP‐1RAs with each other and to DPP4is and sulfonylureas, our study provides strong evidence in support of preferential use of SGLT2is and GLP‐1RAs by most patients with T2D. This is important, as DPP4is and sulfonylureas continue to be used much more frequently than SGLT2is or GLP‐1RAs, with both medication classes less costly and often more accessible to patients, particularly those with Medicare insurance who are ineligible for copayment reduction cards and have higher out‐of‐pocket payments when treated with brand‐name therapies.

For all examined outcomes and across the range of baseline ACME‐predicted CVD risk, we found sulfonylurea drugs to be associated with worse outcomes when compared with GLP‐1RAs, SGLT2is, and DPP4is. Thus, irrespective of baseline CVD risk and even at the lowest end of the moderate CVD risk category, it may be best, if financially feasible, to avoid use of sulfonylureas in patients prioritizing CVD risk reduction. The trade‐offs between clinical benefit and affordability of GLP‐1RAs and SGLT2is need to be addressed and discussed with each patient, weighing the individualized benefits expected with preferential use of these drugs against the financial burden they may pose. These considerations are also important for addressing socioeconomic and racial disparities in diabetes treatment and health outcomes, as low‐income patients and individuals from racially minoritized groups have historically been less likely to be treated with GLP‐1RA and SGLT2i drugs, ²⁶ , ²⁷ , ²⁸ , ²⁹ despite (and potentially contributing to) the higher rates of CVD in these groups. ³⁰ , ³¹

In the absence of randomized controlled trials comparing these medications head‐to‐head, especially in the moderate CVD risk subgroup, robustly designed observational studies provide important evidence for comparative effectiveness in real‐world care settings and heterogeneous, clinically and socio‐demographically, patient populations. Our study is the first to examine HTEs in cardiovascular outcomes across pairwise comparisons of the 4 commonly used T2D medication classes across the full range of moderate CVD risk. It is further strengthened by the size and diversity of the patient population included. In the absence of a longitudinal clinical database for all Americans, the linked OLDW/Medicare fee‐for‐service data may be the closest alternative composed of individuals across the entire United States, from different health systems, regions, and socioeconomic backgrounds.

This study also has important limitations. Even with robust analytic methods and adherence to the target trial framework, ³² observational studies are prone to residual confounding. ³³ , ³⁴ Data on important risk factors for adverse outcomes, most notably social determinants of health, are not available in claims. Our primary analyses followed the intention‐to‐treat approach, which closely aligns with the clinical trials we sought to emulate and best addresses the decisional dilemma facing clinicians and patients initiating a new therapy in clinical practice. While the exposed (on‐treatment) time period was shorter than the intention‐to‐treat follow‐up, our prior analysis found the drug treatment effects to be similar between the intention‐to‐treat and on‐treatment comparisons. ⁵ We also used a claims‐based MACE prediction model ⁷ for these analyses, and the “moderate” risk group identified using this risk calculator may not directly translate to “moderate” risk groups as identified by other models such as the American College of Cardiology/American Heart Association atherosclerotic CVD risk calculator. ³⁵ It will be important to externally validate ACME and compare it with other CVD risk scores available in electronic health record data. Finally, this analysis examined HTEs only by CVD risk and did not incorporate individual variable effect modifiers. Recent work has proposed methods to more thoroughly examine HTEs by incorporating candidate effect modifiers. ³⁶ However, in the absence of good prior information of what these modifiers are, these approaches may be prone to false discovery and overfitting. ³⁷

Conclusions

Understanding the heterogeneity in the cardiovascular benefits of GLP‐1RAs and SGLT2is as compared with DPP4is and the less costly sulfonylureas is essential to inform shared decision making and high‐quality, cost‐effective, person‐centered care in T2D. ²⁴ Our prespecified cardiovascular risk HTE analysis of the comparative effectiveness of GLP‐1RA, SGTL2i, DPP4i, and sulfonylurea drugs among patients with T2D at moderate baseline risk for CVD revealed that SGLT2is and GLP‐1RAs may be the preferred glucose‐lowering agents for CVD risk reduction across all baseline levels of CVD risk, while sulfonylureas are clearly and consistently inferior to the other medications. Moreover, SGLT2is may be preferred to GLP‐1RAs in patients with higher levels of CVD risk, even in the absence of established CVD (the population studied in prior HTE analyses ⁹ , ¹⁰ , ¹¹ , ¹² ). These findings support extending current clinical practice guideline recommendations to suggest preferential use of GLP‐1RAs and SGLT2is in all patients prioritizing cardiovascular risk reduction rather than just those at high baseline CVD risk. As such, it will be essential to address the financial and logistical barriers to obtaining these evidence‐based therapies. This includes eliminating step therapy requirements, reducing out‐of‐pocket cost‐sharing responsibilities, and addressing racial and ethnic disparities in drug prescribing. ²⁶ , ²⁷ , ³⁸ , ³⁹

Sources of Funding

Research reported in this work was funded through a Patient‐Centered Outcomes Research Institute Award (DB‐2020C2–20306). Drs McCoy and Maron are investigators at the University of Maryland–Institute for Health Computing, which is supported by funding from Montgomery County, Maryland, and the University of Maryland Strategic Partnership: MPowering the State, a formal collaboration between the University of Maryland, College Park and the University of Maryland, Baltimore.

Disclosures

In the past 36 months, Dr McCoy has received support from the National Institute of Diabetes and Digestive and Kidney Diseases of the National Institutes of Health, the National Institute on Aging of the National Institutes of Health, the National Center for Advancing Translational Sciences, and the American Diabetes Association for projects unrelated to this work. She also served as a consultant to Emmi (Wolters Kluwer) and Yale/New Haven Health System and has received speaking honoraria and travel support from the American Diabetes Association. Dr Herrin currently receives support from the Centers for Medicare and Medicaid Services to develop measures of quality and equity; the National Institutes of Health, Agency for Healthcare Research and Quality, the Patient‐Centered Outcomes Research Institute, and the American Heart Association for multiple research projects. Dr Ross currently receives research support through Yale University from Johnson and Johnson to develop methods of clinical trial data sharing, from the Food and Drug Administration for the Yale–Mayo Clinic Center for Excellence in Regulatory Science and Innovation program (U01FD005938), from the Agency for Healthcare Research and Quality (R01HS022882), and from Arnold Ventures; formerly received research support from the Medical Device Innovation Consortium as part of the National Evaluation System for Health Technology; and, in addition, Dr Ross was an expert witness at the request of Relator's attorneys, the Greene Law Firm, in a qui tam suit alleging violations of the False Claims Act and Anti‐Kickback Statute against Biogen Inc. that was settled September 2022. Other authors have no conflicts of interest to disclose.

Supporting information

Data S1–S2

Tables S1–S18

Figures S1–S3

References 40–50