Prescribing of direct oral anticoagulants and warfarin to older people with atrial fibrillation in UK general practice: a cohort study

The incidence of OAC prescribing increased from 111 per 1000 person-years to 587 per 1000 person-years between 2003 and 2017. This upward trend was also seen when incidence was stratified by age band and sex (see Fig. 2). The incidence of OAC prescribing in the pre-DOAC era (2003–2007) was 122.6 per 1000 person-years. This increased to 164.2 per 1000 person-years in the era when DOACs were available but not yet recommended by NICE, then more than doubled to 387 per 1000 person-years in the era following NICE recommendation (2013–2017). Stepped Poisson regression (see Additional file 1: Fig. S1) highlighted this change in prescribing of DOACs compared with warfarin with a 6.5% increase each year (Incidence rate ratio (IRR): 1.065, 95% CI 1.06 −1.07) and a more than doubling prescribing of DOACs compared with warfarin (IRR: 2.25, 95%CI 2.14–2.37). Non-parametric test for trend of prescribing of warfarin reported a non-significant value (p=0.89) but prescribing of DOACs showed a significant increasing trend (p=0.008). However, given that the warfarin prescribing was not linear, Poisson regression was also used to separately model the prescribing of warfarin and prescribing of DOACS; the model coefficients were compared and these were significantly different (p<0.001).

From 2012 to 2017, i.e., when both DOACs and warfarin were marketed, the incidence of warfarin prescribing decreased and DOAC prescribing increased rapidly (Fig. 3). Apixaban was the most commonly prescribed DOAC followed by rivaroxaban then dabigatran. Edoxaban was licensed toward the end of the study period in 2015 and therefore had the lowest incidence of prescribing.

The prevalence of OAC prescribing to patients aged ≥ 75 years with a diagnosis of AF increased every year. In 2003, the point prevalence of OAC prescribing was 37.6% increasing to 75% in 2017. The point prevalence of warfarin prescribing to this group decreased from 54.8% in 2013 to 44% in 2017 whilst the prevalence of DOAC prescribing increased from 1.9% in 2013 to 31.1% in 2017.

Switching was calculated separately for those in the incident and prevalent groups. In the incident group, there were 5467 switches in 4566 patients and in the prevalent group, there were 821 switches in 621 patients. The majority of patients in both the incident (86.6%) and prevalent groups (79.2%) switched only once. A small proportion of patients (3.8% of the incident group and 6% of the prevalent group) had 3 or more switches during the study. The most common switch type was warfarin to DOAC, and the least common switch was DOAC to warfarin. Table 2 describes the number of switches and the proportion of each switch type in each group.

Switches were categorised as direct (no gap between OACs) or indirect (a period of time unexposed to any OAC before starting the next one). In both the incident and prevalent groups, the first switch was commonly a direct switch (73% and 78%, respectively). For direct switches, the median time on the index OAC prior to switching was 662 days (IQR 164–1451 days) for the incident group. For indirect switches, the median time on the index OAC was 251 days (IQR 54–869 days) in the incident group. The median unexposed gap between stopping the index OAC and starting the next OAC was 230 days (IQR 40–743 days) in the incident group and 101 days (IQR 35–427 days) in the prevalent group.

Figure 4 shows the risk difference (RD) in prescribing of OACs to patients with different demographics. Patients in the older age groups were less likely to receive an OAC than younger patients. Patients aged ≥ 90 years were 40% less likely to receive an OAC than those in the 75–79 year age group, and this difference was maintained in each time period. Women were slightly less likely to receive an OAC than men (risk difference around 6% in each period). Risk differences were also calculated based on smoking and alcohol status, but results are limited due to missing data (see Additional file 1: Fig S2-S3).

OAC prescribing was similar for CHA₂DS₂-VASC scores less than six (equating to an estimated annual stroke risk of < 9.7% [17]), but patients with a score greater than six (those at the highest stroke risk) were less likely to receive an OAC in all periods compared with those with a score of two or three (period 1: RD −0.06, 95% CI −0.07 to −0.05; period 2: RD −0.04, 95% CI −0.05 to −0.03; period 3: RD −0.06, 95% CI −0.07 to −0.05). Higher HAS-BLED scores (indicating an increased bleeding risk) were associated with a lower proportion of patients prescribed an OAC in all periods, but the difference was greatest in period one for HAS-BLED scores of 5–8 compared with a score of 0–2 (period 1: RD −0.12, 95% CI −0.13 to −0.11; period 2: −0.05, 95% CI −0.06 to −0.04; period 3: RD −0.03, 95% CI −0.04 to −0.01). Full results are shown in Additional file 1: Fig. S4.

Figure 5 shows the difference in OAC prescribing when a co-morbidity is present to when the co-morbidity is not present. The co-morbidities with the largest impact on prescribing were bleeds, dementia, falls and fractures. A history of major bleeding significantly reduced OAC prescribing (period 1: RD −0.09, 95% CI −0.11 to −0.07; period 2: RD −0.11, 95% CI −0.13 to −0.1; period 3: RD −0.17, 95% CI −0.19 to −0.15). Intracranial bleeding had the greatest impact on OAC prescribing and the impact increased over time (period 1: RD −0.12, 95% CI −0.15 to −0.10; period 2: RD −0.18, 95% CI −0.2 to −0.16; period 3: RD −0.25, 95% CI −0.27 to −0.23). The impact of dementia on OAC prescribing also increased over time, in period 1 patients with dementia were 23% less likely to receive an OAC than those without and this increased to 34% in period 3 (period 1: RD −0.23, 95% CI −0.23 to −0.22; period 2: RD −0.28, 95% CI −0.29 to −0.27; period 3: RD −0.34, 95% CI −0.35 to −0.33). A history of falls reduced OAC prescribing by 17% in all periods and fracture by 12%.

Patients with a prosthetic heart valve replacement or moderate-to-severe mitral regurgitation (formally known as ‘valvular AF’) were 5–10% more likely to be prescribed an OAC in periods 1 and 2 (pre-DOAC) than those without these co-morbidites; however, in period 3, there was no difference in OAC prescribing between those with and without these conditions (period 1: RD 0.1, 95% CI 0.07 to 0.14; period 2: RD 0.05, 95% CI 0.02 to 0.09; period 3: RD −0.01, 95% CI −0.04 to 0.03).

Patients prescribed antihypertensives, NSAIDs or statins were over 10% more likely to receive an OAC. Patients prescribed SSRIs or anticonvulsants were 10–14% less likely to receive an OAC. Figure 6 shows the full results.

Results from the univariate and multivariable logistic regression models are shown in Additional file 1: Tables S2-S3. Factors identified as potential confounders and added to the persistence analysis were dementia, age, intracranial haemorrhage, valve disease, a fall in the year preceding OAC therapy, acute kidney injury, a history of falls, a history of fracture, a history of stroke, TIA or thromboembolism.

Figure 7 shows the unadjusted Kaplan-Meier survival estimates in period 3. Failure (i.e., time to treatment discontinuation) was fastest with dabigatran followed by rivaroxaban, warfarin and apixaban had similar time to failure.

The unadjusted Cox model suggested that persistence with DOACs was lower than with warfarin as patients started on DOACs were more likely to stop therapy (HR 1.22, 95% CI 1.15–1.28). This result remained after adjusting for confounders (HR 1.16, 95% CI 1.11–1.23). However, the assumption of proportional hazards was violated with the basic Cox model so a time dependent effect was used to improve the modelling of the baseline hazard function by individual year of study. This showed that in the first and second years of treatment, patients were more likely to stop a DOAC than warfarin (0–1 year: HR 1.25, 95% CI 1.17–1.33; 1–2 years: HR 1.12, 95% CI 0.99–1.28), but from the third year onwards, patients were more likely to persist with DOACs than warfarin (HR 0.75, 95% CI 0.63–0.89). Results when comparing individual DOACs to warfarin in the first year of treatment (adjusted model) showed heterogeneity. Persistence was lower with dabigatran (HR 1.99, 95% CI 1.76–2.25) and rivaroxaban (HR 1.37, 95% CI 1.27–1.48) than warfarin, but apixaban (HR 0.96, 95% CI 0.88–1.05) was similar to warfarin. Persistence with edoxaban past the first year was not calculated due to low patient numbers.

In the first year of treatment, persistence was higher with apixaban than either dabigatran (HR 0.49, 95% CI 0.43 to 0.56) or rivaroxaban (HR 0.7, 95% CI 0.64 to 0.77). Persistence with rivaroxaban was also higher than dabigatran (HR 0.7, 95% CI 0.61 to 0.8).

Numerous studies have confirmed an increase in OAC prescribing over the last decade in the UK for prevention of stroke in AF in the general population [11, 18‐20] and to older people [21]. This change has also been observed at practice level, with general practices with a higher ratio of older people and a higher prevalence of AF more likely to prescribe DOACs [22]; however, there is a risk of ecological fallacy if wishing to extrapolate this practice level data to prescribing to individuals. Whilst some of the increase in prescribing of OACs may be due to the introduction of DOACs, it may also be attributed to the introduction of Quality Outcome Frameworks for AF which incentivised the diagnosis of AF and prescribing of anticoagulants [23, 24]. The increase in prevalence of OAC prescribing to older people has also been seen in other countries [25, 26]. The general increase in prescribing of OACs to patients aged ≥75 years is promising and suggests greater adherence to NICE and European Society of Cardiology (ESC) guidelines [4, 5]; however, prescribing to the oldest old is still significantly lower than for younger patients. Our study and others show that compared with patients aged 75–79 years, patients aged ≥85 years are up to 45% less likely to receive an OAC, and this association remains when other factors such as comorbidities are adjusted for [27‐31]. Terminal illness or palliative care may influence OAC prescribing at all ages, but particularly in the oldest old, however, the number of patients in our cohort with a Read code suggesting a palliative diagnosis at study entry was very small (n = 1686, 1% of total cohort).

Dementia/cognitive impairment [25, 27], a history of falls and bleeding [27, 29, 31] have been shown in previous studies to reduce OAC prescribing. Prior stroke has been associated with increased OAC prescribing [29, 31] but had no effect in our study. This finding is unexpected and may be due to other factors such as age or other comorbidities which were not adjusted for in our univariate model. Other studies have found that OAC prescribing is not significantly affected by contraindications to therapy such as major bleeding or haemorrhagic stroke [32], but this may be due to the fact that these so-called ‘contraindications’ are time dependent and a history of one of these contraindications does not necessarily preclude treatment for life. It is unsurprising that we found prior bleeds, and in particular major bleeds and intracranial haemorrhage to be associated with reduced OAC prescribing. Where serious bleeds have occurred, the risk of further bleeding needs to be balanced carefully with the risk of stroke.

Despite NICE guidance specifically recommending that anticoagulation not be withheld solely because a person is at risk of falls [5], this was not seen in our study. The reduction in OAC prescribing to people with falls remained the same in all three periods (spanning the time pre- and post- the 2014 NICE publication). There is strong evidence that patients on warfarin would have to fall around 300 times a year for the risk of bleeding to outweigh the benefit of stroke prevention [33, 34]. Studies of DOACs also suggest that patients are not at higher risk of brain injury following low level falls [35]; however, falling is still often cited by prescribers as a reason for avoiding anticoagulation [36].

Dementia is not a contraindication to anticoagulant therapy, but historically, it may not have been practical to prescribe warfarin to patients with cognitive impairment due to the complex dosing regimens. DOACs have a simpler dosing regime and can be added to compliance aids, but the gap in OAC prescribing to those with dementia compared with those without has actually increased since the introduction of DOACs. Comorbidities such as AF are common in older people with dementia, but they are often undertreated [37]. Whilst there is strong evidence that anticoagulation significantly reduces stroke risk in older people, few studies have investigated outcomes in patients with dementia [38], so it may be difficult for prescribers to weigh up the risk-benefit ratio in this group.

Studies comparing persistence with warfarin and DOACs have yielded conflicting results. Persistence is defined as the time from medication initiation to discontinuation. Where persistence with medications is compared, hazard or odds ratios are reported for the time to discontinuation. A ratio greater than one indicates that the medication of interest (in this case DOACs) is discontinued sooner than the comparator (i.e., warfarin). Previous studies from the USA and Italy have found persistence with dabigatran [39] and all DOACs [40, 41] to be better than warfarin after 12 months of treatment. A UK study found that persistence with dabigatran was lower than warfarin (HR 1.24, 95% CI 1.08–1.42) but higher with rivaroxaban (HR 0.85, 95% CI 0.77–0.93) and apixaban (HR 0.53, 95% CI 0.46–0.60) [42]. The average age of patients in this study was 74 years compared with 80 years in our study. Another more recent UK study also found persistence with DOACs to be better than warfarin, particularly in older age groups (aged 85 and over) [43], but the study consisted only of those who had attended for an influenza vaccine rather than all patients with AF as in our study. This may have led to selection bias, i.e., only including those who actively engage with healthcare. Neither of these UK studies used comprehensive prescription mapping and both assigned arbitrary end dates to prescriptions. Our study used all available prescription data, test data and clinical data to more robustly estimate exposure time and account for gaps in exposure which would not be seen using the methods described by Lund and colleagues [43]. Other European studies, however, have yielded results similar to ours. A Swedish study found persistence at 1 year to be similar between apixaban (OR 0.88, 95% CI 0.62–1.25) and warfarin but lower with dabigatran (OR 1.81, 95% CI 1.57–2.10) and rivaroxaban (OR 1.50, 95% CI 1.24–1.81) [44]. A German study found the same for persistence at 1 year for DOACs compared with warfarin: apixaban (HR 1.08, 95% CI 0.95–1.24), dabigatran (HR 1.53, 95% CI 1.40–1.68) and rivaroxaban (HR 1.21, 95% CI 1.14-1.29), although they noted that when time was partitioned by first 100 days of treatment that persistence with apixaban was lower than with warfarin [45]. These differences may be due to the methods used to calculate exposure, differences in cohort characteristics or differences in each country’s healthcare systems. Some studies used ‘proportion of days covered’ to assess persistence and assumed each prescription would last 30 days [42]. This method has been shown to work well for fixed-dose medications but may be inaccurate for medications with variable doses, such as warfarin, which could have affected the results [46]. The studies with results most similar to ours used different methods to define exposure to DOACs compared with warfarin to account for the variable warfarin dosing and used INR test results in addition to prescription data. Sensitivity analyses in these studies showed that when INR tests were not used, warfarin persistence decreased [40, 44, 45].

The major strengths of this study are that we used a large and representative sample of older people in the UK and extracted data from both the time prior to and post NICE approval to describe how prescribing has changed over time. We mapped OAC prescribing rather than relying on single prescriptions which allows us to consider switching and unexposed time. We have provided a wealth of data that could be used to inform future studies as it highlights a number of potential confounders that may lead to channelling bias if not considered when comparing outcomes with warfarin and DOACs in this population. A weakness of this study is that the CPRD only contains data of prescriptions and not whether these were dispensed or taken by the patient. Ultimately, this type of study cannot account for the various reasons that prescribers use to determine whether or not to prescribe an OAC. The prescription data for warfarin held in the CPRD does not accurately define the dose taken or how long a supply would last. Periods of exposure are based on estimations from an algorithm that includes clinical and test data (INR) in addition to prescription data to estimate exposure. This provides more robust estimates than prescription data alone. Diagnostic Read codes have been shown previously to accurately identify patients with AF in the CPRD with a low rate of false positives [47]. We further strengthened case identification by only including patients that had more than one diagnostic Read code for AF or evidence to support their diagnosis such as diagnostic codes or a change in prescribing that would support the diagnosis of AF. A limitation of this study is that we did not account for patients whose AF had resolved which could be a reason for not prescribing an OAC. Only a small number of patients (5% of those in the no OAC group) had a Read code suggesting that their AF had resolved during the study so this would not have significantly changed our results had we excluded them. We did not exclude patients with thrombotic disorders which may contraindicate OAC therapy; however, we would anticipate the number of patients with clotting disorders to be small. We have used the CHA₂DS₂-VASc score to define stroke risk in this study, but it should be acknowledged that this score was not recommended until 2010 [15]. The CHA₂DS₂-VASc score replaced the CHADS₂ score as it was found to better discriminate truly low-risk patients [15]. The guidance on when to anticoagulate has also changed over the study period: the 2006 NICE guidelines [48] recommended that those at moderate risk (as per the CHADS₂ score) should consider aspirin or an anticoagulant whereas the 2014 NICE guideline [5] no longer recommends aspirin. We have not considered aspirin use as it is no longer recommended, and our focus was on anticoagulation.