IntroductionRationale

Several major cardiovascular outcome trials (CVOTs) with glucagon-like peptide-1 receptor agonists (GLP1-RAs) have been conducted over the last decade in patients with type-2 diabetes (T2D).1–3 These agonists have proven cardiovascular (CV) safety, and liraglutide, dulaglutide and semaglutide have been demonstrated to reduce the risk of major adverse cardiovascular events (MACE). A recent meta-analysis of these CVOTs concluded that GLP1-RA was associated with a reduction of the risk of MACE by 14% (HR 0.86, 95% CI 0.79 to 0.94, p=0.006) with a non-significant heterogeneity between subgroups of patients with and without CV disease (p=0.127), of CV death by 13% (p=0.016), of all-cause mortality by 12% (p=0.012), of non-fatal stroke by 16% (p=0.007), of hospitalisation for heart failure by 10% (p=0.023) and of the broad composite kidney outcome by 17% (p<0.012).4 These results were confirmed by another meta-analysis of CVOTs.5 The potential cardioprotective effect of incretin-based therapies is attributed to their multiple non-glycaemic actions in the CV system, including changes in insulin resistance, weight loss, reduction in blood pressure, improved lipid profile and direct effects on the heart and vascular endothelium, besides their glucose-lowering action.6 The main observational studies conducted on GLP1-RA therapy in real-life settings have also shown their impact on HbA1c, fasting blood glucose, renal function and body mass index (BMI).7–11

Real-world studies investigating CV outcomes allow comparisons between GLP-1RAs and other glucose-lowering drugs (GLDs) instead of placebo, in the absence of clinical trial inclusion/exclusion procedures, and in populations with more heterogeneous characteristics. A recent meta-analysis of real-world studies on the CV effects of GLP1-RAs versus those of other GLDs showed that GLP1-RAs were associated with a significant reduction of the risk of the composite CV outcome, the MACE and all-cause mortality.12

Four GLP1-RAs—dulaglutide, exenatide, liraglutide and semaglutide—are currently reimbursed in France for T2D.13 Their utilisation was initially authorised by French Health authorities (‘Haute Autorité de Santé’) and reimbursed by the National Sickness Fund in adult patients with T2D, as second or third line of drug therapy in the following situations:

In dual therapy with metformin or sulfonylurea, except for dulaglutide and semaglutide (due to lack of clinical evidence).

In triple therapy with sulfonylurea and metformin or with metformin and insulin, except for exenatide and semaglutide (due to lack of clinical evidence).

The fixed-dose combination of liraglutide and insulin degludec was also recommended in association with metformin in patients whose treatment with the separate combination of liraglutide and basal insulin was optimised and in patients not controlled by the combination of metformin and basal insulin.

Based on these recommendations and other aspects regarding market access procedures, it is likely that the use of GPL1-RAs in patients with T2D is different in France than in other countries.14–16 Patients treated in France also differ from the populations enrolled in CVOTs. The possibility offered by the existence of nationwide claims and hospital database documenting the most relevant clinical data of interest made it interesting to design a retrospective real-world study on the effectiveness of the utilisation of GLP1-RAs as compared with other GLDs in patients with T2D in routine practice regarding CV risk.

Objectives

Primary objective: To assess whether GLP1-RAs in association with metformin are associated with a reduction in CV events and all-cause death in comparison with other classes of medications (DPP4i, sulfonylureas/glinides, basal insulin) also associated with metformin.

Secondary objectives: For patients initiating either a GLP1-RA or other classes of GLD included in the primary objective: (1) to describe the characteristics of patients from the different populations of interest and (2) to describe the persistence of treatment schemes over time.

MethodsData source

The Système National des Données de Santé (SNDS) is a comprehensive nationwide administrative healthcare database in France. It covers 67 million people with state health insurance (more than 99% of the population in France) and allows conducting large-scale observational studies of medical conditions, treatment patterns and outcomes. The SNDS database integrates various administrative, healthcare and social data sources and offers a comprehensive overview of the French population’s healthcare utilisation and outcomes. This includes hospital admissions, outpatient care, drug dispensation, healthcare reimbursements and demographic information.17 The SNDS database allows individual-level data to be linked to perform longitudinal analyses and to follow patients over time. This feature is used to assess long-term health outcomes and to evaluate the effectiveness and safety of different medical interventions.

It also includes sociodemographic, medical and administrative data on beneficiaries (age, gender, specific health coverage for deprived people, International Classification of Diseases, 10th Revision (ICD-10) diagnoses of long-term diseases (LTD) eligible for full reimbursement when they belong to a predefined list of chronic severe conditions, including most CV diseases. A social deprivation index was available for each patient.18

Information on hospitalisations includes primary and secondary diagnoses (ICD-10 codes), procedures performed during the stay, length of stay (total and number of days in an intensive care unit, dates of stay (entry and exit dates are available since 2009), and expensive medications and medical devices that are funded on top of the Diagnosis Related Group (DRG) tariff.

However, the SNDS claims database does not include biological tests or imaging results.

Study design

A population-based cohort study of patients with T2D who initiated treatment with GLP1-RAs in France between 1 January 2016 and 30 June 2021 will be conducted using data from the French National Healthcare Data System (SNDS). The study is expected to be conducted over a 6-month period starting in Q3 2024.

Several cohorts of patients initiating GLP1-RA in association with metformin according to the diabetes treatments received before GLP1-RA initiation (over a 6-month period) and the treatment received simultaneously with GLP1-RA will be constituted and followed up until the end of the study period (31 December 2021). These groups will be compared with control groups for each cohort, allowing a comparison with DPP4i, sulfonylureas/glinides or basal insulin in association with metformin and matched with T2D controls (3:1 ratio) based on the year of drug initiation and treatment regimens before and simultaneously with GLP1-RA.

Table 1 presents comparisons that will be made according to the treatment schemes. Overall, 10 separate analyses (20 groups of patients) will be conducted, allowing the outcomes to be identified separately.

Patients’ selection

To select patients of interest in the SNDS database, the following stepwise process was conducted:

Identification of patients with T2D

The first algorithm will be used to qualify patients with any type of diabetes. This algorithm was developed and validated by the French Health Insurance Schemes (CNAM).19 A patient will be considered to present with diabetes if and only if this patient had received at least three reimbursements for any GLDs during the year of treatment initiation or during the previous year (at least two reimbursements if at least one large pack size was dispensed) or when this patient had been hospitalised at least once with a diagnosis (ICD-10) code: E10 (T1D, E11 (T2D), E13 (other specified diabetes) or E14 (unspecified diabetes)) over these 2 years, or if this patient had beneficiated of an LTD status qualified with the same codes by the National Sickness Fund.

Second, patients with T2D will be identified among patients with diabetes by a classification algorithm already used in previous studies to exclude patients with T1D.20

Identification of patients with T2D initiating a GLP1-RA

Initiation (index date) will be defined as the first delivery of any GLP1-RA treatment over the study period, without any GLP1-RA delivery over the year preceding the index date.

For the 4 GLP1-RAs (dulaglutide, exenatide, liraglutide and semaglutide) reimbursed in France during the period of interest, the scope of the analysis will be restricted to the regimens recommended by French health authorities as follows: (1) for dual therapy with metformin (in cases of intolerance or contraindication to sulfonylurea), (2) in dual therapy with a sulfonylurea, except for dulaglutide and semaglutide, (3) in triple therapy with a sulfonylurea and metformin and (4) in triple therapy with metformin and insulin, except for exenatide and semaglutide. Figure 1 displays how these limitations will be implemented.

Figure 1

Flow chart of cases selection. HAS, Haute Autorité de Santé (High Authority for Health).

Identification of controls

The selection of matched control cohorts will be based solely on the year of drug initiation and treatment regimens before and simultaneously with GLP1-RA in the different selected cohorts, as listed in table 1. Direct utilisation of a propensity score to select controls in the whole database was not possible for regulatory reasons. Only patients with a minimum 6-month period before the initiation of new treatments, as well as 6 months after the index date, were selected. The control subjects will be matched to patients initiating a GLP1-RA using a 3:1 ratio. Increasing this ratio provides minimal additional improvement in statistical power and may present challenges in identifying control subjects within the database when the number of cases is substantial.21

Follow-up and exposure period

The data extraction period from the SNDS covers the period from 1 January 2014 to 31 December 2021, to allow a minimum 2-year lookback period for documenting the history of the disease and chronic comorbidities.

Index date (d0) will be the date of GLP1-RAs initiation or regimens used by controls over the 4.5-year period from 1 January 2016 to 30 June 2021. Figure 2 shows a schematic of these rules illustrated in three cases.

Figure 2

Schematic of the study period.

Initiation of GLP1-RAs will be defined as not having had any treatment with GLP1-RAs previously for at least 1 year. In the case of a switch between two GLP1-RAs, only the first one will be included in the analysis. The discontinuation of GLP1-RA will be defined as no dispensing of GLP1-RA for at least 6 months. In case of treatment discontinuation, a 1-month grace period starting from the last dispensation date will be considered for outcome attribution (the period after the last drug delivery during which observed outcomes could be attributed to the exposure).

The follow-up period will start from the index date (GLP1-RAs initiation) to the end of the study period (31 December 2021). The exposure period will be defined as the period between GLP1-RA initiation and treatment discontinuation plus the grace period, except for death-related discontinuation or end of follow-up.

The same method will be applied to controls, with an index date corresponding to the initiation of DPP4i, sulfonylurea/glinides and basal insulin combined with metformin in the regimens shown in figure 1. To avoid time bias, controls in each cohort of regimens will be selected in the database with a distribution of index dates by calendar year equivalent to the one observed for the corresponding cases.

Adjusted comparisons will be conducted on the probability of CV events or death, and the time to onset of CV events (survival analysis).

Outcomes

The following CV outcomes and all-cause death will be considered successively: all-cause death, transient ischaemic attack, non-haemorrhagic stroke (not fatal), unstable angina, acute heart failure, lower extremity arterial disease, carotid revascularisation, amputations of the lower limbs and revascularisation of the lower limbs. In an exploratory manner, a composite endpoint, including myocardial infarction (MI), stroke and all-cause death, will also be considered. These events will be identified using validated algorithms, summarised in online supplemental materials.22

In an exploratory manner, three composite endpoints will be computed: classical MACE (all-cause death, MI and stroke), MACE5 (all-cause death, MI, coronary revascularisation, stroke and hospitalisation due to heart failure) and major adverse limb events defined as a composite of peripheral artery disease (PAD), percutaneous transluminal angioplasty or peripheral bypass for PAD and amputation.

Patients’ baseline characteristics

The following patient characteristics will be described at the index date in each cohort of interest: baseline sociodemographics (age, sex and region of residence), index of social deprivation,23 patients with any exemption from payment for LTD and associated ICD-10 diagnostic codes, simplified Charlson index (comorbidities in a 1- year period preceding the index date),24 proportion of patients with a history of any atheromatous CV event (unstable angina, non-fatal MI, transient ischaemic attack and non-fatal ischaemic stroke, heart failure) over the 2 years before the index date (initiation of GLP1-RA) or before, GLP1-RA initiation prescribers’ specialty, history of high blood pressure, duration of diabetes (date of first LTD status related to diabetes), concomitant treatment (with antithrombotic and antiplatelet therapies, Angiotensin-Converting Enzyme Inhibitors (ACEI)/Angiotensin II receptor blockers (ARBs), beta-blockers and lipid-lowering agents).

Statistical analysisGeneral method

Quantitative variables will be described using the following descriptive statistics: size, number of missing values, mean, SD, median, minimum and maximum. Categorical variables will be described using the following descriptive statistics: size and percentage of each modality calculated on the expressed responses. Univariate statistical analyses will be subjected to statistical tests, and the type will depend on the nature of the variables analysed. No missing data will be considered. Statistical significance will be defined as p<0.05. Subgroup analyses according to age and gender will be performed.

Survival and multivariate analysis

Survival analyses will be performed using the Kaplan-Meier method. The data will be censored as of 31 December 2021. Main analyses will be carried out both in ‘intention to treat’ and ‘per protocol’. In the intention-to-treat analysis, patients will be followed up until the end of the study period (or death), regardless of treatment changes occurring after the first 6 months of treatment. In the per-protocol analysis, patients will be followed up during the exposure period until discontinuation of the GLP1-RA+grace period (or until the end of the study period or death).

For each comparison of treatment effects on the different outcomes, we will compute a high-dimensional Propensity Score (hdPS), to reflect the probability of being treated with GLP1-RA versus one of the other treatments adjusting for hidden or undocumented confounders. Since 2009, the hdPS method has emerged as an approach that extends traditional PS covariate selection to include large numbers of covariates that may reduce confounding bias in the analysis of large healthcare databases.22 hdPS is an automated, data-driven analytical approach for covariate selection that empirically identifies pre-exposure variables and proxies to include in the PS model.25 Some of the covariates selected by the hdPS algorithm may not be direct confounders but proxies of undocumented confounders like blood pressure, blood glucose, smoking or alcohol consumption.

In the first step, a list of data dimensions within the SNDS database will be created, distinguishing six different subsets of information of independent variables collected in the pre-exposure period (2-year period prior to the index date): (1) all outpatient drugs dispensing (CIP codes), (2) ICD-10 diagnoses related to hospitalisations and LTD registrations, (3) outpatient and inpatient medical and paramedical visits, (4) laboratory testings, (5) medical procedures and (6) medical devices reimbursed. It should be noticed that the hdPS algorithm considers all codes in each dimension without a need to specify their clinical meaning. It creates binary variables indicating the presence of each factor during the pre-exposure period. In this perspective, three binary variables are created for each code indicating at least one occurrence of the code, more than the median number of times, or more than the 75th percentile number of times.

In the second step, these variables are selected according to their greatest potential to adjust for confounding using a formula by Bross that depends on the covariate–outcome and covariate–exposure associations.26 It is planned that the 200 most prevalent codes (in terms of patient numbers) will be selected in each dimension except for the ‘outpatient and inpatient medical and paramedical visits’ dimension, where only the 20 most prevalent codes will be included due to their limited potential numbers. We will rank all variables according to their potential for bias across dimensions and keep a final number of the first 500 variables.

In a third step, treatment effects on the different outcomes measured will be estimated for each GLP1-RA group, through HR and their corresponding CIs (95% CI) using Cox regressions and/or competitive risk regressions when necessary. Models will be adjusted on hdPS and fixed baseline characteristics (sex, age, region, Universal health insurance coverage (CMUc), prescriber’s specialty, index year, diabetes duration, social deprivation index, history of atherosclerosis, history of MI amputation, history of heart failure and history of chronic renal failure).

Potential number of patients

In 2016, 3.286 million patients with diabetes were treated pharmacologically in France, of whom approximately 3.128 million patients with T2D were treated according to the High Authority for Health (HAS) recommendations.27 The latter used different existing epidemiological results to estimate that 5.5% of patients with T2D treated with metformin monotherapy, 8.1% treated with sulfonylurea monotherapy, and 18.2% treated with metformin and sulfonylurea combination therapy had an HbA1c>8% and a BMI≥30 kg/m². The target population for GLP1-RAs is estimated to be approximately 213 000 patients. A conservative estimate of 50% of those or roughly 100 000 patients should represent their fraction using treatment regimens in line with French reimbursement criteria for GLP1-RAs. Their distribution according to the treatment regimens of interest is not currently documented, but it is likely that the smallest group will be the one where GLP1-RAs are combined with basal insulin and metformin. We estimated the size of this group to be 5% (5000 patients). In summary, these estimates were considered appropriate to ensure sufficient statistical power for the planned analyses.

Discussion

The results of our study will be compared with available real-world evidence. Indeed, Caruso et al performed a meta-analysis of the CV effects of GLP1-RAs vs those of other GLDs, as observed in ten real-world studies.12 In summary, this meta-analysis showed that GLP1-RAs have been associated with a significant reduction of the risk of the composite CV outcome by 18% (HR 0.82, 95% CI 0.73 to 0.91, I2=4%) and of MACE by 30% (HR 0.70, 95% CI 0.58 to 0.84, I2=72%). All-cause mortality was also reduced to an even greater extent with GLP1-RAs use in all considered studies regardless of the comparator21–26 (HR 0.61, 95% CI 0.52 to 0.73, I2=88%). Pineda et al found that GLP1-RA initiation significantly reduced the risk of composite CV outcomes by 29%,28 while Yang et al and O’Brien et al demonstrated a 27% and 22% risk reduction versus dipeptidyl peptidase-4 inhibitor (DPP4i) users, respectively.29 30 GLP1-RAs resulted in a significant reduction in MACE risk, ranging from 10% to 45% when compared with DPP4i.29 31–34 A comparison of GLP1-RAs versus SGLT-2i has also been performed in several real-world studies but will not be summarised here as it is outside the scope of this protocol, which excluded SGLT-2i as a comparator as this class of oral hypoglycaemic agents was marketed only recently in France.

The SNDS is a nationwide database covering almost the entire population, which provides high statistical power and an absence of selection bias. Detailed records of all types of reimbursed healthcare consumption in inpatient and outpatient care settings are also available. However, the SNDS has some limitations. The SNDS contains information on reimbursed healthcare consumption only. Self-medication and prescribed non-reimbursed medications are not recorded. Moreover, it is impossible to identify patients who fail to use prescriptions provided by their doctors because such instances are not recorded in the SNDS. In addition, some clinical data are not available in the SNDS, such as biological or imaging results, obesity and the presence of some risk factors (smoking), which may constitute limitations in the comparisons, although we believe that these potential confounding factors should affect cases and controls to the same extent.

The study’s potential limitations also include the difficulty of tracking other potential outcomes of interest linked to diabetes, which would have required the use of non-validated algorithms for their identification, introducing additional uncertainties (ie, peripheral arterial disease).

One interest of our study is related to the method used to provide new inputs on the CV effectiveness of GLP1-RAs in real-life settings in France. Using both a matching approach based on diabetes treatment schemes before and after GLP1-RAs initiation and hdPS method for adjustment. Consequently, the results will be limited to estimate the impact of GLP1-RAs use in some treatment schemes currently reimbursed in France and should not be generalised to the overall use of GLP1-RAs in T2D. This approach ensures the best quality of comparisons at the cost of multiplying the number of results distinguished by therapeutic schemes before and at the time of GLP-1-RA initiation.

This protocol should provide new data in real-life settings in France related to the CV effectiveness of GLP1-RAs in some specific contexts of use.

Ethics and dissemination

In accordance with French legislation, the protocol has been approved by an independent ethics committee (Comité éthique et scientifique pour les recherches, les études et les évaluations dans le domaine de la santé, Paris, France; reference: 8699786, dated 2 June 2022) and has been registered with the French National Data Protection Commission (Commission Nationale de l'Informatique et des Libertés, Paris, France; reference: 922161, dated 26 June 2022). The findings of this study will be published in peer-reviewed scientific journals and presented at international conferences. The data will be consulted via the French National Health Insurance System (Caisse Nationale de l’Assurance Maladie) portal; the investigators’ access will be restricted to the scope of the study. The data will not be extracted from the main database but will be analysed in a dedicated project area on the server. The investigators will comply with the reference framework applicable to the SNDS database (as set out in the government act, dated 22 March 2017). The study protocol was registered at France’s Health Data Hub (www.health-data-hub.fr). The first results are expected in late 2025.