Introduction

Osteoporotic fractures are associated with high mortality and reduced quality of life in an elderly population [1]. Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia in the general population, being more prevalent in the elderly, and associated with an increased risk of mortality and morbidity from stroke and dementia [2,3,4,5].

The vitamin K antagonist (VKA, eg. warfarin) has been used for stroke prevention in persons with AF and has been associated with an increased risk for osteoporotic fracture [6,7,8,9]. By regulating vitamin K, warfarin inhibits the g-carboxylation of several proteins, including coagulation factors II, VII, IX, and X [10]. Preclinical studies have shown that several vitamin K–dependent proteins, such as matrix Gla protein and osteopontin, play a role in bone metabolism [9], and this has led to concerns that warfarin may increase the risk for osteoporotic fracture. These results correlate with clinical findings that propose a connection between warfarin and an increased risk of osteoporotic fractures [8,9,10,11, 14,15,16,17]. Furthermore, patients who are treated with VKAs are subjected to several dietary restrictions that may contribute to a low bone mineral density.

More recently, the non-vitamin K antagonist oral anticoagulants (NOACs) including dabigatran, rivaroxaban, apixaban, and edoxaban have been introduced for use as alternatives to warfarin. Indeed, the NOACs are now recommended over warfarin for stroke prevention in persons with AF mainly because they are at least as efficacious as warfarin in preventing stroke, have lower bleeding risks, and require less monitoring [11, 12]. The NOACs are also associated with a lower potential risk for drug–drug interactions than warfarin [13]. However, data on osteoporotic fracture risks with NOAC use are more limited [14,15,16,17], and it remains unclear which anticoagulant type or NOAC agent should be recommended as the first choice for a patient who is also at risk for osteoporotic fracture.

Given that oral anticoagulants are often prescribed to older adults who have multiple risk factors for osteoporotic fractures [18], further clarity on their associations with fracture risk is needed. This is particularly relevant to persons with AF, who were reported to have a higher risk of osteoporotic fracture and subsequent death after fracture than those without AF [19].

In this nationwide cohort study, we aimed to investigate the fracture risk among patients with AF treated with warfarin or NOACs, and second, to compare the fracture risks between the different NOAC agents.

Materials and methods

All data and materials have been made publicly available at the National Health Insurance Service (NHIS) of Korea. The data can be accessed on the National Health Insurance Data Sharing Service homepage of the NHIS (http://nhiss.nhis.or.kr). Applications to use the NHIS data will be reviewed by the inquiry committee of research support and, once approved, raw data will be provided to the authorized researcher with a fee at several permitted sites.

This study was a retrospective cohort analysis using the national health claims database (NHIS-2016–4-009) established by the NHIS of Korea. The NHIS is the single insurer managed by the Korean government. The majority (97.1%) of Korean citizens are mandatory subscribers to the NHIS, and the remaining 3% of the population are under the Medical Aid program. As the NHIS database contains the information of Medical Aid users, it is based on the entire Korean population [2,3,4, 20,21,22,23].

This study was approved by the institutional review board of the Yonsei University Health System (4-2016-0179), and the requirement for informed consent was waived.

Study population

From the Korean NHIS database covering a population 51.5 million inhabitants, we identified adults (aged 18 years or older) with AF who initiated oral anticoagulant (OAC) treatment (apixaban, dabigatran, rivaroxaban, edoxaban, or warfarin) between January 1, 2013, and December 31, 2016. We defined the date of the OAC prescription as the index date. AF was diagnosed using the International Classification of Disease 10th revision code I48. To ensure diagnostic accuracy, AF was defined as present only when it was a discharge diagnosis or confirmed at least twice in the outpatient department. The AF diagnosis has previously been validated in the NHIS database with a positive predictive value (PPV) of 94.1% [2,3,4, 20,21,22,23].

We excluded patients with valvular heart disease, OAC use prior to AF diagnosis, OAC prescription less than 30 days, catheter ablation for AF, or end-stage renal disease, and those who took unapproved NOAC dosage forms for prevention of AF-related stroke; dabigatran 75 mg; rivaroxaban 2.5 mg and 10 mg. We excluded patients who were exposed to any OACs within the previous year in order to establish an OAC-naive cohort. Patients taking edoxaban were excluded from this study due to their shorter follow-up times (the median follow-up of 5.6 months). Patients who had a record of bone tumors, epilepsy, or seizure before the index date or baseline use of hormone replacement therapy (on or within 90 days before the index date) were excluded to reduce their potential residual effects on fractures (Fig. 1) [24].

Fig. 1
figure 1

Flowchart of the enrollment and analysis of the study population. AF atrial fibrillation, NOAC non-vitamin K antagonist oral anticoagulants, OAC oral anticoagulant; RFCA, radiofrequency catheter ablation

Full size image

Outcomes

The primary outcome was major fracture requiring hospitalization including hip fractures, vertebral fractures, and fractures of the pelvis and acetabulum, which was defined from any discharge diagnoses (details in Additional file 1: Table S1). The diagnosis of major fracture has been validated previously in the NHIS database with PPV of 98% [19]. Patients were followed-up until the occurrence of the study outcome, switching to other OACs, death, emigration, or end of study (December 31, 2016), whichever came earliest.

Inverse probability of treatment weighting for multiple treatment options

To address potential bias due to nonrandomized treatment allocation, inverse probability of treatment weighting (IPTW) for multiple treatment options based on propensity scores was used to construct a weighted cohort of patients who differed with respect to oral anticoagulants but were balanced with respect to other measured characteristics [25]. The weights were derived to obtain estimates representing population average treatment effects with optimal balance between the treatment populations by using generalized boosted models based on 10,000 regression trees. The predictor variables in the propensity score model are presented in Table 1. The patients were considered to have comorbidities when the condition was a discharge diagnosis or was confirmed at least twice in an outpatient setting, similar to previous studies using NHIS data (Additional file 1: Table S1) [2,3,4, 20,21,22,23]. Balance between treatment populations was evaluated by standardized differences of all covariates, using a threshold of 0.1 to indicate imbalance. Characteristics with a standardized difference greater than 0.1 after IPTW were included as covariates in the subsequent regression model.

Table 1 Baseline characteristics before inverse probability of treatment weighting
Full size table

Statistical analysis

Baseline characteristics were expressed as means with standard deviations (SDs) for continuous variables and frequencies (percentages) for categorical variables. Incidence rates of dementia were calculated by dividing the number of events by person-time at risk. We compared the incidences of outcomes using the weighted log-rank test and plotted weighted failure curves. Competing risk regression by Fine and Gray using IPTW as a probability weight was performed to estimate the hazard ratio (HR) of the risk for osteoporotic fractures considering all-cause death as a competing event. The proportional hazards assumption was tested on the basis of Schoenfeld residuals. We performed subgroup analyses according to sex. Propensity scores and weights were separately re-calculated for the patients within the subgroups. To assess whether the observed differences in the risk of the primary outcome could be fully explained by an unmeasured confounder, we calculated the E-value for our HRs [26]. The E-value is defined as the minimum strength of association that an unmeasured confounder would need to have with both treatment and outcome, conditional on the measured covariates, to explain away an observed association [26]. A two-sided P-value of < 0.05 was considered significant. Statistical analyses were conducted using SAS version 9.3 (SAS Institute, Cary, NC, USA) and R version 3.5.3 (The R Foundation, www.R-project.org).

Results

Patient characteristics

A total of 70,481 new anticoagulant users met the inclusion criteria: apixaban (n = 10,879), dabigatran (n = 12,160), rivaroxaban (n = 18,055), and warfarin (n = 29,387) (Fig. 1). The mean age of the cohort was 70.5 years (SD 11.3), ranging from 68.5 years (warfarin) to 73.0 (apixaban) and 41.3% were female (Table 1). Median follow-up was 390 days (interquartile range 187–704 days). After IPTW, all baseline characteristics had standardized differences less than 0.1 (Table 2).

Table 2 Baseline characteristics after inverse probability of treatment weighting
Full size table

Risk for osteoporotic fractures

A total of 2412 fractures were identified (crude event number [weighted rate per 100 patient-years]: apixaban, 263 [2.56]; dabigatran, 273 [2.39]; rivaroxaban, 535 [2.78]; and warfarin, 1341 [3.43]). The crude median time to fracture (interquartile range) was 374 (172–672) days in overall anticoagulant users, ranging from 234 (106–386) days in apixaban users to 705 (337–1047) days in warfarin users (Fig. 2). Compared with men, women tended to have a higher incidence of osteoporotic fractures, regardless of the type of anticoagulant received (Fig. 2).

Fig. 2
figure 2

Risk of major fracture after inverse probability of treatment weighting. Incidence rates are per 100 person-years. CI confidence interval, HR hazard ratio, NOAC non-vitamin K antagonist oral anticoagulant

Full size image

The adjusted cumulative incidences at 6 to 24 months after treatment initiation are shown in Fig. 3. At 24 months, the weighted cumulative incidence of osteoporotic fractures was lower with NOAC use than with warfarin use (3.9% for apixaban, 4.2% for dabigatran, 4.8% for rivaroxaban and 5.4% for warfarin). In men, the weighted cumulative incidence was consistently higher in patients with warfarin use, whereas rivaroxaban and warfarin users had similarly higher incidences of 9.1% and 8.8% in women.

Fig. 3
figure 3

Weighted cumulative incidence curves of osteoporotic fracture

Full size image

Competing risk regression over the entire follow-up suggested that NOAC use was associated with a lower risk for osteoporotic fractures than warfarin use; HR 0.70 (95% confidence interval [CI] 0.57–0.86) for apixaban vs. warfarin, HR 0.69 (95% CI 0.60–0.78) for dabigatran vs. warfarin, and HR 0.79 (95% CI 0.70–0.90) for rivaroxaban vs. warfarin (Fig. 2). However, there were no differences observed between rivaroxaban and warfarin in women (HR 0.87, 95% CI 0.75–1.02) (P for interaction = 0.019). In head-to-head comparisons between NOACs, rivaroxaban use was associated with a higher risk for fractures than dabigatran use (HR 1.15, 95% CI 1.02–1.31) (Table 3). This observation was prominent for women with a HR of 1.22 (95% CI 1.05–1.42). No differences were observed in other head-to-head comparisons between NOACs.

Table 3 Risk of major fracture comparing inverse probability of treatment weighted users of NOACs
Full size table

The corresponding E-values for the point estimates of apixaban, dabigatran, rivaroxaban in comparisons with warfarin were 2.21, 2.27, and 1.85, respectively, which suggests that an unmeasured confounder would be needed to explain away the observed effect estimates only if the confounder was associated with both the treatment and the outcome by a risk ratio of approximately two-fold each, but weaker confounding could not do so.

Discussion

The principal finding of this study was that NOAC use was associated with a lower risk of osteoporotic fractures compared with warfarin use. Second, in head-to-head comparisons between NOACs, rivaroxaban use was associated with a higher risk for fracture than dabigatran, whereas there was no difference between dabigatran and apixaban. To our knowledge, this is the first study that shows significant differences in fracture risk according to individual NOACs, suggesting that rivaroxaban might be associated with a higher risk for fractures than other NOACs.

Accumulating evidence suggests that AF itself is an independent risk factor for osteoporotic fractures [19]. Especially in patients with AF, warfarin use has been suggested to have deleterious effects on bone density and be associated with a higher risk of subsequent fractures [7, 27], but the evidence remains controversial across the literature [28, 29]. The NOACs, which are emerging alternative anticoagulants, have no influence on the synthesis of osteocalcin and patients taking NOACs are not subjected to any dietary restrictions regarding several vegetables which could contribute to a low bone marrow density. A recent meta-analysis pooled the adverse events reported in randomized controlled trials of NOACs and found fewer reports of fractures in NOAC users than in warfarin users [30]. However, previous trials of NOACs were not designed to provide reliable estimates of fracture risks in clinical practice, and a range of population-based studies are needed to inform the risk for osteoporotic fracture for different oral anticoagulants.

Our finding that NOAC use is associated with a lower risk of osteoporotic fractures is consistent with several recent studies [14,15,16,17]. In the present study, we extended prior observations by enrolling a larger number of participants and allowing a longer follow-up. And we used propensity score-weighting for multiple treatment options, accounting for the differences in baseline characteristics of all the four anticoagulants (warfarin and 3 NOACs) simultaneously. This approach makes it possible to generalize the results to the entire population who would be eligible to receive any of the four anticoagulants, which may better reflect real-world clinical practice [17].

Prior studies have reported that warfarin therapy might interfere the process of bone formation. Warfarin not only antagonizes vitamin-K-dependent coagulation cascade but also impairs the γ-carboxylation of osteocalcin and other proteins which contributes to bone mineralization [9, 31]. Binding et al. suggested that the dietary restrictions regarding some vegetables recommended in warfarin users, which were associated with a low intake of folic acid and subsequent hyperhomocysteinemia, could explain the warfarin’s deleterious effects on bone health [14]. Hyperhomocysteinemia is associated with an increase in osteoclast activity and a decreased in osteoblast activity [32]. Furthermore, it is associated with an increase in matrix metalloproteinases that degrade extracellular bone matrix [33]. NOACs act independently of the vitamin-K associated mechanism, and there are no specific dietary restrictions in users of any of the NOACs. Gage et al. reported that long-term (≥ 1 year) warfarin use was associated with osteoporotic fractures in men with AF whereas prescribed warfarin for less than a year did not increase the risk of fracture [7]. Consistently, the median time to fracture after initiating anticoagulation was over a year (374 days) in this study, suggesting that the protective associations of NOACs with lower fracture risk over warfarin could be observed with the proviso that anticoagulants were prescribed for a sufficiently long time.

Evidence is scarce regarding possible differences in the fracture risk by individual NOACs [15, 17]. Prior studies analyzing administrative claim data of Hong Kong and Taiwan reported there were no significant differences in fracture risk in head-to-head comparisons between NOACs, although the strongest beneficial effect estimates were observed for apixaban when comparing individual NOACs with warfarin [15, 17].

To our knowledge, this is the first study to demonstrate significant differences in fracture risk according to individual NOACs, suggesting that rivaroxaban might be associated with a higher risk for fractures than other NOACs. Although the finding might be due to uncontrolled confounding such as NOAC dosing and treatment adherence, it might help to establish NOAC prescription strategies for minimizing the risk of fractures among patients with AF and high fracture risk. Randomized controlled trials investigating comparing the risk of fractures between NOACs are warranted.

Study limitations

The present study has several limitations. First, studies using administrative databases might be susceptible to errors arising from coding inaccuracies. To minimize this problem, we applied the definition that has been previously validated in previous studies using the Korean NHIS sample cohort [2,3,4, 20,21,22]. Second, given its observational nature, causal relationships could not be assessed and the possibility of unmeasured confounders cannot be ruled out. The E-value in this study suggested that a rare unmeasured confounder could explain our observed associations of NOAC use with lower fracture risk compared with warfarin only. Third, we did not have access to information on time in therapeutic range among warfarin users. Thus, our comparisons between NOAC and warfarin users should still be interpreted carefully. Lastly, the study enrolled only Asian patients, and it is therefore unknown whether the results in this study apply to other populations.

Conclusion

In patients with AF, NOAC use may result in a lower risk for osteoporotic fracture compared with warfarin use. Fracture risk does not seem to be altered by the choice of NOAC type, except for rivaroxaban. These associations may help inform benefit–risk assessments when choosing between the different anticoagulant types.