Evaluation of antithrombotic use and COVID-19 outcomes in a nationwide atrial fibrillation cohort

Study design and data sources

We conducted a cohort analysis using the newly established NHS Digital Trusted Research Environment (TRE) for England which provides secure, remote access to linked, person level EHR data for over 56 million people[12]. Available data sources cover primary care, secondary care, pharmacy dispensing, death registrations and COVID-19 tests and vaccines. For this study, we used the General Practice Extraction Service Extract for Pandemic Planning and Research (GDPPR) for demographic and diagnostic data (e.g. a diagnosis of AF) and the NHS Business Service Authority Dispensed Medicines (BSADM) for medication exposure data (e.g. pre-existing AT use) as this is the most accurate available representation of the medication an individual takes. Hospital Episode Statistics (HES), COVID-19 Hospitalisations in England Surveillance System (CHESS), Secondary Uses Service (SUS) and Office for National Statistics (ONS) Civil Registration of Deaths were used for COVID-19 hospitalisation and death. Public Health England’s Second Generation Surveillance System (SGSS) was used to identify COVID-19 test results and the COVID-19 vaccination events dataset was used for COVID-19 vaccine status.

Figure 1 provides an illustrative overview of how cohorts were designed to address the study objectives and the open protocol (https://github.com/BHFDSC/CCU020/tree/main/england/protocol) provides an overview, including all protocol amendments.

• Download figure
• Open in new tab

Figure 1: Illustrative overview of study design

Study populations

Individuals were included in the study if registered with a GP practice in England (at least one record in the GDPPR dataset with a valid person pseudo identifier), >= 18 years old and alive on January 1st 2020, had available sex, ethnicity and GP practice location data (based on most recent, available data across primary care (GDPPR), secondary care (HES) and death registrations (ONS)) and had a diagnosis of AF (coded in GDPPR) with a CHA₂DS₂-VASc score>=2 (calculated from the sum of components[13] coded in GDPPR).

Individuals with contraindications for sub-types of AT (e.g. DOACs in mitral stenosis, prosthetic mechanical valves, antiphospholipid antibody syndrome) were included as they are still eligible for other AT sub-types (e.g. AP, warfarin).

To investigate exposure to pre-existing AT use on COVID-19 related hospitalisation and death, inclusion criteria of a recorded COVID-19 event were applied. A COVID-19 event was defined as any of a positive test (polymerase chain reaction or lateral flow), a coded diagnosis in primary or secondary care or a COVID-19 diagnosis on a death certificate (see Thygesen et al (https://github.com/BHFDSC/CCU013_01_ENG-COVID-19_event_phenotyping/tree/main/phenotypes) for further details and phenotyping algorithms).

All phenotyping algorithms used are available on Github (https://github.com/BHFDSC/CCU020/tree/main/england/phenotypes). Supplementary Figure 1 provides an overview of the numbers of individuals excluded at each stage.

Study variables

Statistical analysis

Descriptive statistics were used to summarise the study population characteristics and were stratified by medication category. Pairwise Pearson’s correlation coefficients were used to check for potential collinearities between covariates. Multivariable logistic regression was used to test associations with pre-existing AT use and calculate the propensity score.

Multivariable logistic regression and Cox regression were used to test differences between exposure groups (any AT vs no AT, AC only vs AP only, DOACs vs warfarin) for COVID-19 related hospitalisation and death. These methods were selected to evaluate potential differences between event (logistic regression) and time-to-event (Cox regression) based analysis. All covariates including the propensity score were included in both methods. For variables with incomplete data (BMI – 9.3% missing), individual values were imputed with the cohort mean.

Two sensitivity analyses were conducted. Firstly, to evaluate the potential impact of different time periods, analysis was repeated for Jan 1^st 2020 – December 1^st 2020, prior to the introduction of vaccines and the December 29^th 2020 cases peak of the second wave[14]. Secondly, to validate the potential effect on COVID-19 specific outcomes, analysis was repeated with COVID-19 hospitalisation and death defined exclusively as the primary recorded diagnosis (coded first in hospital record or death certificate).

Primary results are reported from the multivariable logistic regression models covering the full time period (Jan 1^st 2020 – May 1^st 2021) with the other analyses reviewed for concordance.

Data preparation was performed using Python 3.7 and Spark SQL (2.4.5) on Databricks Runtime 6.4 for Machine Learning with analysis performed using R version 4.0.3. All code for data preparation and analysis is available on Github (https://github.com/BHFDSC/CCU020/tree/main/england/code) and full results available at this microsite (https://alexhandy1.shinyapps.io/at-evaluation-results/)

Ethical and Regulatory Approvals

Data access approval was granted to the CVD-COVID-UK consortium through the NHS Digital online Data Access Request Service (ref. DARS-NIC-381078-Y9C5K). The BHF Data Science Centre approvals and oversight board deemed that this project (proposal CCU020 Evaluation of antithrombotic use and COVID-19 outcomes) fell within the scope of the consortium’s ethical and regulatory approvals. Analyses were conducted by approved researcher (AH) via secure remote access to the TRE. Only summarised, aggregate results were exported, following manual review by the NHS Digital ‘safe outputs’ escrow service, to ensure no output placed in the public domain contains information that may be used to identify an individual[12]. The North East-Newcastle and North Tyneside 2 research ethics committee provided ethical approval for the CVD-COVID-UK research programme (REC No 20/NE/0161).

Evaluation of antithrombotic use

From a total of 55,903,113 individuals registered with a GP practice in England, 972,971 (1.7%) had a diagnosis of AF and a CHA₂DS₂-VASc score>=2 on January 1^st 2020 and 88.0% (n=856,336) had pre-existing AT use, with 74.3% (n=722,737) on AC only. The demographic and clinical characteristics of this cohort are summarised in Tables 1-3. By May 2021, the proportion of individuals on any AT had fallen to 87.7% but AC only had increased to 75.7% (see Figure 2). For individuals on any AT, warfarin prescriptions fell from 24.8% in January 2020 to 17.1% in May 2021 whilst DOACs rose from 60.3% to 69.5% (see Supplementary Figure 2).

• Download figure
• Open in new tab

Figure 2: Individual antithrombotic prescriptions by category January 2020 – May 2021

Ordered by proportion of prescriptions with mutually exclusive categories.

Factors associated with pre-existing AT use vs no AT are shown in Figure 3. Lipid regulating drugs (OR=2.50 [2.47-2.54 at 95% CI]) and antihypertensives (OR=1.90 [1.88-1.93 at 95% CI]) were associated with the highest odds of pre-existing AT use followed by comorbidities in the CHA₂DS₂-VASc score (stroke (OR=1.76 [1.72-1.79 at 95% CI], vascular disease (OR=1.60 [1.56-1.63 at 95% CI]). In contrast, nonsteroidal anti-Inflammatory drugs (NSAIDS, OR=0.39 [0.37-0.40]), liver disease (OR=0.56 [0.53-0.59]) and history of falls (OR=0.80 [0.79-0.82]) were associated with reduced odds.

• Download figure
• Open in new tab

Figure 3: Factors associated with AT vs no AT medication (January 1^st 2020) using multivariable logistic regression

Differences were also observed across demographics, ethnicity, socio-economic status and geographic location, with females (OR=0.91 [0.90-0.92]) and individuals from ethnic minorities and lower socio-economic positions associated with lower odds of AT use (e.g. ethnicity of black or black british vs white (OR=0.68 [0.64-0.72]). In other AT sub-types (AC vs AP and DOACs vs warfarin) results were broadly consistent (see Supplementary Material Figures 3-4), with the primary exception that vascular disease was associated with reduced odds of AC vs AP (OR=0.37 [0.36-0.38)]).

• Download figure
• Open in new tab

Figure 4: Comparison of AT medication exposures on COVID-19 outcomes (follow up to May 1^st 2021) using propensity score adjusted multivariable logistic regression

Antithrombotic use and COVID-19 outcomes

From 972,971 individuals that had a diagnosis of AF and a CHA₂DS₂-VASc score>=2 on January 1^st 2020, 8% (n=77,364) had a recorded COVID-19 event, 3.8% (n=37,418) had a COVID-19 related hospitalisation and 2.2% (n=21,116) died when followed up to 1^st May 2021. The summary characteristics of individuals with a recorded COVID-19 event are summarised in Supplementary Tables 1-3. Mean age (81) and comorbidities (mean CHA₂DS₂-VASc score 4.2) were both marginally higher compared to the full cohort. The proportion of individuals with pre-existing AT use was also marginally lower at 86.7%, but otherwise demographic and clinical characteristics were consistent.

Pre-existing AT use was associated with lower odds of COVID-19 death (OR=0.92 [0.87-0.96 at 95% CI]), but higher odds of COVID-19 hospitalisation OR=1.20 [1.15-1.26 at 95% CI]) (see Figure 4). The same pattern was observed for AC vs AP (COVID-19 death (OR=0.93 [0.87-0.98]), COVID-19 hospitalisation (OR=1.17 [1.11-1.24])) but not for DOACs vs warfarin (COVID-19 death (OR=1.00 [0.95-1.05]), COVID-19 hospitalisation (OR=0.86 [0.82-0.89]).

These results were all directionally consistent across Cox regression analysis and the sensitivity analyses (see Supplementary Material Figure 5-7). Full results are available on this microsite (https://alexhandy1.shinyapps.io/at-evaluation-results/).

Main findings

This study of routinely collected healthcare data from 56 million individuals in England is the largest scale evaluation of AT use to date. In 972,971 individuals with AF and a CHA₂DS₂-VASc score>=2, we observed 88.0% (n=856,336) with pre-existing AT use which was associated with marginal protection against COVID-19 death (OR=0.92 [0.87-0.96 at 95% CI]). Although this association may not be causal, it provides further incentive to improve AT coverage for eligible individuals with AF.

Of the AF cohort analysed, 8% (n=77,364) had a recorded COVID-19 event of which 3.8% (n=37,418) had a COVID-19 related hospitalisation and 2.2% (n=21,116) died. A marginally lower risk of COVID-19 death was observed for those with pre-existing AT use which directionally aligns with the most comparable previous studies[9,11]. AT use was, however, associated with higher odds of COVID-19 hospitalisation. This observation remained consistent when including only hospitalisations and deaths where COVID-19 was the first coded diagnosis. Higher observed risk of hospitalisation could reflect increased health seeking behaviour (both patient-driven or by a clinician) of those with pre-existing AT use or may indicate that any protective effect from AT materialises only in the most serious cases. The same pattern was observed in AC vs AP and supports the findings from Fröhlich et al that AC may offer more protection against death than AP[9]. For DOACs vs warfarin, no difference was observed between groups for COVID-19 death, but DOACs were associated with marginally reduced odds of COVID-19 hospitalisation. Our analysis did not directly investigate the previously reported observation that vitamin K depletion through warfarin is harmful[15] but, more generally, our findings suggest that it is unlikely that warfarin offers protection against severe COVID-19 outcomes compared with DOACs[11].

Although these protective associations across AT sub-types do not prove causality, they provide further incentive to improve AT coverage for individuals with AF that are already at high risk of stroke. Previous evaluations in the UK have estimated that around 15% of these individuals do not take any AT and around 17% take AP only rather than the recommended AC[3,6]. Our evaluation found around 12% on no AT and around 7% on AP only which suggests national level guidance[16] and primary care incentives such as the Quality and Outcomes Framework [17] continue to have a positive impact. Nonetheless, one in five individuals remains on a sub-optimal medication regime. Shifts from warfarin to DOACs observed in this study and others[18] were recommended by COVID-19 guidance[19] and demonstrate the potential impact of rapidly disseminated medications policy using population scale EHR data.

Identifying which factors are associated with AT use is key to further lowering the proportion of individuals on sub-optimal medication. NSAIDs displayed the strongest association with no AT use and likely reflects the association between NSAIDs and increased risk of major bleeding in individuals with AF[20]. For comorbidities, liver disease had the strongest association with no AT use, which is also supported by clinical evidence[21]. However, recent evidence suggests[22,23] more personalised risk calculations for bleeding and stroke may enable more individuals with liver disease to benefit from AT. History of falls was the comorbidity with the second strongest association with no AT use suggesting it remains a key factor in AT medicating decisions and may be overweighted as a proxy for bleeding risk[7,24,25]. In the UK, NICE guidance was recently updated[4] to explicitly address this issue and it will be important to track the impact of this in future evaluations. On demographics, lower odds of AT use was observed in women but this is likely influenced by using NICE’s primary threshold for the CHA₂DS₂-VASc score of 2 for both sexes. The CHA₂DS₂-VASc score allocates 1 point to females and 0 for males resulting in a larger proportion of comparatively health females (e.g. 12% and 25% of females in cohort have vascular disease or diabetes vs 21% and 33% respectively in males). However, demographic differences in AT use across ethnicity and socio-economic status mirror systematic healthcare inequalities that have been reported previously[26,27]. Targeted outreach to these groups will be key to improving AT use further.

Strength and limitations

To our knowledge, this study is the largest scale evaluation of AT use to date. Routinely updated, linked, population-scale EHR datasets provide the statistical power to robustly analyse targeted sub-groups and control for a wide range of potential confounders. The prevalence of individuals with AF and CHA₂DS₂-VASc score>=2 in our cohort is similar to that observed in the Quality and Outcomes Framework[17] which provides an external validation for our dataset. All code is opensource and an updated nationwide evaluation can be rapidly created for future time points.

The study does have limitations. Firstly, the reported associations do not demonstrate causality and residual confounding is unlikely to have been fully eliminated. For example, in-hospital treatment regimens were not analysed so differences in COVID-19 outcomes due to additional targeted anticoagulation regimens[28] or other medications such as dexamethasone[29] cannot be accounted for in our analysis. Whilst we attempted to mitigate confounding through careful cohort selection, covariates and propensity score adjustment, our study design does not control for all potential factors associated with the initiation of AT use which may influence COVID-19 outcomes. Lastly, exposure to AT medication was defined as one or more dispensed prescriptions (recorded in NHS BSA Dispensed Medicines) in the previous 6 months. Other studies have used different time periods and prescription frequency counts[9,11] and adherence was not measured[30]. We purposefully defined a liberal threshold to support evaluation of AT usage up to May 2021 that may have included unusual buying patterns (e.g. bulk buying) caused by the pandemic. The trade-off is that for the COVID-19 outcome analyses it increases the probability of including a minority of “exposed” individuals who had ceased regular, pre-existing AT medication.

Conclusions

Pre-existing AT use may offer marginal protection against COVID-19 death, with AC offering more protection than AP. Although this association may not be causal, it provides further incentive to improve AT coverage for eligible individuals with AF.