The association between atrial fibrillation and dementia: A UK linked electronic health records cohort study

We investigated the association between atrial fibrillation (AF) and dementia, and its subtypes (vascular‐VaD, Alzheimer, mixed and rare dementia), and identified predictors for dementia in AF patients.


| BACKGROUND
Countries worldwide are witnessing an increase in size and proportion of older persons (>65 years old) in the population.In 2021, over 65 s accounted for 18.9% of the UK population, with a 23% increase between 2009 and 2019, outpacing overall UK population growth of 7%. 1 This ageing population, expectedly, is accompanied by an increase in prevalence of older age-related diseases.
Atrial fibrillation (AF) has an estimated prevalence of 2.5% in England, with approximately 1.4 million patients said to be living with the condition in 2016. 2 Globally, AF poses significant public health and economic burden due to its strong association with adverse multi-morbidity (e.g.stroke, heart failure, myocardial infarction and chronic kidney disease) and mortality, 3 and its incidence is on the rise, having recently been described as having reached the level of a 21st-century epidemic. 4n 2019, 850,000 people with dementia were estimated to be living in the UK, with a rise to 1.5 million expected by 2040. 5Similar to AF, dementia has a high public health and economic burden, with the total cost of dementia in England estimated to be £24.2 billion in 2015. 6Prevalent dementia subtypes include Alzheimer's disease (AD) and Vascular dementia (VaD); however, studies have also identified the coexistence of these, termed mixed dementia.With no curative treatments for dementia, this places a strong importance on preventative strategies, including through modifiable prognostic factors across a patient's life.
Research on the relationship between AF and dementia, has produced inconsistent results, [7][8][9] and the association of AF with dementia has not yet been replicated with UK data.Accordingly, the National Institute for Health and Care Excellence (NICE) Clinical Knowledge Summary for AF does not include dementia as a potential complication. 2urthermore, identifying predictive factors, both risk and protective, can further support AF patients understanding of their risk own of dementia, and their individual need for prophylactic intervention to prevent irreversible brain damage.
Using linked UK electronic health records, we aimed to: (1) determine if there is a significant association between AF and dementia, and its prevalent subtypes (VaD, AD, mixed and rare dementia), and (2) identify predictors for dementia in AF patients.

| METHODS
Data from the Clinical Practice Research Datalink (CPRD) was accessed through the caliber platform, comprising deterministically linked anonymised electronic health records data of 6,529,382 individuals, from 727 general practices (GP), between 01 January 1998 and 31 May 2016 across the United Kingdom (UK).
Access to data was approved under the CPRD Research Data Governance operating framework (09/H0810/16), with approval of an AF specific protocol (ISAC 17_205R).We utilised linked data across multiple domains, including primary and secondary care (Hospital Episode Statistics, HES), to provide a comprehensive view of a patient's healthcare experience (Figure S1) CPRD data have demonstrated to be representative in terms of age, sex and ethnicity, of the UK population. 10Research has also shown both AF and dementia diagnoses to have high validity in CPRD. 11,12 patient's study entry date was defined as the latest of their CPRD entry date, up-to-standard GP date, current registration date or 55th birthday.A patient's index date was considered as either this study entry date (for non-AF patients) or AF diagnosis date.Patients were followed up to (end date) the earliest of dementia diagnosis, death or the end of the overall study period (31 May 2016).Mortality data were obtained through linked Office for National Statistics (ONS) Death Registration data.Patients were right-censored if they were not diagnosed with dementia within their study period (Figure S2).An inclusion criteria (Table S1) was applied to all caliber patients to select eligible patients.
Eligible AF patients were randomly matched on a 1:2 ratio to eligible non-AF patients to create a matched cohort study.Matching was performed on age at index date and sex.Patients were matched on age within 5 years, to increase efficiency of the matching algorithm and minimise non-matches.
Variables were extracted from Health Data Research UK (HDRUK) validated phenotype library (Table S2). 13The primary exposure was AF, identified using the validated phenotype. 11Patients who developed dementia were identified using the validated HDRUK phenotypes.Dementia diagnosis date was classified as the earliest clinical code record (VaD, AD, rare or other) in either CPRD or HES.Rare dementia includes Lewy body dementia, Parkinson's disease, Pick's disease, Huntingdon's disease, Creutzfeldt-Jakob disease and HIV-related dementia.If a patient had a subsequent non-'other' dementia diagnosis (VaD, AD or rare), then this was used.If a patient had both VaD and AD clinical codes, they were classed as mixed.Patients phenotyped as having 'possible' dementia were classified as dementia free.Figure S3 illustrates this implementation.
Study covariates were measured at index date and were selected based on their association with AF and dementia, expert opinion and availability within caliber (Table S3).Social deprivation was measured using Index of Multiple Deprivation (IMD) quintiles from ONS and linked through practice ID.
Information on prescribed medications was obtained from CPRD.Patients were classified as being on a medication class at index if a prescription event was recorded within the previous year.Code lists for all variables extracted from the linked data sources can be found in Table S2.
Descriptive statistics for AF and non-AF patients were computed, using means and standard deviation (SD) for continuous variables, and counts and percentages for categorical variables.Survival statistical techniques, primarily Cox Proportional Hazards (CoxPH) regression, were employed to model associations between baseline variables and dementia risk.Proportional hazards were tested using Schoenfeld residuals.If breached, models were stratified by time-varying coefficients.For non-proportional hazards in unadjusted models, weighted Cox regression, was also used to generate an unbiased average HR.This model utilises a weighting function to weight individuals at specific event times proportional to expected number at risk (of event) if censoring did not occur.Weights were truncated at the 95th percentile to reduce effects of extreme weights.
Using the matched cohort, the relationship between AF and dementia was evaluated.Incidence rates (dementia events per 100 person-years) and Kaplan-Meier cumulative incidence curves were computed.In addition, unadjusted and adjusted Cox models produced HRs for the risk of all-cause dementia and dementia subtypes.Sensitivity analyses included censoring patients at first stroke event (ischaemic, subarachnoid haemorrhage or intracerebral haemorrhage) after index and complete case analyses in adjusted models.Sub-analyses performed for age, using the median as the cut-off value.
Using only AF patients in the matched cohort, prognostic factors for dementia were identified.Cox regression models were sequentially adjusted for sociodemographic factors, comorbidities, lifestyle measurements and prescribed medications.Survival time was calculated from AF diagnosis date to their end date (Figure S4).All research was carried out within the secure University College London Data Safe Haven.Data preparation and analyses were performed using r studio (v4.1.2).Table S4 lists key r libraries used.No patients or public were involved in the design, or conduct, or reporting, or dissemination plans of our research.

| Study population and baseline characteristics
Of 6,529,382 patients, 2,200,094 patients were eligible for this study, with 190,428 having a diagnosis of AF during the study period (Figure S5).Of 183,610 AF patients were successfully matched to 367,220 non-AF patients, to create an overall analytical study cohort of 550,830 patients.Of 6818 AF patients were un-matched due to no suitable non-AF matches and were, therefore, excluded.
Table 1 summarises the analytical cohort.The AF cohort had a total follow-up time of 715,107 years, compared to a longer follow-up of 2,677,286 years in the non-AF cohort.Median follow-up time was shorter for AF patients (2.67, IQR .65-6.02 vs. 5.84 years, IQR 2. 26-11.80).The mean age of AF diagnosis was 74 years (SD: 9.43) for men and 79.3 years (SD: 9.63) in women.
Of 73,008 developed dementia within their follow-up period.Among those with specified dementia types, 47.1% had VaD, 46.7% had AD, and 6.2% had rare dementia types.Patients with AF developed VaD most commonly (60.9%), while those without AF developed AD most frequently (51.6%).The majority of patients in the study with recorded ethnicity were white (79.6%).Both AF and non-AF patients were recruited generally proportionally from IMD quintile areas, with a slight skew towards lower deprivation areas.AF patients tended to have increased comorbidities, with 60.6% having multi-comorbidity (two or more morbidities), compared to 39.4% in the non-AF group.A higher proportion of AF patients were classified as obese, than non-AF patients.Conversely, more non-AF patients were recorded as current smokers and drinkers.AF patients were more commonly on prescribed cardiovascular medications.

| AF and all-cause dementia
Incidence of dementia in the AF cohort was 2.65 per 100 person-years, compared to 2.02 in the non-AF cohort.The log-rank test demonstrated strong evidence for a difference in these cumulative incidences (Figure 1).
From an unadjusted CoxPH model, those with AF had a 36% higher hazard rate of dementia (95% CI: 1.34-1.39)versus those without AF (p < .001).As observed in Figure 1, the two cohorts cumulative dementia incidence diverges with time, with a sudden increase observed at around 18 years in the AF cohort.To address this nonproportionality, the model was stratified by the time at which there was a significant change in the HR.This time-stratified model also demonstrated strong evidence the hazard of dementia increased with AF (HR: 1.20, 95% CI: 1.18-1.22;p < .001),further confirmed through the weighted Cox regression model (HR: 1.43 95% CI: 1.41-1.46;p < .001).
When adjusting for all other measured covariates, the increased risk of dementia for AF patients remained.The association remained after censoring for stroke, and stratifying by age, with patients aged ≥77 years having a 20% increased risk compared to patients <77 years (Table S5).

| Dementia subtypes
Associations between AF and dementia subtypes over time were analysed (Figure 2).A time-stratified CoxPH model identified AF patients to have an 76% increased risk of VaD (95% CI: 1.70-1.82,p < .001)versus non-AF patients.This significant association remained when using a weighted Cox model (Table S6).AD was slightly more prevalent in the non-AF cohort (.50 vs. .45per 100 person-years, log-rank p < .001).A time-stratified CoxPH model identified AF patients to have a 19% reduced risk of AD (95% CI: .79-.86).However, no similar evidence of this association was found in the weighted Cox model.Mixed dementia was more prevalent in the AF cohort (.11 vs. .08per 100 person-years).There was strong evidence AF increased the risk of developing mixed dementia in both time-stratified (HR: 1.29, 95% CI: 1.18-1.40)and weighted Cox models (HR: 1.58, 95% CI: 1.43-1.75).
Finally, patients with AF who developed rare dementia had a slightly higher (p = .001)cumulative incidence rate (.07 per 100 person-years), compared to non-AF patients (.06 per 100 person-years), with timestratified, and weighted Cox models both demonstrating no evidence for an increased risk of rare dementia in AF patients.
After adjusting for baseline covariates (Figure 3 and Table S6), the association between AF and VaD remained, and the association of AF and AD was nonsignificant (Figure 3).When censoring for incident stroke, similar estimates were seen across the subtypes.When testing effect modification by age, significant interaction was found for VaD (p = .003),AD (p < .001)and mixed dementia (p = .008)subtypes, with no significant interaction by age in rare dementia (p = .224).patients that had stopped drinking or were on statins on study entry were at higher risk.Variables with evidence for protective effects included BMI (overweight and obese), Asian ethnicity and being on diuretics or beta-blockers on study entry.

| Prognostic factors for dementia
When focusing on VaD, similar predictors were identified (Table 3 & Figure S6).Unlike Cox models for overall dementia, age was not in breach of the PH assumption, and therefore, estimates were derived for this variable.All other covariates met the PH assumption.Hazards of VaD increased with age, IMD (5th quintile), hypertension, diabetes, intracerebral haemorrhage, ischaemic stroke, exdrinker and previous use of statins.Conversely, significant protective factors included BMI (overweight and obese), no smoking habits and being on diuretics or beta-blockers.

| Use of anti-coagulants and dementia
Whilst use of anti-platelets dropped slightly after incident AF, with only 32.2% of patients being on these agents 2 years after the AF diagnosis, the use of anti-coagulants increased from 7.2% to 34.5% within a 2-year time interval.Use of anti-coagulants at the time of AF diagnosis or any time following AF diagnosis was associated with a lower risk of all-cause dementia (HR = .72,95% CI: .58-.88) and VaD (HR = .61,95% CI: .42-.88).

| DISCUSSION
Using a UK representative dataset, we observed that AF is associated with all-cause dementia, and specifically VaD and mixed dementia.No significant association with AD was found.Stroke, subarachnoid haemorrhage, intracerebral haemorrhage and diabetes were identified as the strongest predictors of dementia in AF patients.
The observed associations of AF with dementia remained after censoring for patients with incident stroke, indicating clinical stroke may not be the primary link, a finding supported by an analysis of the Korea National Health Insurance Service-Senior cohort. 14A possible alternative explanation for this relationship is the presence of asymptomatic stroke, with subclinical silent cerebral ischemia in AF patients leading to cognitive decline and dementia onset. 15f the dementia subtypes, VaD demonstrated the highest risk, a finding supported by previous research. 7,14nterestingly, we did not observe an association between AF and AD, which is in line with previous studies. 9,16,17roviding support to our findings, a recent Mendelian randomization showed neutral effects, and hence no T A B L E 2 Estimates of the relationship between selected covariates and dementia in AF patients.causality, of AF-related variants on Alzheimer dementia. 18n the other hand, the same investigation provides support for the association of AF with all-cause dementia and vascular dementia, with a strong association for the latter.Sub-analyses of the data suggest that ischemic stroke and low cardiac output may be some of the mechanisms underlying the association. 18The reported association between AF and Alzheimer's in previous studies 7,14 may be due to ethnicity differences or use and coding of electronic health records.We note that in our cohort nearly 5% of patients were classified as having mixed dementia (diagnosis of both VaD and AD at some point).Contrary to other studies, we analized this subset as a separate group and it found to be significantly associated with AF.We believe that had these patients been coded as having AD, the results could have been biased and suggesting a spurious association, possibly driven by coexisting VaD.We observed that AF was not associated with rare dementia, an unsurprising finding given the heterogenous subtypes within this classification (e.g.frontotemporal dementia, Lewy body dementia or Creutzfeldt-Jakob disease), a finding which has also been confirmed by Li et al. 18 via Mendelian randomization.Strong evidence for patients ≥77 years having much higher risk of mixed dementia was found, with no evidence for those <77 years.As far as we are aware, this is the first observational study to examine relationships between AF and rare/mixed dementia.

Sex
We identified multiple variables that are established dementia prognostic factors, such as sex, ethnicity, multiple deprivation, hypertension, heart failure, subarachnoid haemorrhage, intracerebral haemorrhage, stroke, diabetes and smoking. 19,20Our analyses also suggest that dementia risk reduces as BMI increases according to the results of a pooled analysis of 39 cohort studies. 21Kivimaki et al. 21uggest that the association of BMI with dementia is timedependent, with high BMI recorded >20 years before a dementia diagnosis being considered harmful, but when measured closer to a dementia diagnosis (our average follow-up time from new-onset AF to dementia diagnosis was around 3 years) it seems to have a protective effect due to reverse causation, as weight loss during the preclinical dementia phase may occur.
Previous studies have shown that hypertension burden over time is more strongly associated with subsequent than a one-off measurement at baseline and that subsequent blood pressure control may lower the risk. 22ur data suggests previous or non-drinkers have increased dementia risk, compared to current drinkers.Available evidence on this relationship is mixed, with some studies indicating light to moderate alcohol consumption provides a protective effect, 23 which may explain our results. 23However, reverse causation may play a role in this case, as people may stop drinking prior to their dementia diagnosis due to pre-dementia symptoms or concomitant conditions.Data from the Korean National Health Insurance database suggest that clustering of healthy lifestyle behaviours, defined as no current smoking, alcohol abstinence and regular exercise, has been associated with lower risk of dementia in patients with new-onset diagnosed AF. 24 This is the first study to identify VaD specific predictive factors in the AF population.Age was seen as a significant predictor for VaD, with >10% greater risk per year increase.No significant risk difference between sexes was found for VaD, a finding corroborated by other research. 25ost of the remaining predictors overlapped with the ones identified for all-cause dementia.
Our analyses were not designed for testing the effect of medications on the risk of dementia: we have not assessed duration of exposure or accounted for patients started on any of the drug classes weeks or months after new-onset AF.Treatment with statins, beta-blockers and diuretics at the time of new-onset AF was associated with all-cause and VaD.The increase in risk associated with statins use at baseline may be attributable to statins acting as a proxy for high cholesterol or previously diagnosed atherosclerotic disease.An inverse relationship was found between the use of diuretics and dementia has been previously suggested. 26Similar to other authors we observed an of beta-blocker use at baseline with and lower incidence of dementia. 27Data from a Danish national registers suggests that bloodbrain barrier permeable beta-blockers may reduce the risk of some forms of dementia, like Alhzeimer's. 28owever, this contradicts observations of increased risk of VaD in patients treated with beta-blockers in the Malmö Preventive Project cohort. 29Adding to the uncertainty, a large Mendelian randomization study with phenome-wide association study of drug classes used for the treatment of hypertension failed to establish causality for this association. 30Our data suggest a protective role for anti-coagulants used at the time of AF diagnosis or started following this.Further assessments of the strength and exposure duration following new-onset AF, and having more data on the use of direct oral anti-coagulants may provide more reliable estimates on the association of drugs commonly used to treat AF patients and dementia.

| Clinical implications and future research
Efforts should be made to increase awareness of AF as a risk factor for dementia, specifically VaD.Targeted prevention of AF could reduce dementia burden within society.Given the prognostic factors identified, dementia risk profiles for AF patients can be determined.Patients with increased risk should undergo closer monitoring, with early dementia screening to identify those in the prodromal phase.With an increased risk of VaD specifically, screening for this subtype should be encouraged.These risk profiles could also support identification of higher risk populations for future dementia prevention clinical trials.Hence, our findings have the potential to encourage UK and European policy change, as well as stimulate further research.

| Strengths and limitations
The major strength of our study is the analytical cohort size.With a high number of events per covariate, we can be confident the study has sufficient power to provide precise estimates.Secondly, the data source has been proven broadly representative of the UK population, and therefore, results can appropriately be generalised. 10dditionally, use of linked primary and secondary care data sources reduced risk of disease misclassification.Linked data also allowed for analysis of a wide range of covariates, including sociodemographic factors.Finally, this study had a longer average follow-up period than comparable studies, which given the progressive development of dementia is crucial.
Some limitations need to be highlighted in our analysis.As for any observational study there is a risk for unmeasured confounders (e.g.education, social isolation and genetic factors).Additionally, while patients without a clinical recording were assumed disease free, it is possible some were misclassified, with the quality and of CPRD/HES data reliant on physician entry.Furthermore, our analyses on impact of drugs on the risk of dementia have focused mainly on drugs used at the time of study entry, and we acknowledge that some drugs may have been started following the AF diagnosis.Anti-coagulants, are an important example of this limitation, and, as such, additional analyses were added which suggest a protective role for these agents.Finally, analysis by AF subtype was not performed, as this specificity was not available in the HDRUK phenotype.

| CONCLUSION
This is the largest study to explore the relationship between AF and dementia and the first using UK's electronic health records.AF patients have an increased risk of dementia, independent of stroke, with highest risk of VaD.This study also successfully identified stroke, subarachnoid haemorrhage, intracerebral haemorrhage and diabetes as the strongest predictors of dementia in AF patients.Management and prevention of these could be crucial to reduce the increasing burden of dementia.

F I G U R E 2
Cumulative incidence of (A) vascular dementia, (B) Alzheimer's disease, (C) mixed dementia and (D) rare dementia.Shaded regions indicate 95% confidence intervals.FI G U R E 3Adjusted hazard ratios for all-cause dementia and dementia subtypes (AF vs. non-AF patients).
Baseline characteristics of study participants at index date, stratified by AF status.
T A B L E 1

AF (n = 183,610) Non-AF (n = 367,220)
Table 2 presents the results from CoxPH models for prognostic factors for all-cause dementia.Female, black ethnicity, most deprived (IMD 5th quintile), previous heart failure, history of hypertension, diabetes, previous cranial haemorrhage, ischemic stroke, being underweight and

HR, 95% CI, p-value
Estimates of the relationship between selected covariates and VaD in AF patients.
T A B L E3 (Continued)