This study utilized Centers for Medicare and Medicaid Service (CMS) Medicare data. This database contains medical and pharmacy claims from 100% national Medicare data, which includes Medicare Parts A–D and claim level information like beneficiary demographic information, International Classification of Diseases (ICD) diagnosis, procedure, CMS Healthcare Common Procedure Coding System (HCPCS) codes, diagnosis-related groups, dates of service, reimbursement amounts, provider numbers, etc. It amounts to approximately 38 million fee-for-service beneficiaries.

A retrospective study design was utilized to identify patients with NVAF that switched from warfarin to a DOAC (apixaban, dabigatran, rivaroxaban) from January 1, 2013 to December 31, 2019. The study designated a 12-month baseline period prior to the identification period (January 1, 2012 to December 31, 2019) while the follow-up period was designated from the day after the index date to death, end of study period, discontinuation, switch, or end of continuous medical and pharmacy enrollment, whichever occurred first. Index date was defined as the first DOAC prescription date within 90 days of warfarin discontinuation (Fig. 1).

Study periods (for illustration purposes only, may not be proportional). Bars and lengths of bars are not proportional to time periods. OAC oral anticoagulant

Switching was defined as having ≥ 1 pharmacy claims for warfarin and ≥ 1 pharmacy claims for a DOAC within 90 days after warfarin treatment ended during the identification period. In the primary analysis, only the first switching episode from warfarin to a DOAC was considered for each patient to ensure consistency in evaluating treatment outcomes. Treatment end date was defined as end of days of supply. To ensure consistency in evaluating outcomes and to avoid potential biases from including multiple episodes, only the first episode of switching from warfarin to DOAC was considered for each patient in the analysis. Treatment episodes were defined as the time from the DOAC episode start date to the earliest of DOAC discontinuation, switch, end of study period, or death disenrollment (multiple episodes per patient were allowed, but only the first was considered). Discontinuation was defined as no evidence of index DOAC treatment use for 30 days from the last day of the days’ supply. The discontinuation date was the last day of the days’ supply of the last filled prescription. The follow-up period was censored at 30 days after the index drug discontinuation date.

Patient eligibility was based on selection from the CMS Medicare database according to the following criteria. Inclusion was based on patients having AF diagnosis (ICD-9-CM: 427.31; ICD-10-CM: I480-I482, I4891) with at least 1 pharmacy claim for a warfarin or DOAC during the identification period at any time during the identification period. Patients were allowed to have multiple medications. For patients and episodes where patients were ≥ 65 years of age on the initial oral anticoagulant (OAC) prescription date (treatment date), and for; patients and episodes where patients had a continuous health plan enrollment with medical and pharmacy benefits for ≥ 12 months pre-treatment date (baseline period), no gap was allowed. Patients and episodes where patients had ≥ 1 medical claim for AF prior to or on the treatment date were all included in the study population.

All pregnant patients or those with no follow-up period were excluded from the study. Also excluded were: patients who had episodes of medical claims indicating a diagnosis or procedure code indicative of rheumatic mitral valvular heart disease or valve replacement procedure during the 12 months prior to or on the treatment date; episodes of a diagnosis of venous thromboembolism during the 12 months prior to or on the treatment; episodes of medical claims indicating hip/knee replacement surgery within 6 weeks prior to the OAC prescription date; and episodes of medical claims indicating a diagnosis or procedure code of transient AF (heart valve replacement/transplant, pericarditis, hyperthyroidism, and thyrotoxicity) during the 12 months prior to or on the OAC prescription date. Episodes with > 1 OAC treatment on the index date were also excluded, while patients on edoxaban were excluded from analyses due to the small sample size.

To evaluate switching among patients with NVAF from warfarin, three cohorts were created: switch to apixaban, to dabigatran, and to rivaroxaban. Cohorts used to assess the risks of stroke/SE, MB between DOACs among patients who discontinued warfarin were: apixaban versus dabigatran, apixaban versus rivaroxaban, and dabigatran versus rivaroxaban.

Patient demographics included age, sex, US geographic region, Medicaid dual eligibility and part B low-income subsidy. The Deyo–Charlson–Quan Comorbidity Index [21], CHADS2 score, CHA2DS2-VASc score [22], and HAS-BLED scores [23] were calculated using standard methods. Additionally, several comorbidities beyond the CCI (Charlson Comorbidity Index) were included, such as liver disease, renal disease, prior bleeding history, hypertension, dyspepsia, non-stroke/SE peripheral vascular disease, stroke/SE, transient ischemic attack, anemia and coagulation defects, peripheral arterial disease, coronary artery disease, etc.

Lifestyle factors or exposures such as alcohol use disorder were captured and included in the comorbidities. All comorbidities and lifestyle factors were captured using ICD-9-CM or ICD-10-CM codes on at least one inpatient or outpatient claim during the 12-month pre-index period.

Stroke/SE was identified during the follow-up period using hospital claims that included a stroke/SE diagnosis code as the primary listed ICD-9-CM or ICD-10-CM diagnosis code, and equaled 1 if there was ≥ 1 stroke event during the follow-up period. The first stroke/SE event was stratified into 3 categories: ischemic stroke, hemorrhagic stroke, and SE.

MB was identified in the follow-up period using hospital claims that included a MB diagnosis code as the primary listed ICD-9-CM or ICD-10-CM diagnosis code, and equaled 1 if there was ≥ 1 MB event during the follow-up period. The first MB event was stratified into 3 categories: gastrointestinal hemorrhage, intracranial hemorrhage, and other MB.

Time to stroke/SE or MB was calculated from the day after the index date to the date of stroke/SE or MB. The incidence rate of stroke/SE or MB was calculated per 100 person-years.

Descriptive analyses were conducted for baseline demographic and clinical factors, such as CHA2DS2-VASc and HAS-BLED scores, concomitant medications, and comorbidities across cohorts. Means, standard deviations, median, and inter-quartile range for continuous variables and differences across cohorts were calculated for continuous variables. Frequency and percentages were produced for categorical variables. Differences between groups were compared using t tests for continuous variables, and Chi-square tests for categorical variables. All variables were analyzed descriptively for each drug (apixaban, dabigatran, and rivaroxaban). Numbers and percentages were provided for dichotomous and polychotomous variables. Differences between cohorts (apixaban vs. dabigatran, apixaban vs. rivaroxaban, and dabigatran vs. rivaroxaban) were assessed using McNemar’s test for categorical variables and pairwise t tests for continuous variables.

The 1:1 propensity score matching (PSM) method was used to adjust the baseline patient differences in each pair. Propensity scores were estimated by unconditional logistic regression analyses which incorporated potential predictors of treatment as independent variables in the regression and group status. The propensity score was defined as the probability of being treated with each of the DOACs, based on a set of baseline characteristics in the DOAC cohort. All baseline variables were evaluated as covariates to be matched between treatments using the PSM technique and as covariates to be included in multivariate models.

Cox proportional hazards (PH) models were used to compare the time at risk for stroke/SE and MB between apixaban versus dabigatran, apixaban versus rivaroxaban, and dabigatran versus rivaroxaban cohorts. The proportional hazards assumption was met for most of the models. In instances where the assumption was violated, interaction terms with time were included in the model to account for non-proportional hazards. These adjustments were made in the analysis to ensure that the reported hazard ratios reflect the time-varying nature of the relationship between the covariates and the outcomes where necessary.

All statistical tests were two-sided, with α < 0.05 and performed using SAS software (version 9.4; SAS Institute, Cary, NC, USA).

Since this study did not involve the collection, use, or transmittal of individually identifiable data, Institutional Review Board approval was not required. This study only analyzed de-identified data which are a priori exempt from the Federal Policy for the Protection of Human Subjects (1991) and do not meet the identification criteria necessary to be privileged under the Health Insurance Portability and Accountability Act (HIPAA). Both the data and the security of the offices where the data were kept met HIPAA requirements. All authors had permission to access and use the database utilized in this study.