This study is based on the French national health database [Système National des Données de Santé (SNDS)], which has been widely used for epidemiological studies [21, 22], including for studies involving patients with atrial fibrillation initiating OACs [23,24,25,26,27,28,29]. A complete description of the database is available in previous publications [30, 31]. Briefly, a unique anonymous individual identifier links information from the outpatient care database [DCIR (Données de Consommation Inter-Régime)] and hospital care database [PMSI (Programme de Médicalisation des Systèmes d’information)]. The DCIR database includes outpatient medical care reimbursement, including drugs coded according to the Anatomic Therapeutic Chemical (ATC) classification, status with respect to full reimbursement of care for a severe long-term disease, coded using the International Classification of Disease, tenth revision (ICD-10), as well as demographic information (age, sex, and date of death). The PMSI database includes all hospital diagnosis discharges, recorded using the ICD-10 classification, and medical procedures.

The data access permission policy prohibits making the data set publicly available.

We conducted a longitudinal study on patients aged 75–110 years old, diagnosed with atrial fibrillation, and initiating an OAC treatment (no dispensing in the previous 2 years), either vitamin-K antagonist (VKA) or direct oral anticoagulant (DOAC), between April 2012 and 2016 (inclusion date). The 75-year threshold used to define our study population was defined in line with previous studies and recommendations concerning medication use in older people in France [32,33,34]. We restricted the study population to people affiliated with the French general health insurance scheme for at least 5 years at inclusion so that we could assess medical history. Atrial fibrillation was defined according to previous studies conducted on the SNDS (see Supplementary Methods for definition) [23,24,25, 28, 29]. Patients with contraindications to the use of OACs were excluded (see Supplementary Methods for exclusion criteria). In addition, patients who had received a PPI within 3 months prior to OAC initiation were excluded.

The exposure to PPI was defined as the reimbursement of at least one of the following ATC codes within the 14 days following inclusion: A02BC05, A02BC03, A02BC01, A02BC02, or A02BC04.

The outcome event was the occurrence of hospitalization for UGIB according to the following ICD-10 codes during the follow-up period (any diagnostic position in the hospital record): K250, K52, K254, K256, K260, K262, K264, K266, K270, K272, K274, K276, K280, K282, K284, K286, K290, K920, K921, or I850 (see Supplementary Table 1 for details about the codes). Melena has an upper gastrointestinal predominant origin, but it can also have a lower origin. To reduce the risk of identifying cases of lower gastrointestinal origin, we considered the diagnosis code K921 in the outcome definition unless the diagnosis code K625 (anus or rectum bleeding) was also recorded for the same hospitalization.

Covariates assessed at baseline included sociodemographic characteristics (age, sex, and quintile of deprivation index—a combination of four socioeconomic variables at the smallest administrative unit in France), 5-year medical history (cardiovascular, neurodegenerative, epilepsy or psychiatric diseases, diabetes, chronic kidney disease, liver disease, cancer, coagulation abnormalities, and hemorrhage), and drug exposure in the 4 months before inclusion (antihypertensive drugs, lipid-lowering agents, oral corticosteroids, nonsteroidal anti-inflammatory drugs, antiplatelet agents, anxiolytics and hypnotics, and analgesics). Concomitant dispensing of drugs that increase the risk of bleeding, namely nonsteroidal anti-inflammatory drugs, heparin and antiplatelet agent, and the total number of different ATC codes (0–4/5 to 9/10 drugs or more) at OAC initiation were also considered. The individual risk of bleeding was assessed using a version of the HAS-BLED score adapted for use in the French national health database [28].

In PPI users, the follow-up started on the first date of PPI dispensing within the 14-day period following OAC initiation (index date). This time window was chosen to include patients who had been prescribed a PPI by another physician in the days following OAC initiation. In PPI nonusers, the index date was randomly defined over this 14-day period so that the distribution of time between OAC initiation and index date was similar between PPI users and nonusers.

Patients were followed for 12 months or until the occurrence of UGIB, death, discontinuation of OAC treatment (defined as at least 3 months without OAC dispensing), or switch from a VKA to a DOAC and vice-versa (only for analyses by type of OAC), whichever came first. The time scale was time since start of follow-up (in days).

Cox proportional hazard models were used to estimate hazard ratios (HR) of UGIB between PPI users and nonusers. Different models have been implemented to progressively control for confounding and indication bias: a first model without adjustment to assess the crude association, a second model adjusted for confounders by multivariable Cox regression, and a final and main model controlling for indication bias by inverse probability of treatment weighting (IPTW) adjustment, including weight stabilization [35, 36]. All above mentioned covariates were used for adjustment. The distribution of stabilized weights was analyzed to test the positivity assumption [35]. The balance of covariates between the PPI users and nonusers before and after weighting was assessed by measuring the standardized mean difference. We conducted two parallel analyses, the first with an end of follow-up at 6 months, at the latest, and the second at 12 months. We explored two end-of-follow-up periods, as the risk of bleeding is higher at OAC treatment initiation, especially for VKA, which may require dosage adjustment. In addition, the HR estimate is a weighted average of all instantaneous risks over the study period. Estimating the HR over a shorter period allow us to test for potential variation of the effect over time [37].

In addition to the semiparametric approach of the Cox model, Kaplan–Meier cumulative incidence curves of UGIB adjusted for IPTW were also estimated [38, 39]. Bootstraping based on the standard deviation method was used to calculate the 95% confidence interval (200 replications).

To assess the effect of PPIs by sociodemographic variables and risk of UGIB, we repeated the 6-month analysis in the following subgroups: sex, age categories (75–89 years and ≥ 90 years), HAS-BLED score (score 1 to 2 and score ≥ 3), year of inclusion, and exposition to an antiplatelet agent at baseline.

Several sensitivity analyses were performed. First, we introduced an additional censoring event at PPI discontinuation in the exposed group and at PPI initiation in the unexposed group. In this analysis, we additionally weighted by the inverse of the probability of censoring for PPI discontinuation or initiation, respectively (IPTCW), using all baseline covariates and spline terms for modeling time trends in censuring [40, 41]. Second, to assess the specificity of the effect of PPIs on UGIB, analyses were repeated with nongastrointestinal bleeding as the outcome. The ICD-10 codes used to define nongastrointestinal bleeding are detailed in Supplementary Table 1. Other sensitivity analyses included: (1) defining PPI use based on the drugs dispensed on the same day as OAC initiation (versus within 14 days in the main analysis), (2) excluding patients with at least one PPI dispensation within 6 months before inclusion (versus 3 months in the main analysis), and (3) considering death as a competing risk to estimate the cumulative incidence curves and the risk for UGIB.

Analyses were performed using SAS, version 9.4 software (SAS Institute, Inc).