This retrospective cohort study was conducted between November 1, 2010, and December 23, 2022. Evaluable patients with NVAF and OSA initiating rivaroxaban or warfarin were identified between November 1, 2011, and December 23, 2022 (Online Resource 1; Supplementary Fig. 1). The primary objectives were to compare the incidence of stroke or systemic embolism and major bleeding and compare all-cause HCRU and costs during the follow-up period among patients who initiated rivaroxaban versus warfarin. Secondary objectives included a comparison of the risks of ischemic stroke, intracranial hemorrhage, systemic embolism, and major extracranial bleeds with rivaroxaban versus warfarin. Exploratory objectives were to compare the effectiveness and safety of rivaroxaban versus warfarin in subgroups of patients with NVAF and OSA by age, sex, presence of obesity, diabetes, heart failure, prior stroke, presence of metabolic syndrome (defined as the presence of various metabolic abnormalities, including obesity, hypertension, hyperlipidemia, and type 2 diabetes mellitus [9]), and use of positive airway pressure (PAP) therapy. Adherence to rivaroxaban and warfarin was also evaluated.

We used the IQVIA PharMetrics® Plus adjudicated health plan claims database, which includes information from > 70 contributing health plans and self-insured employer groups throughout the USA covering > 140 million unique individuals since 2006. This database includes medical and pharmacy claims data (costs and descriptive services), patient-level enrollment records, and patient demographics. This database is generally representative of commercially insured individuals < 65 years of age, with 6% having Medicare, 3% having Medicaid, and 1% having other insurance. The mean length of enrollment is ≥ 39 months, and ≥ 47 million patients have ≥ 3 years of continuous enrollment for medical and pharmacy coverage.

Patients newly initiating rivaroxaban or warfarin were identified. Eligible patients had ≥ 1 pharmacy claim for rivaroxaban or warfarin between November 1, 2011, and December 23, 2022, with the date of the first claim being defined as the index date; ≥ 1 medical claim with a diagnosis code for OSA (International Classification of Diseases, 9th/10th Revision, Clinical Modification [ICD-9-CM/ICD-10-CM] codes—327.20, 327.23, 327.29, 780.51, 780.53, 780.57, G47.30, G47.33, and G47.39) in any diagnostic position during the baseline period or on the index date; and ≥ 1 diagnosis of AF (ICD-9-CM/ICD-10-CM codes—427.31, I48.0x–I48.2x, I48.91x) during the baseline period or on the index date. Included patients were aged ≥ 18 years on the index date, had a moderate-to-high risk of stroke (CHA2DS2-VASc score ≥ 2 in men and ≥ 3 in women) per 2019 American Heart Association guidelines [3], and had ≥ 365 days of continuous health plan enrollment prior to the index date. Patients were excluded if they had ≥ 1 of the following: diagnosis code for valvular AF (e.g., presence of mitral stenosis or prior heart valve replacement) at any time during the baseline period, diagnosis code for other indications for oral anticoagulant use (e.g., venous thromboembolism, prophylaxis after knee or hip replacement), claim for an oral anticoagulant during the baseline period, or claim for pregnancy during the baseline period or on the index date. After checking the patients’ profiles, we restricted the analysis to those with an index date on or after 2013 to allow time for a recently approved medication to be prescribed and because early adopters may have different characteristics.

Clinical (effectiveness and safety) and economic (HCRU and costs) outcomes were evaluated using 2 follow-up approaches. The intent-to-treat (ITT) approach evaluated the overall risk of outcomes after initiating treatment. Patients were followed from the index date until the earliest of the following: first clinical outcome event, health plan disenrollment, or last data point. The on-treatment approach evaluated outcomes during the period in which patients were exposed to treatment. Patients were followed from the index date until the earliest of the first clinical outcome event, health plan disenrollment, last data point, or discontinuation of index treatment (defined as a gap of > 60 days between 2 prescription fills of index treatment or a treatment change to another oral anticoagulant).

The primary effectiveness outcome was the incidence (per 100 person-years of follow-up) of stroke or systemic embolism during the follow-up period. Stroke or systemic embolism was defined as ≥ 1 hospitalization with a primary diagnosis of stroke (ischemic or hemorrhagic) or systemic embolism (ICD-9-CM—430.x–432.x, 433.01, 433.11, 433.21, 433.31, 433.81, 433.91, 434.01, 434.11, 434.91, 436.x, 444.01, 444.09, 444.1, 444.21, 444.22, 444.81, 444.89, 444.9; ICD-10-CM—I60.x–I62.x, I63.x, I74.x). The primary safety outcome was the incidence (per 100 person-years of follow-up) of major bleeding after the index date, evaluated using a validated claims-based algorithm [16]. All-cause HCRU and costs incurred during the follow-up period were compared between treatments. HCRU outcomes included inpatient hospitalizations and length of stay (days), emergency department (ED) visits, outpatient visits (hospital outpatient, physician office [primary and specialty care], and skilled nursing facility visits), and prescription fills. All costs were inflated to 2022 US dollars and presented as mean costs per patient per year (PPPY) for total healthcare costs, which included the total medical and pharmacy costs. Medical costs included inpatient hospitalizations, ED visits, and outpatient visits. Pharmacy costs included only all-cause pharmacy expenses.

Secondary outcomes were the incidence (per 100 person-years of follow-up) of ischemic stroke, intracranial hemorrhage, and major extracranial bleeding, evaluated as separate outcomes. Ischemic stroke and intracranial hemorrhage were defined as ≥ 1 hospitalization with a primary diagnosis code of ischemic stroke or intracranial hemorrhage, respectively. Major extracranial bleeding included all major bleeding events from the Cunningham algorithm that did not occur in the head [16].

Exploratory objectives included comparisons of effectiveness and safety stratified by subgroups based on baseline characteristics of age (≥ 65 vs. < 65 years), sex, presence/absence of obesity (ICD-9-CM codes for body mass index ≥ 30 kg/m2), diabetes, heart failure, prior stroke, presence/absence of metabolic syndrome, and use of PAP therapy. Adherence to rivaroxaban and warfarin was measured by the proportion of days covered.

All patients within the database satisfying the eligibility criteria were evaluated. To adjust for potential confounding between rivaroxaban and warfarin cohorts, a multivariable logistic regression model [17] was used to calculate propensity scores, considering demographics (age, sex, race, insurance type, US census region, type of AF diagnosis) and baseline characteristics (comorbidities, continuous or bilevel PAP, sleep study, surgical treatment of OSA, intracranial hemorrhage, stroke or systemic embolism, transient ischemic attack, major bleeding, prior cardiovascular procedures, number of hospitalizations during baseline, CHA2DS2-VASc score, HAS-BLED score, and concomitant medications). Estimated propensity scores were used to weight patients using an overlap weighting approach [18], which assigns weights to patients that are proportional to their probability of belonging to the opposite cohort and adjusts for confounders in all eligible patients. This allows all eligible patients to remain in the analysis, resulting in an exact balance between the cohorts. Sufficient overlap of the propensity score between the cohorts was assessed through the calculation of equipoise.

Demographic and baseline characteristics for each cohort were summarized using descriptive statistics (categorical variables as percentages; continuous variables as means and SDs). Time-to-first-event outcomes were assessed using propensity score overlap weighted Cox proportional hazards regression models using a robust estimator to calculate hazard ratios and corresponding 95% CIs. All-cause HCRU was reported as PPPY, calculated as the number of events divided by total person-years of follow-up. Direct healthcare costs were reported as mean costs PPPY, calculated as the cost incurred during follow-up divided by the total person-years during follow-up. Between-group differences in HCRU and costs were estimated using a Poisson regression model with a log link (or negative-binomial models where appropriate to account for overdispersion). For exploratory subgroup analyses, propensity score models and weighting were rerun for each subgroup including the same variables as the primary analysis. To assess adherence, the proportion of days covered was calculated as the ratio of the number of days covered by the index medication prescription dispensed during the follow-up period divided by the number of days of follow-up. A proportion of days covered threshold ≥ 0.8 was considered adherent, and < 0.8 was considered nonadherent. The proportions of patients adherent to each treatment were compared with an odds ratio and 95% CIs.

All database management and statistical analysis were performed using SAS version 9.4 (SAS Institute, Cary, NC, USA). A P-value < 0.05 was considered statistically significant.