Gastrointestinal bleeding risk with rivaroxaban vs aspirin in atrial fibrillation: A multinational study

2.1 Data sources

This study used electronic medical records of the Clinical Data Analysis and Reporting System (CDARS) of the Hong Kong Hospital Authority and The Health Improvement Network (THIN) database in the United Kingdom. The Hong Kong Hospital Authority manages Hong Kong's public hospitals and is the only public-funder of healthcare in Hong Kong.⁹ It serves a population of >7 million individuals, covers 80% of hospital admissions, and provides medical services through 122 outpatient (specialist and general) clinics.⁹ Records from the Hong Kong Hospital Authority are anonymized and centralized in CDARS and include information on demographics, hospitalization, consultation and death dates, medication dispensing, test results, and diagnoses and procedures coded using the International Classification of Disease, ninth revision, Clinical Modification (ICD-9-CM). CDARS has been used extensively for population-based research^{10, 11} and has demonstrated high coding accuracy in validation studies.^{4, 10}

THIN is a longitudinal primary care database of 15 million individuals from 770 practices. Primary care physicians (PCPs) enter data, which describe patient characteristics, diagnoses, prescriptions, consultations, referrals and laboratory investigations. Diagnoses and procedures are categorized using the Read classification system. THIN data is periodically subject to internal quality checks, has been used extensively for population-based research, including bleeding studies.¹²

Study protocols were approved by the Institutional Review Board of the University of Hong Kong/Hospital Authority Hong Kong West Cluster (UW13-468) and by THIN Scientific Review Committee (19THIN021), respectively. Informed patient consent was not required as data were anonymized.

2.2 Study design and participants

This was a new-user cohort study. Patients aged ≥18 years with a new AF diagnosis between January 1, 2010, and December 31, 2017, were identified in CDARS using ICD-9-CM code for AF (427.3), and in THIN between August 1, 2011, and September 26, 2016, using Read codes for AF (G573.00, G573000, G573200, G573300, G573400, G573500, G573700, G573z00). We excluded patients with AF due to a reversible cause, who had valvular heart disease (Table S1) or who died during their first AF episode. To ensure accuracy of patients' records; in CDARS, patients with missing birth date or sex were excluded. Similarly, in THIN, patients with ≤12 months of medical history prior to AF diagnosis were excluded to ensure the capture of incident AF diagnosis.

The index date was defined as the date of the first prescription for rivaroxaban or aspirin following AF diagnosis. To capture GIB events after treatment initiation, and to reduce possible confounding, we excluded patients who received aspirin or any anticoagulant (warfarin, dabigatran, rivaroxaban, apixaban or edoxaban) within 180 days prior to, or on, the index date. Patients who were exposed to both drugs of interest (rivaroxaban and aspirin) on the index date were also excluded (Figure 1). Patients who used non-aspirin antiplatelet drugs at baseline were included, but this was controlled in propensity score modelling (eg, patients prescribed rivaroxaban plus clopidogrel or aspirin plus clopidogrel were allocated to rivaroxaban or aspirin groups, respectively).

2.3 Outcome and cohort follow-up

The study outcome of interest was GIB which included peptic, duodenal, gastrojejunal ulcers with haemorrhage, bleeding gastritis or duodentitis, gastrointestinal haemorrhage and intestinal haemorrhage (Table S1). Previous data validation has shown high coding accuracy for GIB in CDARS (positive predicted value [PPV] of 100%).¹⁰ Validation of GIB in UK primary care demonstrated PPV of 99%,¹³ and GIB Read codes have a 84 to 86% confirmation rate in THIN, specifically.¹²

Patient follow-up commenced from the index date until GIB occurrence, 30 days after treatment discontinuation or treatment switch (to aspirin or any anticoagulant), death, transfer out of the practice (UK patients), or study end (December 31, 2018, for CDARS analysis and September 26, 2017, for THIN analysis), whichever came first (Figure S1). Treatment discontinuation was determined by the time when patients failed to refill prescriptions within the allowable refill gap. The allowable refill gap was defined as the 95% percentile of the refill gaps observed in each treatment group (CDARS: 31 days for rivaroxaban and 29 days for aspirin; THIN: 5 days for rivaroxaban and 14 days for aspirin).¹⁴

2.4 Covariates

We included age at baseline, sex, index year and the confounders listed in Table S1 as covariates. Missing data in THIN for smoking status (0.2%), alcohol consumption status (6.8%) and body mass index (BMI) (5.9%) was addressed by multiple imputation (MI) using the fully conditional specification algorithm which created 25 imputed datasets. Twenty-five imputed datasets were selected as it is the default within the PROC MI procedure in SAS, it is also more than the suggested number (between 3 and 10¹⁵), is at least equal to the percentage of incomplete cases,¹⁶ and therefore is more likely to enhance precision in effect estimates.¹⁷ To minimize bias and enhance MI precision estimates, all potential confounders outlined were included in the MI model, as well as GIB outcomes and survival time.¹⁸

2.5 Statistical analysis

Statistical analyses were separately conducted in CDARS and THIN, using a common protocol approach. We used the propensity-score (PS) fine stratification weighting¹⁹ to control for confounding. The PS stratification with fine strata results in better covariate balance and increased precision of estimated treatment effects with low exposure prevalence and similar results at relatively higher exposure prevalence compared to other PS methods.¹⁹ Given the low rivaroxaban prevalence in CDARS (3.7%) and relatively higher rivaroxaban prevalence in THIN (43%), this was an optimal method for use in both cohorts. PS was estimated as the predicted probability of receiving rivaroxaban, conditional on covariates outlined, using logistic regression. For each study sample, the fine stratification weight was calculated by creating 50 PS strata and ranked rivaroxaban patients based on their PS. Aspirin users were then assigned to these strata based on their PS. Non-overlapping areas of PS distributions among treatment groups were trimmed. After stratification, weighted regression models were used, in which each rivaroxaban-treated patient received a weight of 1, and aspirin users were weighted in proportion to the distribution of the exposed stratum in which they were assigned.¹⁹ GIB risk between rivaroxaban and aspirin users was compared using Cox proportional hazards regression based on the weighted populations and presented using Hazard Ratios (HR) with 95% confidence intervals (CI). In THIN, analyses were undertaken separately for each imputed dataset (n = 25) and results were combined using Rubin's rule with PROC MIANALYZE in SAS (v9.4).¹⁸

Standardized differences were used to assess covariate balance. A threshold of 0.1 was considered negligible.²⁰ Baseline characteristics were presented as frequencies, percentages or medians (interquartile ranges [IQR]). A two-sided P-value of <.05 was considered statistically significant. Cohort construction was conducted separately in CDARS and in THIN. Statistical analyses were independently conducted on both datasets by two study authors for quality assurance. All data preparation and analyses were undertaken using SAS 9.4 (SAS Institute Inc, Cary, NC).

2.6 Additional analyses

Several additional analyses were conducted to test robustness of results: we accounted for the dosage effect of rivaroxaban by stratifying analyses by the standard daily dose of rivaroxaban (20 mg/day) and reduced daily dose of rivaroxaban (≤15 mg/day). As anticoagulants or aspirin can exacerbate bleeding in underlying GI lesions,²¹ patients with a prior history of peptic ulcer or GIB, and patients receiving gastric acid suppression therapy (proton pump inhibitor [PPI] or histamine type-2 Receptor Antagonist [H2RA]) were separately analysed. Previous data indicate women have increased DOAC-related GIB risk compared to men,²² therefore sex-specific analyses were conducted. Increasing age can also increase GIB risk,²³ therefore, analyses in patients aged ≥75 years were undertaken. Finally, the maximum allowable refill gap was extended to 99th percentile of gaps in both cohorts.¹⁴

3.1 Patient characteristics

Following patient exclusions, the CDARS cohort comprised 28 161 aspirin initiators and 1052 rivaroxaban initiators; and THIN comprised 8053 aspirin initiators and 3496 rivaroxaban initiators (Figure 1). The median (IQR) follow-up time was 209 (49-864) days in CDARS and 174 (86-422) days in THIN.

Median age of the CDARS cohort was 78 years and 14 721 (50%) were women. In THIN, the median age was 74 years and 5377 (47%) were women. Before propensity score modelling, there were differences in the presence of heart failure, vascular and renal disease, and use of ACEI/ARB, beta-blockers, H2RAs and statins between new users of rivaroxaban and aspirin in CDARS; and differences for hypertension, diabetes, prior ischaemic stroke/transient ischaemic attack/systemic embolism, renal disease, and concurrent use of antiplatelet drugs, calcium channel blockers and statins in THIN. After propensity score modelling, both cohorts were well balanced across covariates (standardized differences <0.1). Low-dose rivaroxaban (≤15 mg/day) was used in 27% of patients in CDARS, compared to 17% in THIN (Table 1). For all patients, low-dose aspirin in Hong Kong comprised 80 mg/day and 75 mg/day was used in the United Kingdom.

3.2 Risk of gastrointestinal bleeding

Crude GIB rates were 3.0 and 2.6 per 100 patient-years, for aspirin and rivaroxaban users, respectively, in CDARS; and 1.3 and 2.4 per 100 patient-years, respectively, in THIN. In both cohorts, the adjusted GIB event rate was higher in patients with a history of GIB or ulcer (CDARS: 5.3 [aspirin] and 5.8 [rivaroxaban] per 100 patient-years; THIN: 4.8 [aspirin] and 4.8 [rivaroxaban] per 100 patient-years), and in patients aged ≥75 years (CDARS: 4.1 [aspirin] and 3.2 [rivaroxaban] per 100 patient-years; THIN: 2.5 [aspirin] and 3.2 [rivaroxaban] per 100 patient-years) (Table 2).

In the propensity score adjusted samples, median time to GIB after first prescription for aspirin was 182 (IQR = 44-783) days in CDARS and 146 (IQR = 86-339) days in THIN; and was 408 (IQR = 108-905) days for rivaroxaban in CDARS and 254 (IQR = 122-525) days in THIN. Results for Cox proportional hazards regression analysis showed no difference in overall GIB risk between rivaroxaban or aspirin initiators in both cohorts (HR = 1.04, 95%CI = 0.76-1.42[CDARS]; 1.40, 1.00-1.98[THIN]). For additional analyses, the HR point estimates for CDARS and THIN tended to be in different directions, but overall, no statistically significant differences in GIB risks were observed for additional analyses, including prior use of PPIs or H2RAs (HR = 0.99, 95%CI = 0.60-1.62[CDARS]; 1.66, 0.92-2.98[THIN]), history of GI bleeding or ulcer (HR = 1.12, 95%CI = 0.55-2.28[CDARS]; 1.26, 0.61-2.61[THIN]), sex-specific analyses of women (HR = 0.72, 95%CI = 0.44-1.20[CDARS]; 1.50, 0.90-2.51[THIN]), increasing age (≥75 years) (HR = 0.84, 95%CI = 0.57-1.24[CDARS]; 1.28, 0.83-1.98[THIN]), or for an extended refill gap (HR = 0.96, 95%CI = 0.71-1.31[CDARS]; 1.31, 0.96-1.79[THIN]). Rivaroxaban use was significantly associated with increased GIB risk in men in CDARS (HR = 1.52, 95%CI = 1.02-2.27) (P for interaction with sex = .02), however, no significant difference in GIB risk was found for men in THIN (HR = 1.67, 95%CI = 0.99-2.82) (P for interaction with sex = .77) (Table 3).

When standard dose rivaroxaban (20 mg/day) was compared to aspirin, results demonstrated significantly increased GIB risk in rivaroxaban users in THIN (HR = 1.57, 95%CI = 1.08-2.29). However, no significant difference was observed for GIB risk in CDARS for rivaroxaban 20 mg/day vs aspirin (HR = 1.21, 95%CI = 0.84-1.74). For additional analyses, standard dose rivaroxaban was also significantly associated with increased GIB risk in THIN in patients using PPIs/H2RAs (HR = 1.96, 95%CI = 1.14-3.37), in women (HR = 1.91, 95%CI = 1.08-3.40; P for interaction with sex = .53) and for an extended refill gap (HR = 1.44, 95%CI = 1.03-2.02) (Figure 2A). No differences were observed among all additional analyses for standard dose rivaroxaban vs aspirin in CDARS. Finally, no difference in GIB risk was observed for low-dose rivaroxaban (≤15 mg/day) vs aspirin in THIN or CDARS across all primary and additional analyses (Figure 2B).

4.1 Strengths and limitations

Our study is the first to investigate comparative GIB risk for rivaroxaban vs aspirin in patients with AF in routine care. A key strength of the current study is the utilization of longitudinal data of the largest territory-wide clinical data source in Hong Kong and representative data of the United Kingdom. Use of two different data sources enabled broad investigation of GIB risk in diverse patient groups. Study limitations include: GIB event rates in both cohorts were small, therefore results should be interpreted with caution. Consistent with most clinical databases, CDARS and THIN does not capture over-the-counter aspirin use. However, the Hospital Authority is the only public healthcare provider in Hong Kong, where medications are highly subsidized (85%-98%).²⁸ Similarly, in the United Kingdom, most chronic aspirin use is obtained via prescription as it is free for patients aged >60 years, and there is confirmed minimal unrecorded usage.²⁹ Ultimately, uncaptured over-the-counter aspirin use is unlikely to have impacted results. This study accounted for confounding factors using propensity score methods; however, residual confounding could remain. Finally, we focused our analysis on a new-user cohort, GIB risk may be different among prevalent users.