Screening strategies for atrial fibrillation: a systematic review and cost-effectiveness analysis

The condition should be an important health problem

Atrial fibrillation is a major cause of ischaemic stroke, with a fivefold increase in risk of ischaemic stroke5 and associated increased mortality,5,6,27–29 increased morbidity15 and reduced quality of life. 30 AF incidence and prevalence increases with age, and a recent report with 50 years’ follow-up indicates that the incidence and prevalence of AF are increasing over time, largely because of an ageing population and increased detection. 31 It has been estimated that AF is responsible for about 25,000 AF-related strokes per year in the UK. 6 AF therefore represents an important health problem that is likely to increase over time with an ageing population.

The natural history of the disease, including from latent to declared disease, should be adequately understood and there should be a detectable disease marker for the latent period/early symptomatic stage of the disease

The natural history of AF is a progression from paroxysmal to persistent to permanent AF. 32 Risk factors for AF-related stroke risk in patients are well described by the CHA₂DS₂-VASc score. 11 All types of AF may be either symptomatic or asymptomatic. 33,34 It is expected that, compared with routine practice, a population-based screening programme is likely to detect a higher proportion of asymptomatic cases that would not otherwise present until a stroke or cardiovascular event occurs. We would also expect that, unless an extended screening test is used, the majority of screen-detected cases would be persistent or permanent AF, because only a small proportion of paroxysmal cases would be in AF at the time of testing. Flaker et al. 35 found that stroke risk was higher in symptomatic patients than asymptomatic patients, but that this difference was explained by baseline characteristics.

Cost-effective primary prevention interventions should have been implemented as far as practical

Across the UK there are programmes in place to help prevent modifiable risk factors of AF (MI, angina, diabetes mellitus, hyperthyroidism, stroke or TIA11,19).

There should be a simple, safe, precise and validated screening test, and the distribution of the test values in the target population should be known and a suitable cut-off level defined and agreed

A 12-lead ECG interpreted by at least one cardiologist is considered a gold standard for the diagnosis of permanent or persistent AF. However, in a screening programme quicker/cheaper screening tests are likely to be adopted, with confirmation of a positive result using the gold standard diagnostic test as deemed necessary. In one arm of the Screening for Atrial Fibrillation in the Elderly (SAFE) study, Hobbs et al. 19 used pulse taking as a screening test with a follow-up ECG if an irregular pulse was found, whereas in another arm a 12-lead ECG interpreted by a cardiologist was used directly as a screening test. However, there are a variety of possible screening tests available (e.g. modified blood pressure monitors, one-lead ECG, three-lead ECG, Holter monitor), as well as devices used for other reasons that can also detect AF [pacemakers, implantable cardioverter defibrillator (ICD) and implantable loop recorder devices]. There are also newer, smaller, non-obtrusive technologies currently being trialled (patches, smartphone/watch devices, hand-held devices) that may be of relevance to a screening programme. In this report, we consider only non-invasive tests for which information on sensitivity and specificity compared with a gold standard (12-lead ECG interpreted by at least one cardiologist) is available.

A screening programme would generate patients identified with an irregular pulse on the screening test used (e.g. pulse taking), some of whom may then need a confirmatory 12-lead ECG. There is some evidence that the sensitivity and specificity of diagnostic tests depend on the interpreter [nurse, general practitioner (GP), cardiologist or computer algorithm]. 19,36 Therefore, a confirmatory ECG would add an additional burden to the NHS, either through training GPs to interpret them or through referral to cardiology units.

Long-term continuous Holter monitoring is considered appropriate to detect paroxysmal AF, although there is no accepted gold standard for the diagnosis of paroxysmal AF. This may change with the introduction of newer, non-invasive devices for home monitoring. Note, however, that a 12-lead ECG interpreted by a cardiologist would detect only the small proportion of patients with paroxysmal AF who happened to be in AF at the time of testing. A population-based screening programme with only short-term monitoring is unlikely to detect many paroxysmal AF cases.

The test should be acceptable to the population

The SAFE study reported an approximately 50% uptake rate for screening and, of those identified as having an irregular pulse, 73% had an ECG. In a post-screening questionnaire, only 3.7% reported that screening was not convenient. 19 No adverse results were reported.

There should be an agreed policy on the further diagnostic investigation of individuals with a positive test result and on the choices available to those individuals

We assumed that AF detected using a screening test would be confirmed using a 12-lead ECG, as set out in the NICE clinical guidelines. 11 The gold standard is a 12-lead ECG interpreted by a cardiologist; however, increasingly GPs are being trained to interpret ECGs, with referral to a cardiologist only in unclear cases. For the purposes of our economic model we assumed that, following a screening test, there would be a diagnostic 12-lead ECG interpreted by a GP with referral to a specialist to help with interpretation if necessary. Following a confirmed diagnosis, treatment and care options are set out in the NICE clinical guidelines. 11

There should be an accepted treatment/therapy for patients identified through early detection, with evidence of early treatment/therapy leading to better outcomes than later treatment/therapy

As the morbidity and mortality associated with AF-related stroke are substantial, therapy to reduce AF-related stroke risk earlier rather than later is likely to lead to better outcomes; however, earlier treatment also implies more adverse events (bleeding events). The cost-effectiveness of anticoagulant therapy (warfarin or DOACs) to prevent stroke in eligible AF patients has been established,37 indicating that it is an efficient use of NHS resources. However, the trials informing these assessments were not conducted in screen-detected cases. Mant et al. 38 found that screen-detected AF patients had a lower risk of stroke than diagnosed AF patients, but they found no evidence that the relative efficacy of warfarin compared with aspirin depends on whether AF is screen detected or not. However, the low event rates mean that this result is very uncertain, and the data were consistent with both a beneficial and a harmful effect of warfarin.

Other treatment options include rate control and rhythm control therapies. As screen-detected AF is likely to be asymptomatic, we would assume that a small proportion would be offered rate control until their heart rhythm disorder progresses to be symptomatic. Cardioversion may be offered to younger patients to avoid long-term anticoagulation; however, in an elderly population-based screening programme this would be a small proportion of screen-detected cases.

There should be agreed evidence-based policies covering which individuals should be offered therapy and the appropriate therapy to be offered

The NICE clinical guideline11 clearly sets out who is eligible for anticoagulation and rate and rhythm control according to stroke and bleeding risk scores (see Current practice in the diagnosis and treatment of atrial fibrillation). Screen-detected patients would be eligible for anticoagulation only if the CHA₂DS₂-VASc score is ≥ 2, taking bleeding risk into account using the HAS-BLED score.

Clinical management of the condition and patient outcomes should be optimised in all health-care providers prior to participation in a screening programme

It has been estimated that around 35% of patients with diagnosed AF who are eligible and who do not have a contraindication are not receiving anticoagulation. 39 In addition, the introduction of DOACs has not led to uptake as expected. The potential therapy benefits from screening may, therefore, not be realised.

There should be evidence from high-quality randomised controlled trials that the screening programme is effective at reducing morbidity or mortality

There are currently no randomised controlled trials (RCTs) evaluating a population-based screening programme that have collected morbidity or mortality outcomes. The STROKESTOP trial26,40 that is currently under way aims to fill this evidence gap, but is not due to report until 2019. RCT evidence on long-term outcomes is limited to the follow-up period available and so may not be able to fully answer this question. There is RCT evidence available on the prevalence of screen-detected cases under different screening strategies;19 however, it is currently an assumption that early detection and treatment leads to morbidity and mortality benefits.

There should be evidence that the complete screening programme (test/diagnostic procedures/therapy/treatment/intervention) is clinically, socially and ethically acceptable to health professionals and the public

There have been several publications in the clinical literature calling for a national screening programme for AF,16,41,42 including a consensus statement from the Royal College of Physicians of Edinburgh in 2012. 43 This suggests that such a programme would be acceptable to clinicians. A recent survey44 of health-care professionals found that general practices have access to resources for AF screening and there is enthusiasm among non-GP health-care professionals to train and play a role in AF screening. However, there were concerns about a lack of staff, time and capacity.

Individuals invited to screening may choose to attend or not and so it is reasonable to assume that screening is acceptable in those that take up the screening opportunity. There may be a variety of reasons why some individuals do not attend for screening. It is important that this is an informed choice and that sufficient information is provided on the screening test and what positive and negative results mean, including the need for subsequent diagnostic tests and the treatment and therapy options that are available.

Patients diagnosed with AF are fearful of the consequences of a debilitating AF-related stroke. It is therefore important that they understand their risk of AF-related stroke and are able to make an informed choice about anticoagulation therapy after consulting with their health-care professionals. The options should be fully explained to them based on their eligibility and suitability and the risks and benefits of treatment, and should be reviewed periodically. NICE has produced Patient Decision Aids for this purpose. 45

Warfarin needs to be dose adjusted with regular monitoring to ensure that the patient stays within the recommended therapeutic international normalised ratio (INR) range, making warfarin unacceptable to some patients. 46,47 Warfarin also has interactions with other medications including over-the-counter preparations, foods high in vitamin K and alcohol, which can impact on quality of life, meaning that some patients are reluctant to take warfarin. Recent NICE guidance DG1448 has recommended coagulometers, devices that allow patients to self-monitor the INR, which is an alternative to regular INR monitoring, although some patients have experienced difficulties gaining access and support for these devices [Diane Eaton, AntiCoagulation Europe (ACE), 2016, personal communication]. DOACs do not require regular INR monitoring and have fewer drug and food interactions, making them more acceptable to many patients, although there are some concerns around lack of antidotes if there is a bleed (although dabigatran now has an antidote and others are in development).

Patients need to understand that, although anticoagulation therapy reduces the risk of AF-related stroke, it does not delay or reduce symptoms of AF. The importance of adhering to the medication in the absence of relief from symptoms therefore needs to be stressed repeatedly.

The proportion of eligible patients being prescribed DOACs is increasing, but is still relatively low. 39 If a population-based screening programme were to be put in place, the benefits of the programme would be realised only if screen-detected AF patients were put on appropriate anticoagulation therapy (e.g. not aspirin monotherapy, which is no longer recommended by NICE11). Therefore, there needs to be support at a local clinical commissioning group level to provide access to anticoagulation therapies recommended by NICE to eligible patients detected through screening.

The benefit from the screening programme should outweigh the physical and psychological harm (caused by test, diagnostic procedures and treatment)

The screening tests for AF investigated in this report are non-invasive and unlikely to cause physical harm. The SAFE study recorded the acceptability of screening to individuals and measured anxiety and quality of life before and after screening. 19 The authors found screening to be acceptable to the majority of individuals, but anxiety was higher and quality of life was lower in those who received a positive diagnosis of AF following screening. A 12-lead ECG, used to confirm the diagnosis of AF, is also non-invasive, although it requires the placement of pads on the upper body under clothing. The benefits in terms of reducing the risk of AF-related stroke are likely to outweigh adverse events (bleeding risk) and risk factors for bleeding on anticoagulant therapy are also well described by the HAS-BLED score. 11 However, as described earlier, there is not yet robust evidence of the impact of anticoagulation therapy in a screen-detected AF population.

The opportunity cost of the screening programme (including testing, diagnosis, therapy and treatment, administration, training and quality assurance) should represent value for money

Hobbs et al. 19 found that a systematic opportunistic screening programme is likely to be cost-effective in men and women from the age of 65 years, but their model assumed that benefits in screen-detected and non-screen-detected populations are the same and it was unclear whether or not lack of compliance to therapy, if indicated, was modelled. The cost-effectiveness of OACs in non-screen-detected AF has recently been demonstrated. 37

A recent report for the National Screening Committee24 concluded that there was a need to formally assess the cost-effectiveness of a screening programme for AF, and that such a cost-effectiveness analysis should include both costs of detecting undiagnosed cases of AF and also allow for stroke risk in screen-detected AF to be lower than that in clinically diagnosed AF.

Key questions arising from the World Health Organization criteria for the assessment of a screening programme

Many of the WHO criteria are clearly met, although some questions remain. Key inputs to assess the value of a screening programme are:

• the relative stroke risk for asymptomatic compared with symptomatic AF

• the proportion of these asymptomatic cases that would be eligible to benefit from anticoagulation therapy

• the split between AF type (paroxysmal, persistent, permanent) in screen-detected cases and those missed by screening (false negatives)

• the relative sensitivity and specificity of the range of possible screening tests

• the resources required for performing the screening tests and subsequent diagnostic tests

• the risk factor profiles of screen-detected AF cases

• the proportion of people with AF detected by screening who are eligible for anticoagulation who would actually receive it

• the impact of screening on long-term morbidity and mortality outcomes.

Aim

The aim of this project was to assess the cost-effectiveness of different national screening strategies for AF in older adults. To achieve this aim, we conducted a series of reviews and developed an economic model.

Objectives

• To conduct a systematic review of diagnostic test accuracy (DTA) studies to determine:

  • the diagnostic accuracy of screening tests for detecting AF in adults (≥ 18 years) who have not sought medical attention in a primary or community care setting on account of symptoms associated with AF

  • the diagnostic accuracy in systematic opportunistic, targeted and population screening settings.

• To update a previous systematic review49 of screening strategies for AF to answer the following questions.

  • Does systematic screening increase the detection of AF compared with routine practice?

  • What are the characteristics of those identified with AF by screening strategy?

  • Which combination of screening strategy, screening population and test is the most effective at detecting AF compared with routine practice?

  • What are the potential safety issues and adverse events associated with individual screening programmes?

  • How acceptable is the intervention to the target population?

  • What are the costs associated with systematic screening for AF?

• To develop an economic model to compare the cost-effectiveness of different national screening strategies (including no screening) in a primary care setting in England and Wales, with the following objectives:

  • to review previous economic evaluations of screening for AF

  • to review recent literature on the prevalence, disease progression and risk profiles of AF and screening strategies relevant to a UK population-based screening setting

  • to evaluate the cost-effectiveness of different population-based screening strategies in England and Wales.

Objectives

The primary objective of this review was to determine the DTA of screening tests for detecting AF in adults (aged ≥ 18 years) who have not sought medical attention in a primary or community care setting on account of symptoms associated with AF.

The secondary objective of this review was to determine the DTA of index tests in various (i.e. systematic opportunistic, targeted and population) screening settings.

Review question

• Study design. Diagnostic cohort or case–control study.

• Population. People who had not sought medical attention on account of symptoms associated with AF.

• Presentation. Individuals registered in primary care and/or presenting to primary care or a community centre. Individuals may be invited to screening regardless of medical history (this may be done on the basis of age – systematic screening); on presenting to the GP for issues unrelated to AF symptoms (e.g. flu vaccination – systematic opportunistic screening); or based on their medical history/the presence of risk factors that are associated with AF (targeted screening).

• Prior tests. No prior testing for AF was required for the inclusion of individuals.

• Index test. Any non-invasive test that could be utilised in a primary care setting or in the community.

• Purpose. A screening test to identify people with AF who have not sought medical attention on account of symptoms associated with AF.

• Target disorder. AF.

• Reference standard. 12-lead ECG interpreted by a cardiologist.

Search strategy and selection criteria

MEDLINE, PreMEDLINE, EMBASE, The Cochrane Library, the Centre for Reviews and Dissemination (CRD) and the Science Citation Index were searched without language or date restrictions until January 2015. The search strategy included terms for atrial fibrillation and DTA or atrial fibrillation and diagnosis and the names of specific tests [pulse palpation, finger probe, ECG (single-lead, 12-lead, ambulatory, serial, continuous), Holter, cardiac event recorder, modified blood pressure monitor] (see Appendix 1 for the search strategy run in MEDLINE). Grey literature sources were also searched [Google, National Guideline Clearing House, Guidelines International Network (GIN), Current Controlled Trials, WHO International Clinical Trials Registry Platform, ClinicalTrials.gov, UK Clinical Trials Gateway, OpenGrey, NICE website, King’s Fund, Department of Health, National Institute for Health Research (NIHR) UK Clinical Research Network (UKCRN), PROSPERO, European Society of Cardiology, American College of Cardiology, British Heart Foundation, Atrial Fibrillation Association, Arrhythmia Alliance] with search terms including AF and diagnosis. Additionally, the reference lists of identified systematic reviews and eligible studies were hand-searched to identify potentially relevant studies.

Titles and abstracts were screened independently by two reviewers (AM and GO) to identify potentially relevant studies. At title and abstract screening stage, studies were excluded if the target condition was not AF, the study was not a DTA study, the study was performed exclusively in people with diagnosed or treated AF, or if the index test was invasive, not possible in primary care or did not detect arrhythmia. Full texts of the remaining studies were then obtained and assessed for study eligibility using the full set of inclusion and exclusion criteria. This was also performed independently by two reviewers (AM and GO). Conflicts at each stage were resolved through discussion with a third reviewer (PD).

Data extraction

Data extraction was performed by one reviewer (AM) and checked by a second reviewer (GO). Any disagreements were resolved through discussion with a third reviewer (PD).

The following data were extracted:

• authors, publication year, journal

• study design

• characteristics of study participants including age, sex and ethnicity

• study inclusion and exclusion criteria

• setting

• prevalence of comorbidities (e.g. hypertension, diabetes, renal failure, heart failure, heart disease, previous stroke or MI) or risk of stroke CHA₂DS₂-VASc score

• prevalence of AF and prevalence of AF by type (paroxysmal, persistent, permanent)

• method of participant recruitment

• index test, including how and when the test was performed, the frequency of screening and the length of monitoring, and who performed the test and interpreted the test results

• definition of a positive index test result/cut-off value

• reference test, including definition for positive disease

• whether the readers of the index test and reference standard were blind to the results of the other test and other clinical information available to readers

• number of missing or unavailable test results

• numbers of true positives, true negatives, false positives and false negatives.

When studies reported the DTA of multiple thresholds for the same test or the DTA of more than one type of the same class of device, data for all thresholds/devices were extracted.

In studies in which three groups were included (cases with AF, control subjects with another arrhythmia and sinus rhythm/healthy control subjects), DTA information was extracted for AF cases compared with sinus rhythm/healthy control subjects when possible. When this was not possible (i.e. the analyses pooled the two control groups), data were extracted but the study was not included in the meta-analysis.

Study quality

Study quality was assessed using the QUality Assessment of Diagnostic Accuracy Studies – 2 (QUADAS-2) tool. 51 The QUADAS-2 tool consists of four key domains: patient selection, index test, reference standard and flow of patients through the study and timing of index tests, and reference standard (flow and timing). Each domain is assessed in terms of risk of bias, and the first three are assessed in terms of concerns regarding applicability. The tool was tailored to the review by modifying signalling questions, which help to inform judgements, and developing review-specific guidance. The signalling questions and guidance for this review are detailed in Appendix 2.

Statistical methods typically used to gauge risk of reporting bias may be misleading if applied to meta-analyses of test accuracy. 52 Therefore, reporting bias was not assessed statistically using methods such as funnel plots.

Statistical analysis

Evidence synthesis

Formal evidence synthesis was conducted by fitting a hierarchical summary receiver operating characteristic (HSROC) model53 to estimate the relationship between sensitivity and specificity of the index tests. The HSROC model estimates a summary receiver operating characteristic (SROC) curve allowing for correlations between sensitivity and specificity and heterogeneity between studies. The SROC curve depends on the threshold (cut-off point) for the test, the accuracy of the test and the shape of the curve (Figure 1).

FIGURE 1.

Illustration of the SROC curve.

We attempted to fit both symmetrical and asymmetrical HSROC models; however, there were too few data to identify the shape parameter (regardless of whether this was constant, exchangeable or independent across studies) and so results for symmetrical HSROC models only are reported. We adapted the HSROC model to allow an extra level of variability between interpreters in studies with replicates; however, we again found that there were too few data to estimate this extra level of variability. We therefore reported results from a model that does not estimate between-interpreter variability. We allowed for multiple thresholds by treating these as independent observations but with the same study-specific accuracy. We fitted two models, one with and one without classifying the tests according to the interpreter.

The models were fitted using a Bayesian approach computed in OpenBUGS version 3.2.3 rev 101254 (MRC Biostatistics Unit, Cambridge, UK). Model fit was assessed using the posterior mean deviance,55 and models were compared using the deviance information criterion (DIC). 55 Full details of the model fitted, including computer code, are provided in Appendix 3.

The results of the HSROC models were presented as SROC curves with confidence regions and prediction regions for sensitivity and specificity.

Objectives

The objective of this systematic review was to update the Cochrane review of screening strategies for AF. 49 During the course of our review we identified that the Health Information and Quality Authority (HIQA) in Ireland had published a health technology assessment (HTA) of a national screening programme for AF in primary care. 25 The HIQA HTA project25 included a review of the effectiveness of screening, which updated the Cochrane review49 with the results of a literature search covering June 2012 to June 2015. In addition, we were also informed that the Cochrane review itself was being updated (currently undergoing peer review, Patrick Moran, Health and Quality Association, Dublin, Ireland, 2015, personal communication) and, with the assistance of the review authors, we updated this version of the review while expanding the scope to include head-to-head trials and trials that had not confirmed the presence or absence of AF with a 12-lead ECG interpreted by a GP, a specialist or a suitably trained ECG technician.

Review questions

Our review questions were as follows.

• Does systematic screening increase the detection of AF compared with current routine practice?

• What are the characteristics of those identified with AF by screening strategy?

• Which combination of screening strategy, screening population and test is the most effective at detecting AF compared with routine practice?

• What are the potential safety issues and adverse events associated with individual screening programmes?

• How acceptable is the intervention to the target population?

• What are the costs associated with systematic screening for AF?

The protocol for the systematic review was registered with the NIHR international prospective register of scientific reviews (PROSPERO) prior to executing the literature search strategy (registration no. CRD42014013739). 56 In line with the Cochrane review, which we aimed to update, our PICO (population, intervention, comparator, outcome) framework was as follows.

Search strategy and selection criteria

The Cochrane review team confirmed that all head-to-head RCTs of systematic screening would have been identified by their search strategy and would have been examined at full text, although those without a no screening control arm or without AF diagnosis confirmed by a 12-lead or continuous ambulatory ECG interpreted by a GP, specialist or suitably trained ECG technician would then have been excluded. The Cochrane review team kindly provided us with a list of excluded studies that they thought may be relevant to this review. We examined this list to identify studies meeting our inclusion criteria. In addition, we updated the update by searching MEDLINE and PreMEDLINE, EMBASE, The Cochrane Library’s CENTRAL and CINAHL from July 2015 to December 2015.

An information specialist searched health-care databases and other literature sources to identify published and unpublished literature on human subjects, in any language. Our searches were developed for MEDLINE and then adapted for other literature sources when necessary (see Appendix 4 for details). Reference and citation tracking were undertaken to identify further relevant studies. When necessary, we contacted lead authors for more information on published and unpublished studies that might be relevant. We also sought information on studies in progress.

After removal of duplicates, articles were screened in parallel by two reviewers (GO, AM) applying our eligibility criteria and using a two-stage sifting approach to review article title and abstract and then full text. As the Cochrane review and the ongoing update review excluded comparative studies that did not have a no screening arm (control), two reviewers (GO, AM) reassessed the list of excluded studies provided by the Cochrane review authors for inclusion. In all cases, disagreements were discussed between the two reviewers and, if not resolved, resolution was sought through the involvement of a third reviewer (PD).

Data extraction

We developed data extraction forms and piloted them on two studies. Data were extracted from included studies according to the PICO framework using the piloted extraction forms. Data extraction was conducted by one reviewer and then reviewed by a second reviewer. Disagreements were discussed between the two reviewers and, if not resolved, resolution was sought through the involvement of a third reviewer. We extracted data on the following:

• study details (identifier, study design, location, year, length of follow-up, study funder)

• participant details (number of participants, age, sex)

• intervention details (population screening type, screening test, method of AF diagnosis, timing)

• comparator details

• details relevant to the risk-of-bias assessment

• effect modifiers (demography of the study population, study settings).

Assessment of risk of bias

Risk-of-bias assessments were conducted by one reviewer and then reviewed by a second reviewer. Disagreements were discussed between the two reviewers and, if not resolved, resolution was sought through involvement of a third reviewer. Risk of bias was assessed using the Cochrane tool for assessing risk of bias in randomised trials. 58 This tool assigns the judgement of low, high or unclear risk of bias for each of the following sources of bias: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data and selective reporting. Additional items for assessing the risk of bias in cluster randomised trials were derived from the Cochrane-suggested risk-of-bias assessment tool for Effective Practice and Organisation of Care (EPOC) reviews. 59

Abstracts that were assessed to meet the study eligibility criteria for inclusion were subjected to data extraction only and not to assessment for risk of bias because of a lack of information to enable judgement on some of the criteria.

We planned to use funnel plots to assess reporting bias; however, insufficient numbers of studies were identified for this to be possible.

Data analysis

We have tabulated study-specific characteristics, outcome measures and risk-of-bias assessments, and we provide a narrative overview of the results from the included studies.

We report results for binary outcomes (e.g. proportion of new AF diagnoses) only for studies in which there is a clear denominator. We report results for the proportion of new AF diagnoses among those without a previous diagnosis of AF and summarise relative effects between study arms using the odds ratio (OR). The number needed to screen (NNS) to detect an additional newly diagnosed AF case was calculated as the reciprocal of the risk difference (difference in proportion of new cases detected between study arms). We computed the prevalence of previously diagnosed AF prior to screening by pooling across arms within a study. We calculated the prevalence of AF diagnoses before and after screening to compute the percentage change in prevalence as a result of screening. We examined the acceptability of systematic screening programmes within the screened population in three ways: from the level of uptake achieved, feedback elicited from the participants and health professionals involved and a description of any direct costs associated with screening that were borne by the person to whom the screening programme was offered. Patient characteristics of those diagnosed with AF (age, sex, clinical history, CHA₂DS₂-VASc score, AF type) are described when reported. We report any adverse events associated with screening reported in the studies. We also report information on the incremental cost to the NHS per additional newly diagnosed AF case identified for a given screening strategy compared with routine care when this was reported in RCTs conducted in the UK. We planned the following subgroup analyses when enough data were available: aged ≥ 65 years compared with 65–75 years compared with > 75 years; men compared with women; different ethnic and socioeconomic groups; and community compared with specialist settings. We also planned a sensitivity analysis excluding trials with a high risk of bias if there were enough data.

Meta-analysis was conducted when feasible (enough data on outcomes defined in the same way) to obtain pooled ORs with 95% CIs. We did not anticipate there being a sufficient number of studies to estimate random-effects meta-analysis models and so reported results from fixed-effects meta-analysis only.

Study design

There were 10 studies with a single set of inclusion criteria (cohort or one-gate studies),19,60,63–69 three studies with two sets of inclusion criteria (case–control or two-gate studies)70–72 and two studies of unclear design. 73,74 One of the case–control studies grouped participants into those with known sustained AF, those with other non-AF arrhythmias and those in sinus rhythm. 71

Population and setting

Only four studies were performed in primary care. 19,67,70,72 The other studies were performed in outpatient settings (hypertension outpatients,63 presurgery64), secondary care or higher settings60,65,66,68,69 or a mixture of outpatient and secondary care or higher settings [outpatient clinics including cardiology outpatients and the preadmission clinic (ECGs carried out prior to surgery in patients aged > 40 years);74 outpatient and inpatient settings;60 outpatient hypertension clinic, patients admitted to University Department of Medicine wards and healthy volunteers71]. In one study the setting was unclear (a computerised electrocardiographic system). 73

A secondary objective of this review was to determine the DTA of index tests in systematic (opportunistic, targeted and population) screening settings. To do this we had to assess whether the populations studied represented the populations who would attend if a population-based systematic screening programme was in place. In practice, this was difficult to achieve as it was unclear how the study populations were selected and recruited. We were able to categorise only four studies, three of which represented population systematic screening (one recruited individuals aged ≥ 75 years from primary care,67 one case–control study recruited cases and controls aged ≥ 65 years from primary care70 and one case–control study recruited a convenience sample of people from primary care72) and one that represented targeted screening (patients referred to a hypertension clinic63). However, it should be noted that case–control studies, by design, include individuals who are clearly cases or controls and who do not tend to fully reflect a population-based screening population. We were unable to classify the population in one study in primary care because it was part of a RCT that compared systematic population and systematic opportunistic screening, with participants from both arms feeding into the DTA study. 19

In the included cohort studies, the prevalence of AF varied between 0.93%64 and 32.93%. 68

Age was an inclusion criterion in seven studies. Participants had to be ≥ 18 years in two studies,63,66 ≥ 35 years in one study,71 ≥ 60 years in one study,69 ≥ 65 years in two studies19,70 and ≥ 75 years in one study. 67

Seven of the studies excluded patients with a pacemaker and/or an implanted defibrillator (in some cases only if they were in active pacing mode). 63,65–67,71,72,74

Only seven studies reported any characteristics of the included cohorts19,63,64,66,67,71,72 and, of these, only three reported on comorbidities and/or treatments received. 67,71,72

Index test

The 15 studies contained 56 eligible comparisons of index test with the reference standard. Not all comparisons in the included studies were included because of ineligible cohorts68,73,74 or because the results were not reported in a usable format (e.g. pooling multiple readings of the same scans19 or data being presented per reading rather than per person71).

The index test used was classified into one of eight categories to facilitate the analyses:

• pulse palpation

• photoplethysmography

• modified blood pressure monitor

• single-lead ECG

• > 1- and < 12-lead ECG

• 12-lead ECG

• ambulatory monitoring

• two-stage screening.

Two studies investigated pulse palpation. 19,70 In both studies, the pulse palpation was performed by a nurse, but in one study the accuracy of pulse palpation by nurses with different backgrounds (a nurse who had a background in both community and accident and emergency nursing and a practice nurse) was compared. 70 Pulse palpation may have been investigated as an index test in one other study, but no results for this index test were presented. 67

Photoplethysmography was investigated in one study. 69 This study used a portable device (AFS instrument), which is a finger-probe instrument, as used in pulse oximetry. The resulting pulse waveform was interpreted using fast Fourier transform analysis and the results were presented for three different cut-off values. The data on photoplethysmography for the 0.20 cut-off value was obtained directly from raw data kindly provided to us by Malcolm Lewis (Wales Postgraduate Deanery, Cardiff, 2015, personal communication) and was used to estimate the data for the 0.25 and 0.30 cut-off values (using the number of false negatives and false positives reported in the publication). However, we had confirmation from the authors that what they have termed specificity in their paper is in fact (1 – the proportion of false positives in the whole population). It was therefore possible to obtain the actual 2 × 2 data only for the 0.20 cut-off value and so only this cut-off value was included in the meta-analysis.

A modified blood pressure monitor was used as the index test in three studies. 63,67,71 These devices have been modified to detect pulse irregularity during blood pressure measurement. In all three studies, AF was detected automatically/by an algorithm. In one study, the WatchBP device was used. 67 One study compared two different blood pressure monitors: Microlife BP A200 Plus and OMRON M6. 63 One study used the Microlife BPA100 Plus and different cut-off values for the number of positive results needed to screen positive (readings were taken in triplicate and the cut-off values investigated were positive on the first reading, positive on either of the first two readings, positive on any of the three readings, positive on two of the three readings). 71 This study recruited 27 people with AF, 23 people with non-AF arrhythmias and 23 people with sinus rhythm. 71 We specified that, when possible, when studies were made up of three groups of participants (people with AF, people with another arrhythmia, people in sinus rhythm) we would extract results for people with AF and people in sinus rhythm (because the people in sinus rhythm are more similar to our target screening population). However, in this study the authors pooled the people with a non-AF arrhythmia and the people in sinus rhythm as non-AF cases. This may have impacted on the observed specificity of the test. Because of this, although all results are reported in Figure 5, we excluded the data derived from this study71 in the meta-analysis.

Single-lead ECGs were used in five studies. 66,67,70,72,74 Various different devices were used. The Omron (model HCG-801), which requires placing one electrode on the individual’s bare chest while the index finger of the right hand holds the device, was included in two studies,66,67 interpreted by a cardiologist. The Merlin ECG event recorder, which resembles a watch that the individual covers with the palm of his or her right hand to take a reading, was included in one study,67 interpreted by a cardiologist. Bipolar ECG using just a single bipolar lead from a 12-lead ECG machine was included in a single study,70 interpreted by nurses with different backgrounds (training/experiences with ECGs). The AliveCor device, which takes a single-lead ECG recording when fingers from each hand are placed on electrodes on the back of an iPhone (Apple Inc., Cupertino, CA, USA) case, was used in one study,74 with interpretation by an optimised algorithm. The MyDiagnostick device, in the form of a rod with a metal handle on both ends containing electrodes, was included in a single study72 and was also analysed automatically.

Greater than 1- and < 12-lead ECGs were used in two studies. 64,65 In one study 12-lead ECGs were reconstructed from V2 and V5 leads or V1 and V4 leads and interpreted by an algorithm. 65 In one study three-lead tele-ECGs were interpreted by two different cardiologists. 64

Twelve-lead ECGs interpreted by different diagnostic algorithms were compared with 12-lead ECGs interpreted by a cardiologist in six studies, with more than one algorithm tested in some cases. 19,60,65,68,73 One of these studies also investigated the effect of the ECG machine by comparing the diagnostic accuracy of an algorithm for ECGs recorded using two different machines (GE Marquette system MAC5000 and Bionet EKG3000). 60 Twelve-lead ECGs interpreted by nurses with different backgrounds (experience/training with ECGs) and a GP were compared with 12-lead ECGs interpreted by a cardiologist in one study. 70 We excluded one arm of a study investigating the accuracy of 12-lead tele-ECGs because of the index test being a variant of the reference standard (sensitivity and specificity were 100%). 64

The results of one of the studies that reported the DTA of 12-lead ECGs interpreted by a diagnostic algorithm were reported in two publications. 19,36 In these publications, people could have a positive, negative or uncertain result on this test. The number of people obtaining an uncertain result was slightly different in the two publications. The numbers extracted and reported in this review are from the later publication. 36 In addition, in this publication individuals obtaining an uncertain result were considered to have obtained a negative test result to calculate the sensitivity and specificity of the test. In practice, those with an uncertain test result may be referred for further testing (effectively treated as test positives). If the uncertain results were considered to be positive results it would increase the sensitivity but decrease the specificity.

None of the included studies used any form of ambulatory monitoring.

Two studies investigated some form of two-stage screening. One study modelled a two-stage screening process in which individuals who had a positive result with the modified blood pressure monitor (WatchBP) went on to have a single-lead ECG with autoanalysis (Omron). 67 One study looked at six different combinations of pulse palpation by different nurses (nurse A, who had a background in both community and accident and emergency nursing and who had experience of taking and interpreting ECGs; and nurse B, a practice nurse with no additional ECG training) followed by a bipolar or 12-lead ECG interpreted by either a nurse or a GP for those who screened positive:70

1. pulse palpation assessed by a nurse (A) followed by a bipolar ECG interpreted by a nurse (A) for those who screened positive

2. pulse palpation assessed by a nurse (A) followed by a bipolar ECG interpreted by a GP for those who screened positive

3. pulse palpation assessed by a nurse (A) followed by a 12-lead ECG interpreted by a nurse (A) for those who screened positive

4. pulse palpation assessed by a nurse (A) followed by a 12-lead ECG interpreted by a GP for those who screened positive

5. pulse palpation assessed by a nurse (B) followed by a bipolar ECG interpreted by a GP for those who screened positive

6. pulse palpation assessed by a nurse (B) followed by a 12-lead ECG interpreted by a GP for those who screened positive.

Reference standard

The reference standard in all studies was a 12-lead ECG interpreted by a cardiologist, apart from one study that used an ECG classified by a clinician and validated by a researcher (a cardiac electrophysiologist),68 which was judged to be equivalent to interpretation by a cardiologist.

Risk of bias: patient selection

Four studies were judged to be at low risk of bias for patient selection. Three of the cohort studies reported that consecutive or randomly selected individuals were included62,65,66 and, although no details were provided about the method of inclusion, it seemed possible that it was consecutive in one other study. 69

The three studies with two sets of inclusion criteria (case–control or two-gate studies)70–72 and the two studies of unclear study design73,74 were judged to be at high risk of bias.

The other studies were judged to be at an unclear risk of bias because of concerns over or lack of information on the method of enrolment and/or concerns over or lack of information on exclusion criteria. 19,60,64,67,68

Risk of bias: index test

The majority of the studies were scored as being at low risk of bias on the index test domain. However, one study was judged to have an unclear risk of bias for some of the index tests investigated because it seemed possible that the test interpreters for some of the index tests were aware of the results of another of the index tests. 67 Three studies were judged to be at high risk of bias on the index test domain, mainly because the cut-off point used to screen positives was not prespecified. 68,69,71

Risk of bias: reference standard

No studies were judged to be at high risk of bias for this domain. Studies were assessed as being at unclear risk of bias because they did not explicitly report that the interpreters of the reference standard were unaware of the results of the index test60,65,71 or because the index test was interpreted by a clinician and validated by a researcher (a cardiac electrophysiologist). 68

Risk of bias: flow and timing

All studies were judged to be at low risk of bias for the domain of flow and timing.

Applicability: patient selection

Only those studies performed in primary care were judged to have a low level of concern with regard to applicability. 19,67,70,72

Applicability: index test

Studies in which the index test was interpreted by a cardiologist were judged to be less applicable (high level of concern regarding applicability) than studies interpreted in primary care/by an algorithm. 64,66,67

Applicability: reference standard

All of the studies were judged to have a low level of concern regarding applicability apart from one, because the index test was interpreted by a clinician and validated by a researcher (a cardiac electrophysiologist). 68

Study year

There were too few data to perform a subgroup analysis for recent studies. Only one study reported before 2000 (in 199273). The study by Slocum et al. 73 was also of unclear design and setting, and it was unclear which screening approach it reflected. Therefore, this study was excluded from the various subgroup analyses reported in the following sections.

Study design

Study design was classified into cohort, case–control and unclear. Ten out of 15 studies (34 observations) were cohort studies, and three out of 15 studies (20 observations) were case–control studies. We conducted analyses including cohort studies only (our preferred study design), but found that the results were robust to restricting to cohort studies only. It should be noted that single-lead ECG – nurse, single-lead ECG – GP, 12-lead ECG – nurse and 12-lead ECG – GP were studied only in non-cohort studies. Tests studied exclusively in cohort studies included photoplethysmography and modified blood pressure monitors.

Screening approach

Three out of 15 studies (29 observations) most closely reflected population-based systematic screening, one out of 15 studies (two observations) most closely reflected targeted screening and the remainder were unclear. We conducted analyses restricting to studies most closely reflected by population-based systematic or targeted screening, which are those most relevant to our decision question. The results were robust to restricting to these populations; however, some tests, including photoplethysmography, were not studied in systematic or targeted screening studies and so the results for these tests could be less applicable to our decision question.

Setting

Four out of 15 studies (31 observations) were carried out in a primary care setting, 10 out of 15 studies (24 observations) were carried out in a non-primary care setting (outpatients or secondary care and above) and in one study (one observation) the setting was unclear. We present the results of subgroup analysis for the primary care setting only as this is most relevant to our decision question. No differences were found from the base-case results but, again, it should be noted that photoplethysmography was studied only in the non-primary care setting and so the results for this test may be less applicable to screening in a primary care setting.

Mean age/sex

The mean age of individuals was reported in only seven out of 15 studies (27 observations). Mean age ranged from 61.4 to 79.7 years, with the majority of studies in the range 65–75 years. Because of the high level of missing data on age and the similarity in mean age across studies it was not possible to perform subgroup analyses based on age. The proportion of male participants was reported in the same subset of studies that reported mean age, varying from 47% to 66%. It was therefore also not possible to report the results according to sex.

Prevalence of atrial fibrillation and atrial fibrillation subtypes

It was possible to estimate the prevalence of AF in the included studies and this ranged from 0.9% to 53%, with no distinct groupings. We included prevalence as a continuous covariate in a meta-regression, adding a regression coefficient for the effect of AF prevalence on false positive rate and true positive rate respectively. We excluded case–control studies from the regression as they do not provide meaningful estimates of prevalence. No data were reported on type of AF in any of the included studies and we were therefore unable to report the results according to AF type.

Prevalence of comorbidities/stroke risk score, length of monitoring for ambulatory tests and cut-off values for tests with quantitative reading

Only two studies reported information on comorbidities67,72 and only one reported information on CHA₂DS₂-VASc score. 72 It was therefore not possible to report the results according to comorbidities/stroke risk score. No data were reported on length of monitoring for ambulatory tests and so no subgroup analysis was possible. The only study that reported results for different cut-off values was that by Lewis et al. 69 but, as previously noted, the results were reliable only for a single cut-off value in this study (the one with the best sensitivity). It was therefore not possible to perform a subgroup analysis on the cut-off values used.

QUADAS-2 risk of bias domains

We present the results restricting the analysis to low-risk studies only for each domain separately:

• risk of bias according to patient selection

• risk of bias according to index test

• risk of bias according to reference test

• applicability according to patient selection

• applicability according to index test

• applicability according to reference test.

All studies were rated as being at low risk of bias on the flow/timing of tests domain and so no sensitivity analysis was necessary for this domain.

We found that the results were robust to all of these sensitivity analyses; however, it should be noted that there was a lack of data to fully explore these. In particular, for most tests there was at least one domain in which no observations were graded as being at low risk of bias/applicability. The exception was modified blood pressure monitors, for which most of the observations were at low risk of bias/applicability. Photoplethysmography was the only test to have no observations at low risk of bias for the index test, suggesting that we should be cautious when interpreting the results for this test.

Summary of results

This review of 15 studies of screening tests for detecting AF found that most tests had a sensitivity (probability of detecting AF in patients with AF) in excess of 0.9. In support of the view that screening could be carried out in primary care, 12-lead ECG interpreted by a GP had a sensitivity of 1 (95% CrI 1 to 1) and also a high specificity of 0.97 (95% CrI 0.84 to 1). Photoplethysmography also had a sensitivity of 1 (95% CrI 1 to 1), but with a lower specificity of 0.87 (95% CrI 0.52 to 0.99). Specificity was in general lower than sensitivity for all of the tests and was lowest for pulse palpation by a nurse (specificity 0.79), 12-lead ECG interpreted by a nurse (specificity 0.84) and photoplethysmography (specificity 0.87). Tests with the highest DOR were the 12-lead ECG (regardless of interpreter), between 1- and 12-lead ECG (automatic or cardiologist interpretation), two-stage tests and single-lead ECG interpreted by a GP, with all of these tests having similar DORs. However, the DOR gives equal weight to true positives and true negatives, whereas for a screening programme test sensitivity takes precedence over specificity (if the test is not too burdensome), so that AF cases are not missed. False positives should be picked up subsequently by a confirmatory diagnostic test. We are therefore interested in finding a test that primarily has good sensitivity; specificity is a secondary consideration to avoid too many individuals being referred for a confirmatory diagnostic test.

In general, for a given interpreter, the results for single-lead ECGs were less accurate and more variable than those for ECGs with more than one lead. Nurse interpretation of single-lead ECGs performed similarly to single-lead ECGs with other interpretation methods, but nurse interpretation of 12-lead ECGs did not perform as well as other interpretation methods. Automatic interpretation did not have a consistent impact on test accuracy, with automatic interpretation of single-lead ECGs having a high sensitivity but variable specificity. In contrast, automatic interpretation of ECGs with more leads had good specificity but variable sensitivity.

The different two-stage screening strategies all had very high specificity, but sensitivity was high only when a 12-lead ECG was used as the second-stage test. This supports the need for a 12-lead ECG as a diagnostic test. Interestingly, a second-stage 12-lead ECG test interpreted by a (trained) nurse performed similarly to that interpreted by a GP.

Strengths and limitations

Our review followed a published protocol, registered on PROSPERO. 56 Inclusion was restricted to studies that best represented a population-based screening programme conducted in primary care. A comprehensive search strategy was used and the risk of bias and applicability of studies was assessed. The effect of interpreter as well as the screening test utilised were both investigated. Preplanned subgroup and sensitivity analyses were conducted when data allowed.

Analyses were limited by the small number of studies that met the inclusion criteria and hence the lack of statistical power to detect meaningful differences. For example, not all combinations of screening test and interpreter were included in the studies and, when present, few observations were reported. For this reason tests were grouped in the statistical modelling. In particular, > 1- and < 12-lead ECGs included 3-lead ECGs as well as reconstructed ECGs resulting from different sets of leads. Similarly, two-stage screening strategies encompassed varying combinations of tests in the two stages.

There was substantial heterogeneity in the test/interpreter combinations when there were repeated observations, and it is likely that the same heterogeneity would be seen for the other tests if multiple observations were available. Furthermore, we would expect more heterogeneity in test accuracy in practice than is seen in research studies.

None of the included studies was assessed to be at low risk of bias across all domains and to have a low level of concern regarding applicability. In particular, only four studies were conducted in primary care. Although we had preplanned a wide range of subgroup analyses and sensitivity analyses, there were enough data available to perform only some of these. The way that the data were reported was in some cases challenging. Some studies were designed so that the same ECGs were read by different individuals, whereas in other studies different individuals read different sets of ECGs. Some studies presented the results for patients with AF, individuals in sinus rhythm and patients with non-AF arrhythmias, but some combined the sinus rhythm and non-AF groups when reporting results, which will underestimate test accuracy as applied to the general population in which the proportion of other arrhythmias is much lower. Although photoplethysmography performed well, only a single study69 used this method, and this study was able to provide data only for a single cut-off point, which was chosen retrospectively for having the best sensitivity. The index test in this study was rated as being at high risk of bias and the applicability of patient selection was unclear. We are therefore cautious in interpreting the results from this screening test and would ideally want the results to be replicated in a new cohort of individuals.

Findings in the context of previous research

European Society of Cardiology guidance recommends pulse palpation as the first step for AF screening. 32,75 We found pulse palpation to have the lowest DOR, driven by its low specificity.

Guidance from NICE (medical technologies guidance MTG13)76 recommends the use of modified blood pressure monitors for the detection of AF in individuals undergoing diagnostic testing or monitoring for hypertension. We found that modified blood pressure monitors had high sensitivity, although specificity was not as high as with other tests.

During the course of this review, two systematic reviews were published. One reported the accuracy of methods for detecting an irregular pulse and suspected AF;77 the other reported the accuracy of methods for diagnosing AF using 12-lead ECG. 78 In addition, a systematic review of the DTA of automated blood pressure monitors for opportunistic AF detection,79 and a review and meta-analysis of the DTA of the algorithm for AF detection during automated blood pressure measurement80 have been published. These reviews are summarised in Table 7.

The SAFE study19,84

This study compared systematic population screening, systematic opportunistic screening and no screening. The study employed a two-stage randomisation process. The first step was a balanced cluster randomisation (stratified based on Townsend score quartiles and practice size) of 50 practices in the UK to either the intervention (screening) arm or the control (no screening) arm. This was then followed by a random assignment of individuals aged ≥ 65 years in the intervention centres to either systematic opportunistic or systematic population screening. Individuals aged ≥ 65 years were also identified in the control arm.

Individuals randomised to systematic opportunistic screening had a flag placed in their notes to encourage practice staff to take a pulse recording during routine consultation. If their pulse was found to be irregular individuals were invited to attend a screening clinic. Individuals in the systematic population screening arm were all invited by post to attend a screening clinic. The screening clinic was the same for both screening arms (i.e. they differ only in how individuals were selected for invitation to the screening clinic). At the screening clinic consent was obtained, baseline characteristics were recorded and a 12-lead ECG was performed, which was interpreted by two blinded cardiologists, with interpretation by a third blinded cardiologist in the case of disagreement. An acceptability questionnaire was also delivered to patients.

This study also identified ‘high-risk patients’ with cardiac failure, hypertension, rheumatic heart disease, previous MI, angina, diabetes mellitus, hyperthyroidism, previous stroke or previous TIA (using practice disease registers) among the individuals randomised to systematic population screening. If these patients were selectively invited to screening, this would represent a targeted screening programme. The results of targeted screening (using the whole population screening as the denominator) were compared with the results of no screening.

Patients with previously diagnosed AF were included in the study, but the number of such patients was reported so that the number of newly diagnosed cases could be ascertained. To identify previously diagnosed cases of AF, patient notes were searched for ‘G573 AF/flutter’, ‘327 ECG supraventricular arrhythmia’, ‘181 palpitations’, ‘digoxin’, ‘amiodarone’, ‘verapamil’, ‘sotalol’, ‘metoprolol’, ‘warfarin’ and ‘aspirin’. Case notes of patients identified were reviewed and a diagnosis of AF was accepted if there were hospital letters referring to AF or confirmatory ECGs within the previous 5 years.

The primary study outcome was the number of newly diagnosed cases of AF during the 12 months of follow-up. Newly diagnosed cases were identified using the same computer searches as performed at baseline.

The STROKESTOP study26,40

The STROKESTOP study aims to test the hypothesis that screening 75- and 76-year-olds for AF will reduce stroke incidence cost-effectively. All 75- and 76-year-olds in two regions of Sweden were randomised to either the systematic population screening arm or the no screening arm.

Participants who were randomised to the screening arm were sent an invitation by post. At the screening centre, participants were asked about their medical history and an index ECG (using a hand-held single-lead ECG) was taken. Participants without a history of AF who were in sinus rhythm were given a hand-held single-lead ECG and instructed to use it twice daily and whenever palpitations were felt over a period of 2 weeks. The ECG transmitted the results to a database where they were interpreted by a nurse. All abnormal ECGs were referred to the investigating cardiologist. When results were unclear, referral for interpretation by a consensus group was used.

An interim publication has reported the AF detection rates in individuals in the screening arm who presented at the screening clinic. 40 Both the number of people with known AF and the number of people with newly detected AF were reported. Results for the control arm will be obtained from a National Patient Registry (Emma Svennberg, Danderyds Sjukhus AB, Stockholm, Sweden, 2015, personal communication) but are not available currently.

The primary outcome of this study is stroke prevention resulting from early identification of AF through screening, and subsequent treatment, with 5 years’ follow-up. However, the study is still ongoing and not due to report until 2019. Because the results are currently available only for the screening arm, this study did not contribute data to our statistical analysis.

The Morgan and Mant82 study

This study compared systematic opportunistic screening with systematic population screening. Four general practices in the UK were chosen to represent each quartile of the small area standardised mortality ratio of the geographical area served and approximately 750 individuals aged ≥ 65 years from each practice were randomly selected and then randomised to either the systematic opportunistic or the systematic screening arm.

In the systematic opportunistic screening arm a flag was inserted into the patients’ notes so that if they consulted with their GP or nurse for any reason during the duration of the study their pulse was taken. If their pulse was found to be irregular the nurse or GP could choose to request a confirmatory ECG. In the reported results from this study it was assumed that there was a clinical diagnosis of AF if the pulse was irregular but a confirmatory ECG was not requested. In the systematic screening arm, all individuals were invited by post to a screening appointment with a nurse who took a radial pulse reading for at least 20 seconds to record evidence of irregularity. A two-lead ECG reading interpreted by a GP was used for validation. The primary study outcome was the number of AF cases after 6 months’ follow-up, but the study also reports the number of cases with no previous evidence of AF in their medical records. However, the number of people with and without AF in their medical records at baseline was not reported.

The two arms differed in the validation of AF cases: in the systematic screening arm, a two-lead ECG interpreted by a GP was used for the validation of new cases of AF, whereas in the systematic opportunistic screening arm ECG validation took place at the discretion of the nurse or GP taking the pulse, with some cases identified by clinical diagnosis only and others based on an ECG (number of leads and interpreter unspecified). This difference in validation could have influenced the reported results.

The DOFA-AP study81

This study compares systematic opportunistic screening with a control group (termed ‘usual care’). Details of the study are currently only available in the protocol81 and in an unpublished abstract shared with us by the ongoing Cochrane update review authors (P Moran, Health Information and Quality Authority, Ireland, 4 November 2015, personal communication). Eligibility criteria for the study were participant aged ≥ 65 years and attending their health centre in Spain for any reason without a previous diagnosis of AF.

In the systematic opportunistic screening arm, it is reported that individuals underwent ‘active screening for AF, regardless of the reason of their visit’ (p. 4). 81 In the control arm an individual’s pulse was measured if he or she presented with symptoms of sequelae associated with AF. AF was confirmed using an ECG, but the ECG type and the information regarding the interpreter were not provided.

However, we have concerns about the interpretation of the results of the study. It is possible that to be eligible for inclusion in the control group individuals had to be symptomatic, as the protocol reports that individuals presenting with AF symptoms (dyspnoea, chest pain palpitations or dizziness) or AF-induced sequelae (stroke, TIAs, peripheral embolism, heart failure or other associated health disorders) were eligible. Another statement in the protocol that supports this suggestion is ‘since the probability of finding patients meeting the eligibility criteria is much lower for the [control group (usual care arm)] than for the [intervention arm], stratified randomization will be used to assign at least 200 primary care health professionals to the two groups, in the ratio of a 2 : 3 ratio’ (p. 4). 81 This suggests that the study may in fact be better described as a comparison of targeted testing of symptomatic individuals (control arm) with systematic opportunistic screening (intervention arm).

From the reported results in the abstract, the denominators are also not clear from the information available.

The study result is also unintuitive (the authors reported more AF cases in the control arm than in the screening arm), although this could be explained by differential inclusion criteria for each arm. Because of the lack of information about the denominator for the control arm (both the actual number and what it represents, i.e. the number randomised or the number randomised who presented with symptoms or sequelae), the results from this study were excluded from our analysis.

The EARLY study83

This study compared targeted screening for AF with usual care (no systematic screening). Individuals in a primary health-care centre in Spain were selected from electronic records on the basis of not having a diagnosis of AF but having at least one of the main risk factors for AF. Individuals were then randomised to either an early detection programme or the control arm. Individuals in the early detection programme arm were invited by telephone to participate in the study. The early detection programme involved visits every 6 months over a 2-year period, at which an ECG, a physical examination and a medical history were taken/performed. On the first visit, advice on pulse taking was also given. The primary end point of the study was the proportion of individuals with newly diagnosed AF at 6 months. A secondary outcome was the number of individuals diagnosed with AF during the 2-year follow-up. Cases of AF were determined from the medical records system.

Many individuals were excluded from each arm for different reasons. Of the 2000 people randomised to screening, 616 were not found, 425 declined to participate, 78 were no longer assigned to the health centres, three were deceased, 153 fulfilled at least one exclusion criterion (were disabled, had a pacemaker, did not really have a risk factor, had AF at baseline) and 262 were not contacted. Of the 2000 people randomised to the control group, 42 were no longer assigned to the health centres, six were deceased, 38 fulfilled at least one exclusion criterion and 1449 were not contacted.

Although the study reported that an intention-to-treat analysis was performed, it considered only the individuals included in the study. In the intervention group, individuals could decline to participate, but in the control group they could not. The study actually compared the incidence of AF in all people eligible for screening but who received usual care with the incidence of AF in a self-selecting subgroup of people who both were eligible and attended screening.

On the basis of these methodological limitations, we felt that it was inappropriate to include the results from this study in our statistical analysis.

Newly diagnosed cases of atrial fibrillation among individuals without atrial fibrillation at baseline

Of the five RCTs included in this review, only one (the SAFE study19,84) provided usable data on the relative effects of screening on the proportion of newly diagnosed cases of AF among individuals without AF at baseline. The results of this study are presented in Table 11.

This SAFE study examined systematic population screening and systematic opportunistic screening compared with no systematic screening (usual practice). Both systematic population screening and systematic opportunistic screening increased the odds of detecting new cases of AF among individuals without known AF at baseline compared with usual practice (systematic population screening OR 1.57, 95% CI 1.09 to 2.26; p = 0.017; systematic opportunistic screening OR 1.59, 95% CI 1.10 to 2.29; p = 0.013). The NNS for one screened individual with newly diagnosed AF to be identified by 12 months was 173 individuals for systematic population screening and 167 individuals for systematic opportunistic screening.

The study also modelled a targeted screening strategy, considering what would happen if only individuals at high risk in the systematic population screening arm were invited for screening. Targeted screening of high-risk individuals had a similar odds of detecting AF as usual practice (OR 1.12, 95% CI 0.75 to 1.66; p = 0.58), and the NNS was 833 individuals. These numbers were calculated from the number of people identified through screening who were at high risk plus the number of people in the systematic screening arm who were identified as having AF outside the screening programme. As this was a modelled outcome, it is possible that other cases of AF in low-risk individuals who were detected by the screening programme would have been identified outside the screening programme if a systematic population screening programme had not been in place and therefore the odds of detecting AF in a targeted screening programme could be slightly higher.

Results from Morgan and Mant,82 who compared systematic opportunistic screening with systematic population screening, could not be used because the number of people with AF at baseline was not reported. However, if we assume that the number of people with a previous diagnosis of AF was balanced between the groups we can assess the relative merits of systematic opportunistic screening compared with systematic population screening. Of the 1502 people randomised to systematic opportunistic screening there were seven newly identified cases of AF over the 6 months of follow-up, whereas of the 1499 people randomised to systematic population screening there were 12 new cases of AF (OR systematic population vs. systematic opportunistic 1.72, 95% CI 0.68 to 4.39; p = 0.25).

Although the SAFE study provides strong evidence that screening identifies more cases of AF than usual care, neither the SAFE study19,84 nor the study by Morgan and Mant82 provides strong evidence that one screening approach is superior to another.

Prevalence of atrial fibrillation at baseline

The prevalence of AF in the absence of screening could be calculated from only one RCT (the SAFE study19,84). Averaging across the study arms, the prevalence of AF before screening was 7.22% (95% CI 6.80% to 7.63%).

Prevalence of diagnosed atrial fibrillation at the end of screening and the change in prevalence of diagnosed atrial fibrillation

The change in prevalence of diagnosed AF after screening could be calculated only from the one RCT from which the baseline prevalence of AF could be calculated (the SAFE study). 19 The prevalence of diagnosed AF increased after the trial in all arms, including the control arm (Table 12), reflecting incident AF over the trial period. Compared with the control arm, the prevalence of diagnosed AF changed by approximately an additional 50% in the systematic opportunistic and systematic population screening arms, but targeted screening saw a similar change in diagnosed prevalence to that in the control arm.

Acceptability of systematic population screening programmes

In screening programmes that sent invitations to a defined population, one measure of the acceptability of screening is the uptake of screening among those invited. Table 13 shows the uptake in studies that investigated systematic population and systematic targeted screening (number who attended screening divided by the number sent invitations). Studies with methodological or reporting limitations that precluded inclusion in other parts of this review have been included here. The fact that patients with a previous diagnosis of AF were not excluded from the SAFE,19,84 STROKESTOP26,40 and Morgan and Mant82 studies should be considered when interpreting these results. With the exception of the study by Morgan and Mant,82 uptake of screening was fairly consistent across the studies, with approximately 50% of those invited attending for screening. In Chapter 5 (see Uptake of systematic population screening) we report uptake rates for single-arm studies as well as RCTs and combine the results in a meta-analysis.

It is more difficult to judge the acceptability of systematic opportunistic screening, as systematic opportunistic screening relies on the person visiting the GP/screening centre, the nurse/GP offering the screening test and the individual consenting to it. In the SAFE study, 69.2% (3278/4738) of people randomised to systematic opportunistic screening who had their notes flagged had their pulse palpated. 19,84 Of the 361 people found to have an irregular pulse, 238 (65.9%) attended the screening clinic, where a 12-lead ECG was performed, although it should be noted that, of the 123 individuals who had an irregular pulse who did not attend for an ECG, 56 (45.5%) had already been diagnosed with AF. 19,84 The SAFE study also reports that there was variation at the practice level in both the proportion of individuals who had a pulse taken (from 33.5% to 93.0%) and the uptake of ECGs when the pulse was found to be irregular (from 14.3% to 87.5%).

In the study by Morgan and Mant,82 30.5% (439/1437) of people randomised to systematic opportunistic screening who had their notes flagged had their pulse palpated. In this study, the number of individuals who had their pulse palpated also varied by practice, from 8.0% in one practice to 52.1% in another. In both studies it is unclear how many eligible individuals visited their GP.

Only the SAFE study19,84 reported the reasons why individuals refused to participate in systematic screening. Of the 904 individuals who replied to decline screening, 38.9% gave no reason, 9.4% replied ‘can’t get to surgery’, 9.2% replied that they were ‘already in NHS’, 7.9% had ‘had a recent ECG’, 6.9% did not attend because of illness, 6.4% did not attend because of ‘old age’, 4.2% were ‘not interested’, 3.0% had ‘mental health problems’ and the remainder gave other reasons. All individuals undergoing an ECG in the SAFE study were asked to complete a screening acceptability questionnaire. Of those who completed the questionnaire, 94.4% felt that the letter/information sheet explained the tests properly, 3.7% considered screening to be inconvenient, 95.4% thought that health screening was important, 4.8% would have liked someone to discuss it more with first, 3.2% would have liked to talk about the tests with a GP first and 4% would have liked to come to a clinic appointment for more information.

The SAFE study recorded private individual costs associated with attending for screening, reporting a mean cost of £3.13 (95% CI £2.97 to £3.29) (2005 price year). Converting to 2016 prices (inflation factor 1.37, source ONS)87 gives a mean cost of £4.29 (95% CI £4.07 to £4.51).

Characteristics of all screened individuals and screen-identified atrial fibrillation patients

Of the five included RCTs, four19,40,82,83 reported the characteristics of the individuals who participated in the screening processes and the individuals who were identified as having AF through screening. The reported characteristics varied across studies, for both the individuals invited to screening and those identified as having AF through screening. Tables 14 and 15 show the characteristics of all screened individuals and screen-detected AF patients in the included studies respectively.

The mean age and sex distributions of the individuals invited to screening were comparable in the three studies that reported these individual characteristics. None of the reported characteristics of the patients identified to have AF through screening was comparable across the studies.

Adverse events with systematic screening

Few data were reported on adverse events associated with systematic screening, the rate and severity of complications or adverse events associated with the ECG or other forms of AF testing, psychological distress, change in quality of life and impact on well-being measured using a validated scale. The SAFE study19,84 was the only study that reported these outcomes.

A random sample of individuals randomised to the screening arms of the trial was sent the postal version of the EuroQol-5 Dimensions (EQ-5D) (to measure quality of life)88 and the shortened Spielberger anxiety questionnaire89 on study entry and again at the end of the screening period (approximately 17 months later), when it was also sent to all participants who had screened positive. In addition, all participants were asked to complete the Spielberger anxiety questionnaire immediately after screening.

The EQ-5D scores were similar across the systematic population screening and systematic opportunistic screening arms at baseline and also at 12 months’ follow-up. Similar results were found for anxiety scores.

However, based on the end-of-study questionnaire, individuals who were screen positive had a higher mean anxiety score (38.1, 95% CI 35.9 to 40.4) than individuals who were screen negative (34.6, 95% CI 32.4 to 36.8). EQ-5D scores were also significantly different, with screen-positive individuals reporting a lower quality of life score.

Cost-effectiveness of screening

Two RCTs19,82 reported the cost-effectiveness of systematic screening. The SAFE study19 reported the incremental cost per additional newly diagnosed AF case for systematic opportunistic, targeted and systematic population screening relative to no screening. The Morgan and Mant study82 reported the incremental cost per additional newly detected AF case for systematic population screening relative to systematic opportunistic screening. The results are shown in Table 16 in 2015 prices (using inflation factor 1.67 to convert the costs in the SAFE study19 from 2003 prices and inflation factor 1.77 to convert the costs in the Morgan and Mant82 study from 2002 prices (source ONS Consumer Price Index: Medical Services). 87 It should be noted that the SAFE study19 conducted a detailed cost assessment, whereas the Morgan and Mant82 study calculated the cost for the time of a short appointment with a nurse as £6 (2002 prices), which is £10.62 in 2015 prices. The SAFE study also included costs incurred by individuals attending for screening.

Subgroup and sensitivity analyses

The SAFE study19,84 provides results by sex and age (65–74 years and ≥ 75 years). Table 17 shows the proportion of newly diagnosed AF cases among those undiagnosed at baseline by age and sex. In the control arm, the rate of newly diagnosed AF was similar for men and women. For men, all screening strategies had a higher odds of newly diagnosed AF than usual practice and this was higher for systematic opportunistic and systematic population screening than for targeted screening. For women, there is no clear evidence that that there are any differences in new AF diagnoses under any of the screening strategies compared with usual practice. The ratio of ORs for women compared with men measures the ‘interaction’ effect of sex on the relative efficacy of the different screening arms. There is clear evidence of an interaction effect, with the ORs for the screening strategies compared with usual practice much closer to 1 for women than for men. This effect is stronger in the systematic population screening and targeted screening arms and less strong in the systematic opportunistic screening arm.

For age, there is clearly an increase in the proportion of newly diagnosed AF cases in the ≥ 75 years subgroup compared with those aged 65–74 years. However, the relative effectiveness (ORs) for the different screening interventions compared with the control was very similar for the different age ranges, with ratios of ORs estimated as close to 1, indicating no interaction between screening efficacy and age.

These analyses were also performed in the Cochrane review, which found strong evidence of a difference between sexes for systematic population screening (χ² = 6.64, p = 0.001) but not for systematic opportunistic screening (χ² = 2.95, p = 0.09). There was no strong evidence of a difference in AF detection between the two age groups. These findings are in line with our results.

There were too few data to perform any other subgroup or sensitivity analyses.

Summary of findings

Both systematic population and systematic opportunistic screening strategies were found to be effective compared with no screening. Both systematic population and systematic opportunistic screening had similar efficacy with an estimated 170 individuals needed to screen to detect one additional AF case compared with usual practice. Systematic screening targeted to high-risk individuals found a similar number of new cases as usual practice. One reason for this may be that individuals with comorbidities are more likely to undergo an increased number of visits to primary care and, as part of investigating and managing these conditions, will be screened in earnest or by default for AF.

Uptake of systematic population screening was typically around 50% although it was as high as 70% in one study. There was variability in uptake between practices, with uptake being between 22% and 70% in the SAFE study. 19 Reasons for not attending for screening were varied, although older age and decreased mobility were common themes. The proportion of individuals having their pulses checked under systematic opportunistic screening varied across studies (30% and 66%) and between practices within studies (from 8% to 93%). The proportion of individuals consulting with their GP was not reported and so it is unclear how much these uptake rates are driven by consultation rates, GPs offering pulse palpation and uptake of pulse palpation by individuals. Of those with an irregular pulse who did not have a previous diagnosis of AF, approximately 18% did not attend for an ECG test, although again this was variable across practices.

Subgroup analyses based on the SAFE study19,84 indicated that both systematic population and systematic opportunistic screening were more effective for men than for women, and this was especially the case for systematic screening. The Cochrane review49 discusses possible explanations for this, concluding that ‘differences in effect observed between the subgroups of men and women could be due to a combination of higher prevalence and greater rates of participation among men’ (p. 14). 49 The efficacy of screening was not found to vary with age, despite AF prevalence being strongly associated with age group.

The findings from the subgroup analyses have implications for a screening programme, in particular whether such a programme would be effective in women and whether or not there are interventions that could encourage women to participate in screening. Reasons given for not attending for screening included poor mobility, thus provision of transport to and from practices may increase participation. Systematic opportunistic screening (when a journey to the GP has already been arranged) appeared to be more effective for women, and this may be a consideration in determining the format of a population-based screening programme. Most evaluations of population-based screening have included individuals aged ≥ 65 years. The results of the subgroup analysis for age suggest that the relative efficacy of screening is likely to be unchanged if the age threshold is increased. However, the absolute rate of detection of new cases (per person screened) and cost-effectiveness are likely to increase with age if the proportion of screened individuals with undiagnosed AF increases with age.

Strengths and limitations

Although five studies were identified, because of studies having reported only interim results or concerns over methodology, only two had data extracted, with only one reporting the number of people with AF at baseline and therefore used to study the primary outcome. The SAFE study19,84 therefore remains the primary source of evidence for a comparison of different screening strategies. This situation should change, as three ongoing studies26,40,85,86 are due to report within the next 3 years. In particular, the STROKESTOP study26,40 aims to report the impact of screening on long-term outcomes (stroke and mortality), for which there is currently no robust evidence.

The SAFE study, on which most of our results are based, was a well-conducted study at low risk of bias on all domains except blinding of participants/personnel (which is not possible) and blinding of outcome assessment (unclear risk of bias). One issue with the SAFE study was that the baseline prevalence of AF was slightly higher in the control group than in the screening group. This could potentially introduce bias because, if more AF cases were previously diagnosed in the control practices, there may be fewer cases that could be diagnosed subsequently through screening or routine care. This was explored in the SAFE study using an individual participant analysis controlling for differences in baseline prevalence, which showed that the conclusions of the SAFE study were robust to adjustments to baseline prevalence. 19,84

The SAFE study19,84 found the incidence and 12-month prevalence of diagnosed AF to be similar between systematic opportunistic and systematic population screening. However, it should be noted that a large proportion of AF cases in the systematic opportunistic arm were detected outside the screening programme (i.e. in primary care but not as a result of activating the flag). Of the 74 new cases of AF identified in the systematic population screening arm, 22 (30%) were diagnosed outside the screening programme over the 12 months of the study. This proportion was greater for the systematic opportunistic screening arm, in which of the 75 newly identified cases, 44 (59%) were diagnosed outside the screening programme. One explanation for this is that GPs in the systematic opportunistic screening arm changed their usual practice to check for AF more frequently (the ‘Hawthorne effect’, where behaviour changes when under observation). If this is the case, then the full benefits seen in the systematic opportunistic arm may not be realised outside the context of a RCT. In contrast, Morgan and Mant82 concluded that systematic population screening is superior to systematic opportunistic screening, although there is no strong evidence to support this based on comparisons of new cases of AF. The Morgan and Mant study82 is complicated by the fact that an unreported number of patients with a previous diagnosis of AF were included. Also, only approximately 30% of people randomised to systematic opportunistic screening had their pulse assessed and, of those with an irregular pulse not referred for an ECG, it was assumed that a clinical diagnosis was made (which could have introduced detection bias).

In a screening programme it is likely that screening is offered repeatedly at given intervals (e.g. every 5 years). However, none of the RCTs identified in our review considered repeated screening, only a one-off screen. We therefore do not have any evidence on the effectiveness of such a screening strategy.

Findings in the context of previous research

A systematic review90 of cohort studies (including 26 prospective cohorts, two retrospective cohorts and two RCTs) that used an ECG or pulse palpation to identify AF in the general ambulant population (excluding studies performed in people with specific comorbidities such as stroke, hypertension or cardiac disease) found that the overall incidence of previously unknown AF was 1.0% (95% CI 0.89% to 1.04%), increasing to 1.4% (95% CI 1.2% to 1.6%) in those aged ≥ 65 years. The authors observed that incidence was similar in GP/outpatient clinics and community-based populations. The AF prevalence was 2.3% (95% CI 2.2% to 4.4%), increasing to 4.4% (95% CI 4.1% to 4.6%) in those aged ≥ 65 years. The denominator for these estimates was the number screened and not the number randomised (even for the RCTs), as used in our review. Despite this difference, the SAFE study19,84 found the 12-month incidence of unknown AF, with either opportunistic or systematic screening, to be a comparable 1.6% for individuals aged ≥ 65 years. However, the prevalence of AF after screening in the SAFE study19,84 was 8.8% in the control group and 8.4% in the screening groups, much higher than in the systematic review by Lowres et al. 90

Screening strategies

We assumed the presence of an established screening programme, with cohorts of individuals invited to attend for screening when they reach a certain age.

We compared the following population-based screening strategies in the base case:

• no systematic screening (current practice)

• single systematic population screen, inviting all individuals meeting the screening criteria

• single systematic opportunistic screen, inviting individuals meeting the screening criteria when they consult with their GP.

We did not include a targeted strategy in which high-risk individuals are screened, based on the findings from the SAFE study,19 which found that this strategy did not perform any better than no systematic screening.

In the base case, we report the results for one-off screening at ages 55, 60, 65, 70, 75 and 80 years, exploring various different screening tests (see Screening tests).

In sensitivity analyses, we explored repeated screening every 5 years for various initial screening ages (55, 60, 65, 70, 75 and 80 years) and final screening ages (55, 60, 65, 70, 75 and 80 years), which gave a total of 21 repeat screening strategies. We assumed that the optimal screening test and recruitment method (systematic opportunistic or systematic population) would be used for the repeat screening.

Population

The target population for systematic screening was the general population in England and Wales of a given age without a diagnosis of AF.

Screening tests

We included all screening tests identified in the systematic review (see Tables 2 and 3) for which we had sensitivity and specificity estimates compared with a gold standard (12-lead ECG interpreted by at least one cardiologist). The exception was the two-stage screening tests because we assumed that a diagnostic test would be given to those testing positive on a screening test, and so in effect all of the strategies represent two-stage screening. We acknowledge that this gold standard is appropriate only for persistent/permanent AF; however, we felt that persistent/permanent AF is likely to be the primary target of a population-based screening programme. Furthermore, estimates of the costs and benefits of anticoagulation are based on trials that do not generally include paroxysmal AF patients. However, we acknowledge that some paroxysmal AF cases may be detected and treated and we included this in our model.

We assumed that all individuals with a positive screening test result would receive a 12-lead ECG diagnostic test to confirm the diagnosis. The 12-lead ECG is likely to be performed in primary care, with a diagnosis being made in primary care when the result is clear and referral to a cardiologist made in a proportion of cases when the result is unclear.

Screening tests were defined according to the technology used and the interpreter (including algorithms). The following diagnostic tests were compared in the model (see Tables 2 and 3):

• 12-lead ECG:

  • interpreted by GP

  • interpreted by nurse

  • interpreted by algorithm

• single-lead ECG (including iPhone and other hand-held devices):

  • interpreted by cardiologist

  • interpreted by GP

  • interpreted by nurse

  • interpreted by algorithm

• > 1- but < 12-lead ECG:

  • interpreted by cardiologist

  • interpreted by algorithm

• pulse palpation (various methods):

  • interpreted by nurse

• modified blood pressure monitor (various devices):

  • interpreted by algorithm

• photoplethysmography:

  • interpreted by algorithm.

Although we have included tests interpreted by a cardiologist (as they were included in the DTA review), we note that cardiologists are unlikely to interpret screening tests in a primary care setting.

Outcomes of the economic evaluation

The economic model was evaluated using a probabilistic analysis in which the model inputs were simulated from distributions that reflected parameter uncertainty in the model. 91 Results from the model are averages (means) over these simulations. We present the mean total incremental costs, mean total incremental quality-adjusted life-years (QALYs) and mean incremental net benefit (INB) at a willingness to pay of £20,000 per QALY92 for each strategy relative to no screening (current practice). The INB is calculated as the incremental QALYs multiplied by £20,000 minus the incremental costs. Positive values of the mean INB indicate strategies that are cost-effective at the £20,000 per QALY threshold. We aimed to maximise the mean INB. We also present the incremental cost-effectiveness ratio (ICER), which is the mean incremental costs divided by the mean incremental QALYs. The INB and ICER are related, so that a positive mean INB would have an ICER of < £20,000 and would be considered cost-effective at a willingness to pay of £20,000 per QALY. Conversely, a negative mean INB would have an ICER of > £20,000 and would not be considered cost-effective at a willingness to pay of £20,000 per QALY. Uncertainty in the optimal strategy is represented using cost-effectiveness acceptability curves (CEACs) and by displaying the simulations in the incremental cost-effectiveness plane (incremental QALYs plotted against incremental costs). The CEAC plots the probability that each strategy is most cost-effective against willingness to pay per QALY. For clarity we omit strategies from the CEAC plots that have a < 10% probability of being most cost-effective at any willingness-to-pay threshold. To compare different screening strategies by age at first screen, and number of repeated screens, we multiply the per-person INB by the population size that will benefit (for a given age cohort). Costs and outcomes over a lifetime time horizon were included because treatment for AF aims to reduce future events. Future costs and benefits are discounted at an annual rate of 3.5% in line with recommendations of the UK treasury. 93

Summary of previous economic evaluations of screening for atrial fibrillation

True positive on the initial screening test

Of those who are tested, a proportion will be true positives (screen-positive results that are confirmed using a 12-lead ECG gold standard diagnostic test), and these individuals may benefit from early detection of AF. All of these individuals will accrue the cost of the initial screening test plus a gold standard 12-lead ECG test interpreted by a GP, except a proportion of unclear cases that are additionally interpreted by a cardiologist.

For those individuals with a stroke risk CHA₂DS₂-VASc score of ≥ 2, anticoagulation therapy is recommended unless contraindicated (bleeding risk using the HAS-BLED score, dementia, epistaxis, frailty or patient preference). However, in a proportion of cases, individuals not contraindicated for anticoagulation are not prescribed anticoagulants by their GP. In the base case we assumed that all of the individuals not contraindicated and prescribed anticoagulation therapy use DOACs. We assumed that the most cost-effective treatment (apixaban) in the DOAC model37 is used for all patients who receive DOACs, although it should be noted that similar results would be obtained using other DOACS. 37 In the sensitivity analyses we assumed that 50% use DOACs and the remainder use warfarin (INR range 2–3). We assumed that individuals with a CHA₂DS₂-VASc score of ≤ 1, in whom OACs are not recommended,11 and patients contraindicated to, not prescribed or choosing not to take OACs, would receive aspirin. This was based on clinical advice, confirmed by Quality and Outcomes Framework (QOF) statistics,3 that these patients will likely receive aspirin in practice, despite aspirin not being recommended in the latest NICE clinical guidelines. 11

Regardless of anticoagulation therapy, we assumed that all symptomatic patients would receive a standard beta-blocker for rate control, which incurs a cost and improves quality of life benefit. The costs and benefits of rate control were assumed to be independent of anticoagulation therapy. Rhythm control is not currently recommended except for patients younger than those who would be invited to a screening programme, and so we did not include it in our model.

True negative on the initial screening test

Of those tested, a proportion will be true negatives on the initial screening test. These individuals incur the cost of the initial screening test but do not receive any further incremental benefits or costs from the screening strategy compared with current practice (S0, no screening).

False positive on the initial screening test

We assumed that individuals with a false-positive result on the initial screening test will be given a diagnostic 12-lead ECG test, which correctly identifies that they do not have AF (i.e. the confirmatory test is 100% specific). All of these individuals incur the cost of the initial screening test plus the cost of the diagnostic 12-lead ECG test interpreted by a GP and, in a proportion of unclear cases that are additionally interpreted by a cardiologist, an additional expense will be incurred. As with the true negatives, these individuals do not receive any further incremental benefits or costs from the screening strategy compared with current practice (S0, no screening). It should be noted that some false positives may be diagnosed with another condition (e.g. atrial flutter) and may receive some benefit from an AF screening programme. However, we do not capture this benefit in our model.

False negative on the initial screening test

Of those tested, a proportion will be false negatives on the initial screening test, that is, they have AF but are not referred for a diagnostic 12-lead ECG test and gain no benefits from treatment. This includes individuals with paroxysmal AF who are not in AF at the time of the screening test. These patients incur the cost of the initial screening test but no future incremental costs and benefits compared with current practice (S0, no screening).

Prevalence of atrial fibrillation in the UK by age and sex

The prevalence of AF strongly depends on both age and sex. 31 We searched for studies reporting prevalence in a general population (both diagnosed and undiagnosed) broken down by age and sex in Western European countries, with data collected from 2000 onwards. Table 19 summarises the studies identified and Figure 19 shows plots of prevalence against age for men and women separately. All studies show similar patterns, although there are some differences between studies, with, in particular, the study by Baena-Díez et al. 129 finding a lower prevalence of AF in older people than the other studies. We chose to use the prevalence estimates from Norberg et al. 2 as this was the most recent and largest study providing the most detailed results. The prevalence estimates from Norberg et al. 2 are shown in Table 20.

FIGURE 19.

Estimated prevalence of AF by age and sex plotted for the studies identified in our literature review relevant to a contemporary Western European population. (a) Men; and (b) women.

Population size eligible for screening at each age

The size of population eligible for screening at each age was estimated as the total number of men and women in the associated 5-year age range as reported by the Office for National Statistics (ONS)130 for England and Wales in 2014 divided by 5. These estimates are reported in Table 21.

Uptake of systematic population screening

Seven studies reported the uptake of systematic screening in which individuals were invited to screening by letter. These studies are summarised in Table 22. Uptake of screening varied between 53% and 76% across studies/populations, with both the minimum and the maximum uptake seen in UK general practice populations. We pooled the data from all seven studies using a random-effects meta-analysis, reported in Table 22. However, uptake of screening is likely to depend on age. The only study that reported uptake by age is the SAFE study19 (Table 23). The results for the < 75 years age group are in line with the results from the studies reported in Table 22, whereas uptake in the older age group is lower. We used the figures in Table 23 to inform beta distributions for use in our model.

Proportion of initial invitees requiring a reminder letter

Two studies reported the proportion of reminder invitations sent during systematic screening (Table 24). This proportion varied from 0.4 to 0.6. We used the results from the SAFE study as being the most representative of a UK population and a normal distribution was used to reflect uncertainty in this estimate.

Pulse taking in flagged individuals in opportunistic screening

Two studies reported the number of flagged individuals who had pulse palpation under a systematic opportunistic screening strategy (Table 25). There is a large discrepancy between these results. We used the results from Hobbs et al. 19 to inform beta distributions in our base-case model as this was the more recent, larger study and the results were broken down by age group (< 75 and ≥ 75 years). Both of these studies used paper flags, whereas in current practice computerised flags are likely to be more feasible and sustainable for recurrent opportunistic screening; however, GPs may be less likely to respond to computerised flags because of ‘alert fatigue’. 133 We used the results from Morgan and Mant82 in a sensitivity analysis (assuming the same rate in both age groups as the rates were not reported separately), to represent a lower uptake of systematic opportunistic screening.

The proportion of 12-lead electrocardiograms interpreted by a general practitioner that require further interpretation by a cardiologist

We did not identify any information on the proportion of 12-lead ECGs read by a GP that would require further interpretation by a cardiologist and so took advice from our clinical experts. On the basis of this, we assumed that 10% of ECGs read by a GP would require a cardiologist interpretation to confirm the diagnosis. We assumed a 95% CI from 1% to 20%.

Rate of diagnosis of atrial fibrillation in the absence of screening

The SAFE study compared systematic and opportunistic screening against a no screening control. 19 Table 26 shows the diagnosis rate in the SAFE study no screening arm. The rate of detection in the absence of screening is low and it is likely that in many cases diagnosis of AF occurs only after a stroke or another acute event. Our DOAC model37 already tracks the incidence of adverse events in individuals with AF not on anticoagulation therapy and assumes that anticoagulation therapy is initiated following a stroke in those who are not contraindicated. Therefore, to avoid double counting a diagnosis of AF following acute events, we assumed that there would be no ‘serendipitous’ diagnosis of AF in the no screening arm.

The proportion of atrial fibrillation that is undiagnosed

It is possible to estimate the proportion of AF that is undiagnosed from screening studies (Table 27). However, these estimates are not directly comparable with each other because different screening tests are used. Furthermore, we might expect these to be underestimates because a one-off ECG is unlikely to detect paroxysmal AF. In the two studies with prolonged monitoring,22,40 higher undiagnosed proportions were seen than in many studies with one-off monitoring. A recent report by Public Health England1 used prevalence estimates from a Swedish study2 (that comprehensively searched ECG records in a variety of sources together with patient records) and applied these estimates to the demographics of GPs in England to obtain predicted AF cases in England. Comparing this prediction with the figures reported in the QOF3 for England GPs gives an estimate of the proportion of AF that is undiagnosed of 0.348. Given that this estimate is in line with the most intensive screening study,22 it has been adjusted to match demographics in England and prevalence estimates are available broken down by age and sex, and we used this estimate of the proportion of undiagnosed AF in our model.

The proportion of atrial fibrillation that is paroxysmal, persistent and permanent

Although several studies provided information on the proportion of patients with paroxysmal, persistent and permanent AF, these were primarily studies of diagnosed patients, which will underestimate the proportion of paroxysmal patients quite dramatically. Instead, we were interested in the proportion of paroxysmal AF in a systematic screening population (Table 28). The proportion of AF of each type varied widely across screening studies. Engdahl et al. 22 and Arronnson et al. 100/Svenberg et al. 40 used a continuous monitoring screening test in population-based screening that was able to detect paroxysmal AF. Both of these studies detected a higher proportion of paroxysmal AF than persistent or permanent AF, suggesting that paroxysmal AF is more prevalent in people with undiagnosed AF, but that most screening studies are not able to detect it. For example, among the new AF cases detected by Sanmartín et al. ,138 only 6% were paroxysmal AF. In our model we assumed that 6% of screen-detected AF is paroxysmal AF and the remaining 94% is permanent or persistent AF. In sensitivity analysis, we varied the proportion of paroxysmal AF that can be detected.

For the proportion of all undiagnosed AF that is paroxysmal we pooled the data from Engdahl et al. 22 and Arronsson et al. 100/Svenberg et al. 40 using a fixed-effects meta-analysis (see Table 28).

The proportion of atrial fibrillation that is asymptomatic (by atrial fibrillation type)

We were interested in the proportion of patients, by AF type, who are asymptomatic in a population-based screening population (Table 29). Two of the studies136,139 gave results for screen-detected AF based on electrocardiography (and hence mostly persistent and permanent AF). These studies found a high proportion of AF that is asymptomatic (0.8 and 0.49 respectively), as would be expected in a screening context. Another study14 reporting 1 year of follow-up of newly diagnosed AF patients found a similarly high proportion of asymptomatic patients (0.87). In patients with diagnosed chronic AF (persistent or permanent), there was a consistent finding of approximately 30% of patients being asymptomatic. 14,140,141 The proportion of paroxysmal AF that was asymptomatic was higher (estimates of 0.4142 and 0.714).

We pooled the data from Deif et al. ,136 Smyth et al. 139 and Nieuwlaat et al. 14 in our model using a fixed-effects meta-analysis (see Table 29).

Disease progression from paroxysmal atrial fibrillation to persistent/permanent atrial fibrillation

Table 30 summarises studies reporting information on the rate at which paroxysmal AF patients progress to persistent and permanent AF. The majority of the studies report the rate at which paroxysmal AF progresses to permanent AF. The results are fairly consistent across studies, with the only UK study143 giving a very similar estimate to the largest study conducted in the US (Holmqvist144). Holmqvist144 also provides the figures to estimate the rate that individuals progress from paroxysmal to chronic (persistent or permanent) AF, which is what is required in our model. Given the agreement between Holmqvist144 and the UK study, we chose to use Holmqvist144 to estimate the rate of progression from paroxysmal to chronic AF in the economic model.

Hazard ratios for stroke and mortality risk by whether asymptomatic or not

Xiong et al. 124 presented a systematic review of studies comparing mortality and stroke outcomes in symptomatic and asymptomatic patients. Because symptomatic patients may receive different anticoagulation therapy and other treatments from asymptomatic patients, we present here the results from the RCTs included in the study by Xiong et al. ,124 in which allocation of symptomatic and asymptomatic patients across trial arms should have been balanced. The results from these two RCTs35,140 included in the study by Xiong et al. 124 are summarised in Table 31.

Both studies show that symptomatic patients are more likely to have a composite outcome of mortality and other events than asymptomatic patients; however, Flaker et al. 35 show that, when adjusted for other risk factors, this effect diminishes (HR 1.14, 95% CI 0.87 to 1.50). Flaker et al. 35 found a similar result for all-cause mortality, with an adjusted HR of 1.07 (95% CI 0.79 to 1.46). Because the risk of stroke and mortality in our model already accounts for risk factors (CHA₂DS₂-VASc score, previous MI, previous stroke), the adjusted estimates are of relevance to our model and, as these suggest that there is no evidence that risk is any different in asymptomatic patients from symptomatic patients, we do not allow for differential risk according to symptomatic status in our base case. In sensitivity analysis, we assumed that the event rate in screen-detected patients is lower than that for routinely detected AF using the unadjusted HR for symptomatic AF (representing non-screen-detected AF) compared with asymptomatic AF (representing screen-detected AF) found by Flaker et al. 35 of 1.31 (95% CI 1.04 to 1.65).

Hazard ratios for stroke and mortality risk by atrial fibrillation type

To obtain estimates of the HRs for events for different AF types, we need to avoid confounding by treatment, as permanent AF is more likely to be treated than paroxysmal AF. We selected studies in which treatment allocation was the same across AF type, mostly post hoc analyses of pooled RCT arms, but also one study in which the results were broken down by treatment (Table 32). Only one study147 gave the HR for mortality for persistent relative to paroxysmal AF, but this was in close agreement with the HR for stroke in that same study, which supports an assumption that the HRs are the same for both stroke and mortality outcomes. The majority of the studies gave results for stroke comparing persistent with paroxysmal AF, showing in general an increased risk of stroke in persistent compared with paroxysmal AF. Vanassche et al. 9 also presented the HR for permanent compared with paroxysmal AF, which suggested that there is a ‘dose–response relationship’, with risk increasing from paroxysmal to persistent to permanent AF. We pooled the HRs from the studies reporting HRs (and CIs) for stroke in persistent compared with paroxysmal AF. Because there were only three such studies, we pooled them using a fixed-effects model. In our model we do not distinguish between persistent and permanent AF; for at least one of the studies, the result for persistent AF also included permanent AF. We therefore used the HR for persistent compared with paroxysmal AF in our model. We assumed that the HR is the same for mortality as it is for stroke.

The proportion of screen-detected atrial fibrillation with a CHA₂DS₂-VASc score of ≥ 2

Table 33 details the results from six studies providing information on the proportion of AF patients with a CHA₂DS₂-VASc score of ≥ 2 in a screening or community population. The results are given separately for known (diagnosed) AF patients and screen-detected AF patients. We pooled the data from all four studies reporting results for screen-detected AF using a fixed-effects meta-analysis for use in our model.

The proportion of screen-detected atrial fibrillation cases with a previous history of myocardial infarction

Table 34 shows the results from studies reporting the proportion of screen-detected AF cases with a previous history of MI. We also show the results for studies of newly diagnosed AF; however, it should be noted that these patients are more likely to be symptomatic at diagnosis than screen-detected AF patients. This is reflected in the results, with a much higher proportion of newly diagnosed AF cases having a previous history of MI than screen-detected or asymptomatic AF cases. The proportion of patients with a previous MI was 5.4%135 and 0%138 in the two studies of screen-detected AF. We pooled the data from both studies reporting results for screen-detected AF using a fixed-effects meta-analysis for use in our model.

The proportion of screen-detected atrial fibrillation cases with a previous history of ischaemic stroke (or transient ischaemic attack)

Table 35 shows the results from studies reporting the proportion of screen-detected AF cases with a previous history of ischaemic stroke or TIA. We also show the results for studies of newly diagnosed AF. The results do not show a consistent pattern of previous stroke risk in newly diagnosed AF patients compared with screen-detected AF. The results are variable even when restricting to screen-detected AF. We pooled the data from all seven studies (eight arms) reporting results for screen-detected AF using a random-effects meta-analysis to account for heterogeneity, reported in Table 35. However, the most relevant study to a UK screening population is the SAFE study,153 the estimate for which is only just contained in the CrI around the random-effects mean. We elected to estimate the history of previous stroke in a UK screen-detected AF population by pooling the systematic population and systematic opportunistic arms of the SAFE study for use in our economic model (reported in Table 35). We assumed a normal distribution for this history of previous stroke. In sensitivity analysis, we used the random-effects mean across all studies reporting screen-detected AF.

Anticoagulation therapy for screen-detected atrial fibrillation with a CHA₂DS₂-VASc score of ≥ 2

Our model required information on the proportions of screen-detected AF patients with a CHA₂DS₂-VASc score of ≥ 2 who are prescribed OACs, who are contraindicated to or who decline OACs and who simply do not have OACs prescribed. The studies providing information relevant to this are summarised in Table 36. Only one study126 provided information on the HAS-BLED bleeding risk score in AF patients with a CHA₂DS₂-VASc score of ≥ 2, finding 2.49% with a HAS-BLED score of ≥ 3. It should be noted that, although the HAS-BLED score can inform clinical decision making on whether or not a patient is suitable for OACs, other factors will also be taken into account. A more practical measure of those who are not suitable for OACs is the reported proportion contraindicated to, or preferring not to take, OACs. Three studies3,118,154 reported this information in patients with a high stroke risk (albeit using the CHADS₂ rather than the CHA₂DS₂-VASc risk score), with all three studies giving very similar estimates (see Table 36). The QOF report3 represents the most recent evidence covering the majority of GPs in England, with an estimated 12.94% of patients contraindicated or declining OACs (‘exceptions’ rate from the QOF report3), which we used in our model.

There is a wide variation in the proportion of AF patients with a CHA₂DS₂-VASc score of ≥ 2 who are prescribed OACs (see Table 36). Again, we used the figures from the QOF report3 as they provide the most contemporary relevant evidence in England, giving an estimated 74.30% of AF patients who are prescribed OACs. It should be noted that this leaves 12.76% of patients who are not contraindicated but who do not receive OACs. In our model base case, we assumed that all patients not prescribed OACs (12.94% contraindicated or declining OACs plus 12.76% not prescribed OACs for other reasons) receive aspirin.

It may be that patients with screen-detected AF are less likely to be prescribed OACs than patients with AF detected routinely, in which case the uptake of OACs may be lower than 74.30%. Alternatively, implementation activities may lead to an increase in uptake of OACs in the future. We explored the robustness of our results to this in sensitivity analyses, with the uptake of OACs reduced to 50% or increased to 87.0% (so that all who are not contraindicated or declining OACs receive OACs).

Uptake of DOACs has been slow, with some studies reporting that DOACs represent as little as 3.79%156 of all OACs prescribed for AF. Figure 20 shows the trends in prescribing of OACs in England over time. It can be seen that there has been a steady trend of increased prescribing of DOACs. Figure 20 includes all patients prescribed OACs. We hypothesise that the trend towards increased DOAC use is driven by prescribing in newly diagnosed patients, rather than by switching treatment in patients who have historically been prescribed warfarin. This view is supported by the STROKESTOP study,40 which found that, of those with newly diagnosed AF who were started on OACs, 73% were prescribed DOACs. For our model we therefore assumed in the base case that the proportion of OACs that are DOACs in newly diagnosed AF patients is 75%. We varied this to 50% and 100% in sensitivity analyses. We assumed that the OAC received is apixaban (most cost-effective in the DOAC model37) but note that the results would be similar if other DOACs were used instead.

FIGURE 20.

Trends over time in the proportion of OACs prescribed that are warfarin or DOACs in England, using data from Powell-Smith A, Goldacre B, OpenPrescribing.net, EBM DataLab, University of Oxford, 2017.

Incremental quality-of-life benefits and costs from rate control compared with no rate control

Incremental lifetime costs and QALYs associated with rate control (compared with no rate control) are assumed to be fixed costs and benefits obtained by patients who are prescribed rate control as a result of screening. In practice, some of these patients would eventually have been detected with AF and given rate control, so our model will overestimate these benefits. We performed a sensitivity analysis assuming an annual rate of detection of AF in the absence of screening or an AF-related stroke.

The model developed by Lord et al. 116 provides information on the incremental costs and QALYs over a lifetime for rate control compared with no rate control for persistent AF. The authors found an incremental cost of –£260 (i.e. cost saving), which we inflated to –£274.55 at 2015 prices, and incremental QALYs of 0.0133 for rate control. We assumed these values in our model for symptomatic patients detected by screening.

Sensitivity and specificity of the screening tests

The assumed values for the sensitivity and specificity of the screening tests were the mean values from our HSROC model reported in the systematic review of DTA studies (see Tables 4 and 5).

Screening costs

The costs of the screening tests consist of the capital cost of any equipment required, the cost of consumables for each test performed and the cost of clinician time to administer and interpret the test results. In the case of systematic population screening there will also be costs associated with the screening invitation.

Equipment costs vary widely across screening tests, from approximately £1500 for a 12-lead ECG to £100 for a modified blood pressure monitor, £24 for photoplethysmography and no cost for pulse palpation. The capital cost per patient will depend on, among other factors, the lifespan of the device and the throughput of patients for AF screening. Patient throughput will, in turn, depend on practice size and demographics, screening uptake and the frequency of screening. The calculation is further complicated for multipurpose devices such as ECG machines, which will not be used solely for AF screening. In practice, even the relatively high purchase cost of an ECG machine will be spread widely over the large number of patients who use it. In our base-case analysis, we assumed that the capital cost of equipment per patient screened is zero and that the screening test costs consist of invitation costs, the costs of consumables and the costs of clinician and administrative time taken to administer and interpret the test results. We discuss the likely impact of this assumption and the issue of investment in equipment, particularly for screening strategies (e.g. photoplethysmography, modified blood pressure monitors) for which GPs do not currently have the equipment, in the discussion section of the next chapter.

We used the results from the SAFE study19 to inform the costs of systematic population and systematic opportunistic screening. The cost of invitation to systematic population screening consists of the costs of the invitation letter, postage and 1 minute of administrator time. Estimates for these costs are based on the trial data from Hobbs et al. ,19 but we inflated from 2003 prices to quarter 1 2015 prices using the ONS Consumer Price Index for Medical Services (DKC3). 157 The quarter 1 2015 prices were chosen to maintain consistency, with costs used to generate the results of the DOAC model. 37 These costs are summarised in Table 37.

Test costs

We used PSSRU 2014/15 data123 to estimate staff costs for administering and interpreting the screening and diagnostic tests. An hour of face-to-face contact with a nurse was assumed to cost £56 (including qualification costs) (see Table 37). An hour of general medical services activity by a GP was assumed to cost £129 (including qualification costs but excluding direct care staff costs). An hour of cardiologist time was taken as the cost of 1 hour of ‘consultant medical’ time, which was £137 (including qualification costs).

We used the times for each test estimated by Hobbs et al. 19 for the times to complete systematic opportunistic screening. We also used their estimated interpretation time of 1 minute for all ECGs interpreted by a nurse, GP or cardiologist. Systematic population screening would require a dedicated appointment with a nurse. A typical practice nurse consultation is estimated to last approximately 15 minutes;123 however, for routine screening the consultation time could be considerably less. Staff times assumed for each test are provided in Table 38.

If the initial screening test was a 12-lead ECG then the diagnostic test cost would be only the cost of the 1-minute interpretation by a GP or cardiologist, which is £2.15 and £2.28 respectively. If the initial screening test was something else (e.g. photoplethysmography) then the diagnostic test cost would be the full cost of a 12-lead ECG with administrator and nurse appointment time. We did not include the cost of a follow-up appointment with a GP or cardiologist for a true-positive 12-lead ECG diagnostic test as this was included in the management costs of the long-term model (see Long-term costs and benefits).

Following Hobbs et al. ,19 all test costs included 2 minutes of administrator time. Administrator time was priced at £10.93 per hour in Hobbs et al. 19 and was inflated to £18.22 (quarter 1 2015 prices) using the ONS Consumer Price Index for Medical Services (DKC3),157 giving a cost of £0.61 for 2 minutes for each test.

We assumed that each lead of the ECG machine required one pad and that a > 1- and < 12-lead ECG, with an average of 2.5 leads, required 2.5 pads. Following clinical advice we assumed that no alcohol wipes were used for the ECGs.

The total estimated costs and their underlying assumptions for systematic opportunistic and systematic population screening are reported in Table 38.

Long-term costs and benefits

Anticoagulation therapy (directly acting oral anticoagulant model37)

A previously developed Markov model37 was used to estimate the long-term costs and outcomes for patients with screen-detected AF (true positives on initial screen, confirmed by 12-lead ECG). The model includes a HR for events (stroke, systemic embolism, TIA) affected by AF type (paroxysmal relative to persistent or permanent). We combined persistent and permanent AF in the model because the RCTs informing the efficacy of anticoagulation therapy predominantly combined persistent and permanent AF patients. The model depends on age, sex, previous history of ischaemic stroke or TIA and previous history of MI.

The discrete-time Markov multistate model used a cycle length of 3 months, as in other recent models. 158–160 We ran the model for a cohort for a given starting age and using a lifetime time horizon. The model structure is illustrated in Figure 21. Each treatment strategy has the same model structure but different costs, utilities and event probabilities. From any state, a patient can have a clinically relevant (extracranial) bleed, an intracranial haemorrhage, an ischaemic stroke or a MI, all of which have long-term consequences that are modelled. Patients can also experience a TIA or systemic embolism, which are transient states, or can discontinue or switch treatment because of any event, or die. Patients on apixaban who experience a bleed or intracranial haemorrhage will be switched to warfarin (INR range 2–3); patients experiencing a further bleed or intracranial haemorrhage are then switched to no treatment. Patients initially on no treatment or aspirin who experience a stroke and who are not contraindicated to DOACs and have no history of a bleed or intracranial haemorrhage are assumed to initiate apixaban treatment.

FIGURE 21.

Discrete-time Markov model structure for screen-detected AF patients. B, major bleed; ICH, intracranial haemorrhage; S, stroke.

Transition probabilities were derived using HRs for apixaban and aspirin relative to warfarin estimated from a systematic literature review. 37 The warfarin arms were used to estimate baseline hazards. We relied on previous meta-analyses to estimate the relative effect of warfarin (INR range 2–3) compared with no treatment. 161 Evidence from the literature was used to estimate the effect of previous events on stroke, mortality, MI, systemic embolism, TIA and bleed risk. 162,163 Average drug costs were based on the British National Formulary March 2015 update,164 and it was assumed that oral apixaban and aspirin incurred no monitoring or administration costs. Acute management costs for systemic embolism, MI, TIA and clinically relevant bleeding were obtained from 2013/14 NHS reference costs. 165 Acute and long-term management costs for ischaemic stroke and intracranial haemorrhage were taken from a study of AF patients on a UK stroke registry. 166 Utilities were identified from a previous NICE technology appraisal submission on rivaroxaban,159 which included a systematic literature search for evidence on EQ-5D utility index scores in health states related to AF.

The model was modified to include management costs for AF reported published in the literature. 110 Costs were available for a 12-week initialisation period and annual ‘maintenance’ periods thereafter, summarised in Table 39. These costs included the costs of blood testing, medication, primary and secondary care visits, further diagnostic testing and inpatient visits. We adjusted the costs by removing the blood testing and medication costs, which were already accounted for in our model. We also inflated costs to quarter 1 2015 prices157 to maintain consistency with the other model costs. We used a log-normal distribution with the mean and SD in Table 39 (‘adjusted total’) for the initialisation and maintenance periods.

TABLE 39 - Assumed non-anticoagulation therapy costs in the economic modela

Costs are provided for a 12-week initialisation period and subsequent annual ‘maintenance’ time periods, based on Kassianos et al. 110

Component	Initialisation period (£)		Maintenance period (£)
	Mean	SD	Mean	SD
Blood testing	5.12	9.22	7.17	11.27
AF medications	9.22	10.25	50.20	50.20
Total	964.14	1120.90	436.48	611.68
Adjusted total	949.80	1120.99	379.10	613.84

Summary of model inputs

Table 40 summarises the model parameters and assumed values and distributions and the evidence sources used for each.

Screening test

We found that results from the evidence synthesis of the DTA studies were robust to the various sensitivity analyses that we performed. However, there was a high degree of uncertainty in the estimates of sensitivity and specificity. Also, the photoplethysmography screening test was based on a single study that did not take place in a primary care setting and the study was rated as being at high risk of bias for the index test. Furthermore, photoplethysmography and modified blood pressure monitors are not currently available in most primary care settings. We therefore performed a sensitivity analysis of the comparison between age at screening, using the most cost-effective screening test and omitting (1) photoplethysmography and (2) photoplethysmography and modified blood pressure monitors.

Repeat screening

In sensitivity analyses, we explored repeated screening every 5 years for various initial screening ages (55, 60, 65, 70, 75 and 80 years) and final screening ages (55, 60, 65, 70, 75 and 80 years), giving a total of 21 repeat screening strategies. The model structure and input parameters were as for the base case; however, the total proportion of diagnosed or undiagnosed AF at each age (see Table 40) was adjusted to allow for the AF detected at previous screening rounds. We considered a single-age cohort that may be screened at various ages as it gets older. We assumed that total (diagnosed and undiagnosed) prevalence increases with age in this cohort according to the patterns reported in Table 20, which reflects increasing incidence with age. We subtracted from this total prevalence the proportion of patients diagnosed in previous screening rounds and also the proportion of patients diagnosed in the absence of screening, assumed to be a 1% diagnosis rate per year based on the control arm of the SAFE study19 (see Rate of diagnosis of atrial fibrillation in the absence of screening), which was varied to 5% in sensitivity analysis. Comparisons across different screening strategies related to this age cohort and therefore we needed to consider the cost and QALY implications for the cohort as a whole to make fair comparisons. This was achieved by using population estimates reported in Table 21 to obtain population costs and QALYs for each screening interval, discounted at 3.5% back to the time of the initial screen. The total population costs and QALYs for each strategy were the sum of the costs and QALYs for each screening interval for a given strategy, with the last screening interval representing the remaining lifetime time horizon.

Uptake of systematic opportunistic screening

In a sensitivity analysis, we used a lower uptake of systematic opportunistic screening (0.305, 95% CI 0.282 to 0.329), as reported in Morgan and Mant,82 which was assumed to be the same for all age groups; this compares to the base case in which we used the results from the SAFE study19 (< 75 years age group: 0.699, 95% CI 0.681 to 0.716; ≥ 75 years age group: 0.670, 95% CI 0.648 to 0.691).

Uptake of electrocardiograms in those with a positive screening result

In the base case, we assumed that all those with a positive screening test result would be offered and would accept a diagnostic ECG to confirm the diagnosis. The SAFE study19 found that only 72.5% of those with an irregular pulse and no previous AF diagnosis agreed to have an ECG test. We therefore conducted a sensitivity analysis assuming that only 72.5% of those with a positive screening test result will agree to have an ECG. The remaining 27.5% are assumed to be a mixture of true positives and false positives (in the same proportion as for those who agree to have an ECG).

Diagnosis in the absence of screening or atrial fibrillation-related stroke

In the base case, we included an annual rate of diagnosis in the absence of screening of 1% based on the control arm of the SAFE study. 19 We increased this to 5% in sensitivity analysis, as assumed by Aronsson et al. 100

Previous history of stroke in UK screen-detected atrial fibrillation

We investigated the robustness of our conclusions to using an estimate of the previous history of stroke in UK screen-detected AF based on the random-effects meta-analysis of screening studies (0.112, 95% CrI 0.072 to 0.170); in the base case we used the estimate of the previous history of stroke from the SAFE study19 alone (0.074, 95% CI 0.032 to 0.116).

Proportion of oral anticoagulants prescribed that are directly acting oral anticoagulants

In the base case, we assumed that 75% of OACs prescribed are DOACs. Prescribing practice is changing, with the DOACs becoming the OACs of choice for newly diagnosed AF. However, warfarin is still prescribed, with some patients/practitioners preferring the regular monitoring of warfarin. We performed a sensitivity analysis of the proportion of OACs that are DOACs, using assumed values of 50% and 100%.

Uptake of oral anticoagulants

In the base case, 74% of patients who are eligible for OACs receive them, 13% are contraindicated or do not wish to take them and the remaining 13% do not receive OACs but could do. In sensitivity analyses, we explored the impact of a reduced uptake of OACs of 50% and also an improvement in the uptake of OACs, so that all 87% of patients who are not contraindicated or who prefer not to take OACs receive OACs.

Hazard ratios for stroke and mortality risk for routine-detected atrial fibrillation compared with screen-detected atrial fibrillation

In sensitivity analysis, we assumed that the event rate in screen-detected patients was lower than that for routinely detected AF using the unadjusted HR for symptomatic AF (representing non-screen-detected AF) compared with asymptomatic AF (representing screen-detected AF) found by Flaker et al. 35 of 1.31 (95% CI 1.04 to 1.65), instead of a HR of 1, as assumed in the base case.

Summary of results from the review of studies on the natural history of atrial fibrillation screening

We identified several studies that estimated the prevalence of AF by age and sex. All showed similar trends, although there was variation in the overall prevalence level, most likely because of the method of measurement used together with population differences. The studies with the most intensive ascertainment of AF diagnosis gave the highest estimates of prevalence and were in line with each other. 2,125 We used the estimates from the most comprehensive and recent of these studies2 in our model to represent the overall prevalence of AF (both diagnosed and undiagnosed) by age and sex.

Uptake of systematic population screening was fairly consistent across studies, with a mean estimate of 64% (95% CrI 54% to 73%), although there was a strong effect of age seen in the SAFE study,19 with a lower uptake of screening in the ≥ 75 years age group of 43% (95% CI 41% to 45%) compared with uptake in the < 75 years age group of 61% (95% CI 59% to 63%). In contrast, there was a high degree of variability across studies in the uptake of systematic opportunistic screening, ranging from 30% to 70% (regardless of age), raising concern that the high level of opportunistic screening achieved in the SAFE study19 may not be seen in practice. Furthermore, with the introduction of computerised flags for a range of health issues, GPs may be less likely to respond to computerised flags as a result of ‘alert fatigue’. 133

The proportion of AF that is undiagnosed was estimated to be 35%; however, this estimate was very variable across screening studies, reflecting the method of screening and test used as well as differences in populations. As expected, the higher estimates came from studies with more intensive screening tests (e.g. long-term continuous monitoring). It is clear that if the objective is to detect paroxysmal AF then long-term continuous monitoring is necessary. Whether or not this is feasible in a population-based screening setting, however, is unclear.

As expected, our results suggest that a high proportion of screen-detected AF is likely to be asymptomatic. Asymptomatic AF has a lower risk of stroke and mortality; however, after adjusting for other risk factors, there is no evidence that the risk of stroke or mortality depends on whether AF is asymptomatic or not. 35 We explored the robustness of our conclusions to this assumption in a sensitivity analysis.

We estimated that paroxysmal AF progresses to chronic (persistent or permanent) AF at a rate of 0.15 per year. There is evidence of a ‘dose–response’ relationship, with stroke and mortality risk increasing as AF progresses from paroxysmal to persistent to permanent AF.

Strengths and limitations

Our review of the natural history of AF and of AF screening covered a very broad set of review questions to identify evidence relevant to our economic model. Our review focused on evidence relevant to screening populations in a primary care setting in the UK. When appropriate, we pooled results from studies in meta-analysis. We also took care to account for ‘confounding by treatment’ in estimating mortality and stroke risk according to type of AF and whether symptomatic or not.

Because of the scale and scope of this review, it was only possible for a single reviewer (NJW) to review the literature, and so it was not possible to discuss inclusion with another reviewer. We ran searches from 2000 onwards only, and so we may have missed relevant evidence published before this date. However, our focus was on understanding the epidemiology of AF and screening for AF in 2016, and so we wished to find the most contemporary evidence available.

Many inputs for the economic model relied on a single trial (the SAFE study19). Our model tracks AF diagnosed as a result of acute events, such as stroke, but does not estimate serendipitous detection of AF in the absence of screening in the base case. There was no evidence on the proportion of 12-lead ECGs interpreted by a GP that are referred to a cardiologist to help with interpretation, and we had to rely on expert opinion. There was scant information on patient characteristics of screen-detected AF cases, and there was no data on the joint distribution of different risk factors; we therefore had to assume that they were independent (which is unlikely, as, for example, those with a previous history of MI may be more likely to have a previous history of stroke as well).

We assumed that the HR for stroke comparing chronic AF with paroxysmal AF was the same as that for mortality, although there was evidence that this assumption was reasonable. 147 We assumed that the stroke and mortality rates were the same in screen-detected AF as in routinely detected AF, once risk factors have been accounted for, based on the results from Flaker et al. 35 We explored the robustness of our conclusions to this assumption in a sensitivity analysis.

If primary care screening were to be adopted by the NHS, then it would be necessary for all practices to acquire and maintain an ECG machine. We did not include this capital cost in the model, but we note that this would be a consideration in the decision to introduce a national screening programme. We took a NHS perspective, excluding the costs of participants, carers and other agencies. Attending for screening is likely to have a financial impact on participants in terms of travel and time to attend screening, particularly if they have to take time off work to attend.

Findings in the context of previous research

The Irish HIQA study25 used a German study125 to estimate AF prevalence, whereas we used a Swedish study;2 however; as noted, the results from the two studies were very similar (see Prevalence of atrial fibrillation in the UK by age and sex).

Previous systematic reviews were identified for only two of our review questions (prevalence of AF90 and cardiovascular risk for symptomatic vs. asymptomatic AF124). Our findings were in line with the results from these reviews.

Screening test

Because the photoplethysmography screening test was based on a single study69 that did not take place in a primary care setting, and the study was rated as being at high risk of bias for the index test, we performed a sensitivity analysis of the comparison of different ages at screening for a single one-off screen using the most cost-effective screening test omitting photoplethysmography, which was the modified blood pressure monitor. The results were very similar (see Appendix 11, Using a modified blood pressure monitor as the screening test) and the conclusions were unchanged.

Photoplethysmography and modified blood pressure monitors are not currently available in most primary care settings. Therefore, we conducted a further sensitivity analysis using the screening test that was most cost-effective (pulse palpation interpreted by a nurse) when restricting to tests that are currently typically available in primary care (i.e. excluding photoplethysmography and modified blood pressure monitors). The results were very similar (see Appendix 11, Using pulse palpation interpreted by a nurse as the screening test) and, although INBs were slightly lower than in the base case, the overall conclusions were unchanged.

Repeated screening

We explored the impact of repeated screening on a given age cohort by varying the initial and final screening ages, and assuming screening every 5 years, reporting all results for the age cohort population (see Chapter 5, Repeat screening). We found that screening every 5 years from age 65 years to age 80 years had the highest probability of being cost-effective, with screening every 5 years from age 70 years to age 80 years having the second highest probability of being cost-effective (Figure 26). These strategies also have the highest expected INB (Table 48). The optimal strategy from the base case (initial screen at 80 years with no repeat screening) had a low probability of being the most cost-effective strategy and a relatively low INB when repeated screening strategies were included in the decision. Because the results of a single screen were very similar for photoplethysmography, modified blood pressure monitors and pulse palpation by a nurse (see Screening test), we present only the results for pulse palpation by a nurse when comparing different repeated screening strategies, as nurse palpation has no capital costs and is currently available in primary care. The conclusions were unaltered using different screening tests.

FIGURE 26.

Cost-effectiveness acceptability curves for screening strategies with repeated screening every 5 years for various initial and final screening ages, all using opportunistic pulse palpation by a nurse as the screening test. No screening represents no screening at any age. The results are based on incremental costs and QALYs for a single age cohort population. Screening strategies that have a < 10% probability of being cost-effective at any willingness-to-pay threshold are not shown for clarity.

Uptake of systematic opportunistic screening

In this sensitivity analysis we assumed a lower uptake of systematic opportunistic screening, as reported in Morgan and Mant82 (0.305, 95% CI 0.282 to 0.329), assuming that uptake was the same for all age groups. The main impact was to change the optimal screening method from systematic opportunistic screening to systematic population screening. All other conclusions were unchanged. This suggests that our finding that systematic opportunistic screening is the optimal screening approach is sensitive to the uptake of systematic opportunistic screening. The CEACs for different screening ages are shown in Figure 27.

FIGURE 27.

Sensitivity analysis assuming a lower uptake of systematic opportunistic screening: CEACs comparing different ages for a single one-off screen in an age cohort population. Screening strategies that have a < 10% probability of being cost-effective at any willingness-to-pay threshold are not shown for clarity.

Uptake of an electrocardiogram in those with a positive screening result

We conducted a sensitivity analysis using the assumption that only 72.5% of those with a positive screening test result will agree to have an ECG, compared with 100% in the base case. We found that incremental costs, incremental QALYs and INBs were all lower than for the base case, but overall the conclusions were robust to reducing the uptake of a diagnostic ECG (see Appendix 11, Uptake of an electrocardiogram in those with a positive screening result).

Diagnosis in the absence of screening or an atrial fibrillation-related stroke

For the repeated screening strategies, we conducted a sensitivity analysis assuming an annual rate of diagnosis in the absence of screening or an AF-related stoke of 5%, as assumed by Aronsson et al. ,100 compared with 1% based on the control arm of the SAFE study19 used in the base case. As one would expect, the INB decreases as the diagnosis rate in the absence of screening or AF-related stroke increases (Table 49). When the diagnosis rate in the absence of screening or AF-relative stroke is 5%, it is unclear whether it is more cost-effective to initiate repeated screening at age 65 years or age 70 years (Figure 28); this compares with the base case, in which there was a reasonably high probability that the optimal screening strategy was to initiate repeated screening at age 65 years (see Figure 26).

FIGURE 28.

Cost-effectiveness acceptability curves for screening strategies with repeated screening every 5 years for various initial and final screening ages, all using opportunistic pulse palpation by a nurse as the screening test, for the sensitivity analysis with a 5% diagnosis rate in the absence of screening or AF-related stroke. No screening represents no screening at any age. Results are based on incremental costs and QALYs for a single age cohort population. Screening strategies that have a < 10% probability of being cost-effective at any willingness-to-pay threshold are not shown for clarity.

Meta-analysis of previous history of stroke in UK screen-detected atrial fibrillation

In a sensitivity analysis, we used the results of a random-effects meta-analysis for the proportion of screen-detected AF patients with a previous history of stroke (0.112, 95% CrI 0.072 to 0.170); this compares with 0.074 (95% CI 0.032 to 0.116) based on the SAFE study19 used in the base case. We found no change in the conclusions and only a slight decrease in the INBs (see Appendix 11, Meta-analysis of previous history of stroke in UK screen-detected atrial fibrillation).

Proportion of anticoagulant patients receiving directly acting oral anticoagulants instead of warfarin

In the base case, we assumed that 75% of patients prescribed OACs would be prescribed DOACs. We ran sensitivity analyses on the proportion of OAC patients prescribed DOACs instead of warfarin (INR range 2–3), assuming that 50% and 100% of patients were prescribed DOACs. We found no difference in the conclusions and almost no difference in absolute results from the base case (see Appendix 11, Proportion of anticoagulant patients receiving directly acting oral anticoagulants instead of warfarin).

Uptake of oral anticoagulants

In the base case, 74% of patients eligible for OACs receive them, 13% of patients are contraindicated or do not wish to take OACs and the remaining 13% do not receive OACs but could do so. In sensitivity analyses we explored the impact of (1) a reduced OAC uptake of 50% and (2) improving the uptake of OACs so that all 87% of patients who are not contraindicated or who prefer not to take OACs receive OACs. The overall INB was lower if OAC uptake was lower, but this did not alter our conclusions (see Appendix 11, Uptake of oral anticoagulants).

Hazard ratios for stroke and mortality risk for routine-detected atrial fibrillation compared with screen-detected atrial fibrillation

In sensitivity analysis, we assumed that the event rate in screen-detected patients was lower than that for routinely detected AF using the unadjusted HR for symptomatic AF (representing non-screen-detected AF) compared with asymptomatic AF (representing screen-detected AF) found by Flaker et al. 35 of 1.31 (95% CI 1.04 to 1.65), instead of a HR of 1, as assumed in the base case. Our conclusions were unchanged from the base case (see Appendix 11, Hazard ratios for stroke and mortality risk for routine-detected atrial fibrillation compared with screen-detected atrial fibrillation).

Summary of the results

Our results indicate that systematic screening for AF is likely to be cost-effective. This is because of the relatively large net benefit resulting from anticoagulation therapy for patients identified with AF and the relatively low cost of a screening programme. We found that cheaper screening tests (such as photoplethysmography, blood pressure monitors and pulse palpation by a nurse) were more likely to be cost-effective than other screening tests, which is probably because these tests have good sensitivity as well as because they are cheaper than alternative tests. It should be noted, however, that photoplethysmography and blood pressure monitor devices are currently not typically available in primary care, and so adoption of these tests would incur a substantial upfront capital investment (see Strengths and limitations). Furthermore, the evidence informing the DTA of photoplethysmography was rated as being at high risk of bias for the index test, and there was no evidence available from a primary care context. The next most cost-effective screening test was found to be pulse palpation by a nurse. All of these findings rely on the use of a 12-lead ECG diagnostic test interpreted by a GP (referred to a cardiologist when the diagnosis is unclear) for individuals with a positive screening test result.

Both systematic opportunistic and systematic population screening methods were found to be cost-effective compared with no screening (current practice); however, systematic opportunistic screening was more likely to be cost-effective than systematic population screening in our base-case analyses. This result was found to be sensitive to assumptions on the uptake of opportunistic screening, which, if lower than in the base case, changes the optimal method of screening to systematic population screening.

For a given age cohort we found that, if we were to conduct a single screen, then strategies that use a higher age at screening were more likely to be cost-effective. However, when allowing for the possibility of repeated screening strategies with 5-year intervals, single screens were no longer found to be cost-effective. Instead, an initial screen at age 65 years followed by repeat screens every 5 years until age 80 years was found most likely to be cost-effective. If we assume a high (5%) diagnosis rate in the absence of screening or AF-related stroke, then it may be more cost-effective to initiate screening at age 70 years with repeated screens at 75 and 80 years. It should be noted, however, that the majority of the evidence informing the model was based on individuals aged 65–75 years, and so we are less confident of our results outside this age range.

Strengths and limitations

Our model is based on a comprehensive set of reviews of DTA studies (see Chapters 2 and 3), screening RCTs (see Chapters 2 and 4), the natural history of AF and AF screening (see Chapter 5) and a systematic review, network meta-analysis and cost-effectiveness model of OACs. 37 This allowed us to explore a wide range of screening tests and to incorporate net benefits of recently developed DOACs as well as warfarin.

We explored the optimal age range at which to conduct a single ‘prevalent’ screen and also assessed the cost-effectiveness of repeat screening rounds for a given age cohort. We considered only 5-year screening intervals for repeat screens and could have varied the screening interval in further sensitivity analyses. However, previous studies that have investigated this have found that cost-effectiveness depends on the interval of screening, with an interval of around 5 years found to be acceptable. 25,96,99

We assumed that uptake of screening varies with age by crudely dichotomising into < 75 years and ≥ 75 years age ranges, as reported in the SAFE study,19 which found that uptake of systematic population screening was greater in younger than older individuals. No other studies reported uptake of screening by age. Although we incorporated the effect of age for prevalence and the uptake of screening, many other model inputs were based on studies that primarily included those in the age range 65–70 years. We are therefore more confident in our results for this age range than for other age ranges.

The sensitivity analysis exploring repeated screening is not as robust as the base-case analysis. We did not directly model the incidence of new cases of AF or the progression from paroxysmal to persistent/permanent AF or from asymptomatic to symptomatic AF in the model of repeat screening – we simply used information on population size, prevalence and previous diagnosis of AF at each age. This assumes that the patterns of prevalence seen in the study by Norberg et al. 2 would apply to a single age cohort (i.e. no cohort or period effects). We assumed a 1% annual diagnosis rate in the absence of screening and assumed that this did not change with increasing age. When we allowed a higher diagnosis rate in the absence of screening of 5%, there was a suggestion that initiating a repeated screening strategy at 70 years rather than at 65 years might be cost-effective. These limitations should be considered when making final decisions about repeat screening strategies.

We considered only screening programmes based on age, rather than on other risk factors. The SAFE study,19 however, found that systematic screening targeted at high-risk individuals had a similar rate of identification of new cases of AF as the control (no screening), suggesting that targeted systematic screening is unlikely to be cost-effective. An ongoing controlled before-and-after study (due to report in 2016) in Scotland has trialled the AliveCor app used in primary care to screen high-risk individuals,167 which will provide more evidence on this question.

We have averaged over men and women in our model to enable a single decision to be made for both sexes. The SAFE study19 found that screening was more effective in men than in women. For women, only systematic opportunistic screening was effective, whereas for men both systematic opportunistic screening and systematic population screening were effective. Women were less likely to accept an ECG than men. If a systematic screening programme is introduced, individuals need to be fully informed of the consequences of receiving an ECG to confirm the diagnosis if the screening test is positive.

We assumed that the screening test would be conducted at a single visit to primary care and therefore would detect only a small proportion of cases of paroxysmal AF (compared with more intensive monitoring). We made this assumption because of a lack of evidence on prolonged screening tests identified in our systematic review of DTA studies. The proportion of paroxysmal AF detected by population-based screening would increase if long-term monitoring devices were used, and costs might therefore increase too. It is not clear, however, whether or not this would represent a cost-effective use of resources, as paroxysmal AF patients may be less likely to have a CHA₂DS₂-VASc score of > 1 and be eligible to receive anticoagulation therapy. Our review of natural history studies found that stroke risk is lower in patients with paroxysmal AF than in patients with persistent or permanent AF.

It should be noted that some false positives may be diagnosed with another condition (e.g. atrial flutter), and thus receive some benefit from an AF screening programme. We did not capture this benefit in our model; however, incorporating this would increase the cost-effectiveness of a population-based systematic screening programme.

We assumed that the DOAC used is apixaban because this was found to be most cost-effective in the DOAC model. 37 However, several DOACs are currently prescribed in UK practice and all licensed DOACs were found to be similar in terms of cost and benefits. 37 Our conclusions would therefore be robust to the actual DOAC that was prescribed and we envisage that a variety of DOACs would be prescribed for screen-detected AF, as is currently the case for incidentally detected AF. In the base case, we also assumed that 75% of patients prescribed OACs received DOACs, but in practice more patients may receive warfarin for various reasons, including comorbidity and tolerability. Also, in the future a higher proportion of OACs prescribed may be DOACs. Varying the proportion of OACs that were DOACs from 75% to 50% or 100% did not alter our conclusions on the cost-effectiveness of screening strategies.

We have not included capital costs of the screening test devices in our analyses. However, based on plausible assumptions about the purchase costs, lifespan and patient throughput of ECG, photoplethysmography and modified blood pressure monitor equipment (Table 50), we believe that the inclusion of the capital costs of equipment within the screening test cost would have little impact on our conclusions. The capital cost of the most expensive equipment (£1495 for a 12-lead ECG) is ‘spread out’ over the lifespan of the equipment (e.g. 5 years) among patients undergoing screening for AF and patients for whom an ECG is used for other clinical indications (Table 51). The estimated additional capital cost of an ECG per individual screened would be approximately £1, increasing slightly with screening age as the number of patients screened declines. The capital costs per AF patient screened for a modified blood pressure monitor (from £0.33 to £0.87) and photoplethysmography (from £0.08 to £0.21) were lower still if each practice purchased one device dedicated to AF screening. Alternatively, practices might eventually choose to replace existing blood pressure monitors with modified monitors in each consultation and treatment room and use them for screening AF and other patients.

If a screening strategy using equipment not available in practices is selected, then investment in equipment will be required to achieve nationwide primary care screening. This may act as a barrier to initiating screening. A recent review found that 81% of practices have a 12-lead ECG machine available. 44 Investment would therefore be required for the 19% of practices that do not have a machine available. This would be required under all screening strategies, because we assume that all positive screening tests would be followed by a diagnostic 12-lead ECG. Modified blood pressure monitors or photoplethysmography devices are not typically available in primary care and if these screening tests were adopted then this would represent a large capital investment. In particular, photoplethysmography has been asessed only in a single study, which was not carried out in a primary care context and which was at high risk of bias for the index test. Investment in such a technology would be unwise without further research to better understand its diagnostic performance in a primary care screening setting. We estimate that, for photoplethysmography and modified blood pressure monitor screening strategies, for which all practices would have to purchase equipment and some practices would also have to purchase 12-lead ECG equipment for confirmatory testing, the investment needed would be approximately £2.4M and £3.0M respectively (Table 52). In contrast, for strategies that require only some practices to purchase 12-lead ECG equipment for screening or confirmatory testing (e.g. pulse palpation, ECG) the initial investment would be £2.2M.

Findings in the context of previous research

In line with other studies,19,25,96,100 we found that population-based screening is likely to be cost-effective. We found that systematic opportunistic screening was more cost-effective than systematic population screening, which, again, is in line with previous studies. 19 The SAFE19 study is the only other study we are aware of that looked at different screening tests. The DTA results from the SAFE study19 were included in our review of DTA studies, along with the results from 14 other studies. Our model is therefore based on a more comprehensive evidence base and covers a wider range of screening tests than the SAFE model. 19

Review of diagnostic accuracy studies

We identified 15 studies of screening tests for detecting AF, including 12-lead ECG, single-lead ECG, between 1- and 12-lead ECG, pulse palpation, modified blood pressure monitors, photoplethysmography and two-stage testing. Screening tests varied in whether they were interpreted by a cardiologist, a GP, a nurse or an automatic algorithm, although evidence was not available for every test and interpreter combination. There was a high degree of variability between studies and a high level of uncertainty in the estimates of DTA. In general, most tests had a high sensitivity, in excess of 0.9. Specificity was, in general, lower than sensitivity for all of the tests, and was lowest for pulse palpation by a nurse (specificity 0.79), a 12-lead ECG interpreted by a nurse (specificity 0.84) and photoplethysmography (specificity 0.87). Tests with the highest DOR were the 12-lead ECG (regardless of interpreter), the between 1- and 12-lead ECG (automatic or cardiologist interpretation), two-stage tests and the single-lead ECG interpreted by a GP; all of these tests had similar DORs.

In general, for a given interpreter, the results for single-lead ECGs were less accurate and more variable than ECGs with more than one lead. Nurse interpretation of single-lead ECGs performed similarly to single-lead ECGs with other interpretation methods, but nurse interpretation of 12-lead ECGs did not perform as well as 12-lead ECGs with other interpretation methods. Automatic interpretation did not have a consistent impact on test accuracy, with automatic interpretation of single-lead ECGs having a high sensitivity but variable specificity. In contrast, automatic interpretation of ECGs with more leads had good specificity but variable sensitivity. The different two-stage screening strategies all had very high specificity, but sensitivity was high only when a 12-lead ECG was used as the second-stage test.

Review of randomised controlled trials comparing screening strategies

We identified five RCTs comparing screening strategies for AF; however, only two of these provided data that could be included in our review, and only one was included in our primary outcome (the number of new AF diagnoses). The SAFE study19 therefore remains the main source of evidence on the comparative efficacy of different screening strategies for AF. Systematic population (a 12-lead ECG interpreted by a cardiologist) and systematic opportunistic (pulse palpation followed by a 12-lead ECG interpreted by a cardiologist) screening strategies were found to be similarly effective, with an estimated 170 individuals needed to screen to detect one additional AF case compared with no screening. There was no evidence that systematic screening targeted to high-risk individuals was effective compared with no screening.

Uptake of systematic population screening was typically around 50%, although it was as high as 70% in one study. There was variability in uptake between practices (between 22% and 70% in the SAFE study19). Reasons for not attending for screening were varied, although older age and decreased mobility were common themes. The proportion of individuals having their pulses checked under systematic opportunistic screening varied across studies (between 30% and 66%) and between practices within studies (from 8% to 93%). The proportion of individuals consulting with their GP was not reported, so it is unclear how much these uptake rates are driven by consultation rates, GPs offering pulse palpation and uptake of pulse palpation by individuals. Of those with an irregular pulse who did not have a previous diagnosis of AF, approximately 18% did not attend for an ECG test, although, again, this varied across practices.

Subgroup analyses based on the SAFE study19,84 indicated that both systematic population screening and systematic opportunistic screening were more effective for men than for women, and this was especially the case for systematic population screening. The efficacy of screening was not found to vary with age, despite AF prevalence being strongly associated with age.

Review of the natural history of atrial fibrillation and atrial fibrillation screening

We searched for studies to inform the natural history of AF and screening for AF for use in our economic evaluation. This comprehensive search identified 48 studies that provided data relevant to our model. The prevalence of AF was variable across studies but showed similar trends with age and sex. The studies with the most intensive ascertainment of AF gave the highest estimates of prevalence and were in line with each other. 2,125 The proportion of AF that is undiagnosed was estimated to be 35%; however, this estimate was variable across screening studies, reflecting the method of screening and test used as well as differences in populations. As expected, the higher estimates came from studies with more intensive screening tests (e.g. long-term continuous monitoring). It is clear that, if the objective is to detect paroxysmal AF, then long-term continuous monitoring is necessary.

Uptake of systematic population screening was fairly consistent across studies, with a mean estimate of 64% (95% CrI 54% to 73%), although this was shown to depend on age in the SAFE study. 19 In contrast, there was a high degree of variability in the uptake of systematic opportunistic screening, ranging from 30% to 70% (regardless of age), raising concern that the high level of opportunistic screening achieved in the SAFE study19 may not be seen in practice.

As expected, our results suggest that a high proportion of screen-detected AF is likely to be asymptomatic. Asymptomatic AF has a lower risk of stroke and mortality; however, after adjusting for other risk factors, there is no evidence that the risk of stroke or mortality depends on whether AF is asymptomatic or not. 35 We found that our results were robust to this assumption. We estimated that paroxysmal AF progresses to chronic (persistent or permanent) AF at a rate of 0.15 per year. There is evidence of a ‘dose–response’ relationship, with stroke and mortality risk increasing as AF progresses from paroxysmal to persistent to permanent.

Economic evaluation

Our results indicate that both systematic opportunistic screening and systematic population screening followed by DOAC therapy when indicated are likely to be cost-effective compared with no screening (current practice). However, systematic opportunistic screening was more likely to be cost-effective than systematic population screening as long as the proportion of flagged individuals who have their pulses checked observed in the SAFE study19 is realised in practice.

We found that photoplethysmography, modified blood pressure monitors and pulse palpation by a nurse were more likely to be cost-effective than other screening tests because they are cheaper than other screening tests while have adequate test sensitivity. This finding relies on the use of a 12-lead ECG diagnostic test interpreted by a trained GP (with referral to a cardiologist when the diagnosis is unclear) in individuals with a positive screening test result.

For a single screen of a given age cohort, we found that strategies that use a higher age of screening were more likely to be cost-effective. However, when allowing for the possibility of repeated screening strategies with 5-year intervals, single screens were no longer found to be cost-effective. Instead, an initial screen at age 65 or 70 years followed by repeat screens every 5 years until age 80 years was found to be most likely to be cost-effective, provided that compliance with screening and treatment does not decline with increasing age.

Screening tests

Although we were able to explore screening test and interpreter combinations, we were restricted by the evidence that we identified in our systematic review of DTA studies. Only a small number of studies met our inclusion criteria, not all screening test and interpreter combinations were included and, when present, few observations were reported, leading to a lack of statistical power to detect meaningful differences. None of the included studies was assessed to be at low risk of bias across all domains and to have low concerns with regard to applicability. In particular, only four studies were conducted in primary care. Although we had preplanned a wide range of subgroup analyses and sensitivity analyses, enough data were available to perform only some of these.

Although photoplethysmography had very high test sensitivity, this estimate was based on a single study69 that was not based in a primary care setting, and for which the index test was rated as being at high risk of bias and the applicability of the patient selection was unclear. Similarly, a 12-lead ECG interpreted by a GP was evaluated in only a single study,70 which was at high risk of bias according to patient selection. We are therefore cautious in interpreting the results from these screening tests.

Setting

We restricted our analyses to a primary care setting, but opportunistic screening could potentially be conducted in other settings, for example community centres, such as pharmacies,104 or in secondary care such as in elderly patients attending for minor surgery. 136

Comparative evidence

For systematic opportunistic screening, a greater proportion of the 75 newly identified cases were diagnosed outside the screening programme (44/75, 59%) than within it (31/75, 41%). One explanation for this is that GPs in the systematic opportunistic screening arm changed their usual practice to check for AF more frequently (the ‘Hawthorne effect’, in which behaviour changes when under observation). If this is the case, then the full benefits seen in the systematic opportunistic arm may not be realised outside the context of a RCT.

None of the RCTs identified in our review considered repeated screening, only a one-off screen. We therefore do not have any comparative evidence on the effectiveness of repeated strategies.

We have considered screening programmes based only on age, rather than on other risk factors. The SAFE study,19 however, found that systematic screening targeted at high-risk individuals had a similar rate of identification of new AF cases as no screening, suggesting that targeted systematic screening is unlikely to be cost-effective. A controlled before-and-after study evaluating the AliveCor ECG app used in primary care to screen high-risk individuals is ongoing in Scotland (due to report in 2016),167 which will provide more evidence on this question.

Other arrhythmias

Some false-positive patients may be diagnosed with another condition (e.g. atrial flutter), and therefore receive some benefit from an AF screening programme. We did not capture this benefit in our model; however, incorporating this would increase the cost-effectiveness of a population-based systematic screening programme.

Paroxysmal atrial fibrillation

Continuous monitoring devices are showing promise for detecting paroxysmal AF. 22 Engdahl et al. 22 found that, in patients who screened negative on a 12-lead ECG but who had two or more risk factors based on the CHADS₂ score, a period of 2-week monitoring using a hand-held monitoring device identified 7.4% of paroxysmal AF cases. This study specifically targeted high-risk individuals, however, and so it is unclear how large the yield would be in a general population. In the general population, paroxysmal AF patients may be less likely to have a CHA₂DS₂-VASc score of > 1 and may be eligible to receive anticoagulation therapy. Also, our review of natural history studies found that stroke risk is lower in patients with paroxysmal AF than in those with persistent or permanent AF. If such devices were to be used in a screening context, there would need to be clear instructions for their use, and there would also be concerns about the devices not being returned.

Screening age

The studies informing many model inputs were primarily carried out with those in the age range 65–70 years, and we are therefore more confident about our results in this age range than in other age ranges.

We did not directly model the incidence of new cases of AF or the progression from paroxysmal AF to persistent/permanent AF or from asymptomatic AF to symptomatic AF in the model of repeat screening – we simply used information on population size, prevalence and previous diagnosis of AF at each age. This assumes that the patterns of prevalence seen in the study by Norberg et al. 2 would apply to a single age cohort (i.e. no cohort or period effects). We assumed a 1% annual diagnosis rate in the absence of screening, and that this did not change with increasing age. When we allowed a higher diagnosis rate in the absence of screening of 5%, there was a suggestion that initiating a repeated screening strategy at 70 years rather than 65 years might be cost-effective.

Economic model

Our review of the literature for inputs to the economic model was the most comprehensive search of the literature to date. We combined the results from multiple studies using meta-analysis when possible. However, despite this, many inputs to the economic model relied on the results from a single trial. 19 There was no evidence on the proportion of 12-lead ECGs interpreted by a GP that are referred to a cardiologist to help with interpretation, and so we had to rely on expert opinion. There was scant information on patient characteristics of screen-detected AF cases and no data on the joint distribution of different risk factors, so we had to assume that they were independent (which is unlikely, as, for example, those with a previous history of MI may be more likely to have a previous history of stroke as well).

We took a NHS perspective, excluding the costs of participants, carers and other agencies. Attending for screening is likely to have a financial impact on participants in terms of travel and time to attend screening, particularly if they have to take time off work to attend.

If a nationwide primary care screening strategy using equipment not available in practices is adopted, then investment in equipment will be required. This may act as a barrier to initiating screening. A recent review found that 81% of practices have a 12-lead ECG machine available. 44 Investment would therefore be required for the 19% of practices that do not have a machine available. This would be required under all screening strategies, because we assume that all positive screening tests would be followed by a diagnostic 12-lead ECG. Modified blood pressure monitors or photoplethysmography devices are not typically available in primary care and, if these screening tests were adopted, this would represent a large capital investment. Investment in photoplethysmography would be unwise without further research to better understand its diagnostic performance in a primary care screening setting. We estimate that, for photoplethysmography and modified blood pressure monitor screening strategies, for which all practices would have to purchase equipment and some practices would also have to purchase 12-lead ECG equipment for confirmatory testing, the investment needed would be approximately £2.4M and £3.0M respectively (see Table 52). For strategies that require only some practices to purchase 12-lead ECG equipment for screening or confirmatory testing (e.g. pulse palpation, ECG), the initial investment would be £2.2M.

Screening test accuracy

Our study was restricted to comparing screening tests that had been evaluated in DTA studies that used a 12-lead ECG interpreted by a cardiologist as the reference standard. Only 15 studies met our inclusion criteria, and of those only four were carried out in a primary care or community setting, which is where we envisage screening taking place. When replicates existed of the same screening test/interpreter combination, heterogeneity in results was observed. This suggests a need to replicate the results based on a single study or few studies. In particular, the DTA results for photoplethysmography were based on a single study69 rated as being at high risk of bias for the index test and conducted outside a primary care setting. Robust evidence on the performance of this device in a primary care screening setting, for a range of cut-off values, would be of merit. Such a study should also record information on hygiene, use in people with skin conditions, ease of use, durability, length of test time and sensation.

We found that a 12-lead ECG interpreted by a GP had a very good diagnostic performance compared with a 12-lead ECG interpreted by a cardiologist/specialist. However, this was based on a single study70 that was at high risk of bias with regard to patient selection. A study to replicate this result in a screening population in a format that reflects how screening is likely to be implemented (i.e. ECGs interpreted by a trained GP with referral to a cardiologist/specialist in cases that are unclear) may be of value, although resources may be better directed towards training GPs in the interpretation of ECGs for diagnosing AF. Such a study would also be able to provide information on the proportion of ECGs that are unclear and that would thus be referred to a cardiologist/specialist, which was an assumption in our economic model.

The development of new devices to detect AF is a fast-moving area, and a horizon-scanning exercise to identify new literature as it becomes available, and to update our review, would be of value. It should be noted that we assumed that only a small proportion of paroxysmal AF would be detected by the screening strategies that we considered. However, the availability of newer technologies that allow for intensive testing over a period of time would allow a higher proportion of paroxysmal AF cases to be detected. More intensive screening tests are likely to be more costly and to incur initial capital investment, although this would not necessarily be the case, for example in the case of smartphone apps. Studies comparing the DTA of smartphone apps with that of a 12-lead ECG interpreted by a cardiologist/specialist reference standard are needed. The benefits of detecting paroxysmal AF through screening also need to be better understood, to help evaluate more intensive screening tests.

Comparative effectiveness of screening strategies for atrial fibrillation

The primary comparative evidence of effectiveness of screening strategies remains the SAFE study. 19 This is a well-conducted study, but it evaluates only a single screen for a given age range and does not follow up to measure long-term outcomes of screening or the value of repeat screening. To fully evaluate the benefits of a screening programme and change practice, there is a need for a comparative study that measures long-term outcomes (stroke and mortality) for a screening strategy compared with a no screening strategy. The ongoing STROKESTOP study should provide evidence on this over a longer follow-up period than previous studies;26,40 however, the results are specific to the format of screening used (intermittent ECG over 2 weeks). Because an intensive screening test was used in the STROKESTOP study, a large proportion of those with newly detected AF will have paroxysmal AF. This study should therefore provide valuable evidence on the impact of screening on long-term outcomes for paroxysmal AF patients. The benefits of early detection of AF through screening will depend on the subsequent medical care and anticoagulation therapy used. The generalisability of the results of the STROKESTOP study will depend in part on how similar the use of anticoagulation therapies for AF patients is between Sweden and the UK.

Given the huge number of potential screening strategies (test, interpreter, age at initial screen, screening interval, age at final screen, systematic opportunistic screening, systematic population screening, systematic targeted screening, setting) it is not possible to design a primary research study to evaluate all combinations. Some form of modelling/simulation is therefore necessary to evaluate different options to identify those that are most likely to be effective so that they can be evaluated further in a primary research study. The design of such a primary research study may take the form of a cluster randomised factorial trial and make use of linked electronic health records. This is likely to be challenging because of the fast-moving development of screening test technologies. Uncertainties in the economic model that a RCT could inform include the uptake of screening by age and sex, patient characteristics of those with screen-detected AF (including CHA₂DS₂-VASc score), the proportion of patients diagnosed in the absence of screening and reason for diagnosis (event, symptoms or other) and anticoagulation therapies used (including the proportion of OACs that are DOACs).

We have focused on screening in a primary care setting. Further research into alternative settings including secondary care (e.g. outpatient clinics) and community settings (e.g. pharmacies) would be of value.

Repeat screening

To thoroughly assess repeat screening strategies, good evidence is required on the incidence of new cases of AF (and the proportion of these that are undiagnosed), progression from paroxysmal to persistent/permanent AF, progression from asymptomatic to symptomatic AF and the diagnosis rate in the absence of screening that is not a result of an AF-related stroke. A large epidemiological study with linked electronic health records may be the most efficient way to provide some of this evidence; however, such a study may not identify undiagnosed AF that could be identified by screening. A nested follow-on study to a cluster RCT of screening could be conducted in those who were screened and who tested negative, by randomising individuals to a repeat screen after 5 years or not. This would provide evidence on the incidence of AF that is diagnosed and undiagnosed and the yield of a repeat screen. All previous models of repeat screening (including ours) use routinely collected data on the prevalence/incidence of AF, which likely underestimates undiagnosed AF cases.

Implementation of screening

The economic model results were sensitive to the proportion of flagged individuals who have their pulses checked. This proportion was estimated to be very different (30% and 66%) in the two studies that reported it19,82 and a high degree of variability was seen across practices. 19 Research on the effectiveness of implementation strategies to improve the proportion of pulses that are taken among flagged individuals would be of value if a systematic opportunistic screening strategy were recommended.

Screening method

Our findings support the use of systematic opportunistic screening for AF, with evidence that a screening strategy with an initial screen at age 65 years and repeated screens every 5 years until age 80 years is likely to be cost-effective. Operationalising opportunistic screening would require some training of clinicians, possibly using age-triggered prompts in the electronic medical record and simple practice-level protocols to ensure that patients in whom AF is suspected based on the screening test are offered an ECG. The practice protocol, based on national guidance, would need to specify who to invite, operationalised through the Read coding of patients to exclude individuals with a previous diagnosis of AF, as it can be distressing to patients to be invited to screening for a condition that they know they have. Similar systems are in place for other screening programmes (e.g. breast and cervical screening), suggesting that such a system could be implemented for AF screening.

Wright et al. 173 found that a multifaceted intervention programme aimed at health professionals in primary care led to an increase in detection of AF cases, highlighting the need for an implementation strategy to embed screening in general practices. Several studies performed screening alongside influenza vaccination,117,118 but yield was fairly low. Another study174 looked at changes in AF detection rates in practices that implemented the NHS Health Checks programme compared with practices that did not implement the NHS Health Checks programme and found no difference in detection rates. This is perhaps not surprising, as AF was not an aim of the NHS Health Checks programme, although it could be in the future.

Screening test

As discussed earlier, we do not consider photoplethysmography to be appropriate for a screening study until further evidence on its diagnostic performance has been collected. The next most cost-effective screening tests were modified blood pressure monitors and nurse pulse palpation. Modified blood pressure monitors are not available in most practices. If modified blood pressure monitors were to be recommended as the screening test in a national screening programme, an investment in the devices would be required (see Chapter 6, Strengths and limitations). This could be implemented over a period of time by replacing existing blood pressure monitors at the end of their lifespan with modified devices. As well as the capital cost, training would be required in their use, although the automatic interpretation algorithm would potentially be simpler to interpret than pulse palpation. A screening programme could therefore allow both screening tests as options, with practices without a modified blood pressure monitor using pulse palpation by a nurse for screening. Potentially, a nationwide screening programme could be rolled out in a stepped or otherwise randomised fashion, enabling additional evidence to be collected on the optimal screening strategies.

A recent review44 found enthusiasm for screening in primary care but a recognised need for the training of nurses. Barriers to AF screening were listed as lack of time, workload, lack of appointments, staffing levels, access to equipment, funding, waiting times at anticoagulation clinics and the cost of anticoagulation therapies. These resource issues would need to be addressed if a national screening programme were to be adopted. Taggar et al. 44 found that health-care assistants and nurses were mostly responsible for conducting ECGs, with interpretation by GPs, but these staff reported that they would benefit from training. In the future, pulse palpation and use of modified blood pressure monitors and ECG machines could become a core part of training and continued professional development for nurses and health-care assistants. The review44 found that only 81% of practices had an ECG machine. Under all of our screening strategies, we assumed that a diagnostic 12-lead ECG interpreted by a trained GP (with referral to a cardiologist/specialist in cases that are unclear) would be conducted in those who screen positive, and so an additional investment in an ECG machine would be required in the 19% of practices that do not already have one. This investment would also be used for other indications as well as AF, and so would bring additional benefits.

Other arrhythmias

A consequence of screening for AF is that other arrhythmias, such as atrial flutter, may be detected. This is not captured in our economic model. Assuming that routine care for patients with other arrhythmias is cost-effective, then the detection of these patients through screening would only increase the cost-effectiveness of a screening programme. It is important that individuals who are invited to be screened understand that other arrhythmias, as well as AF, may be detected and that, if they are diagnosed, they are fully informed of the risks and treatment options.

Future developments

New devices to detect AF are becoming available and are under evaluation (patches, smartphone/watch devices, iPads, hand-held devices), as well as devices used for other reasons that can also detect AF (pacemakers, ICDs and implantable loop recorder devices). These innovations may be of relevance to a screening programme when their diagnostic performance is better understood.